Microsoft, Meta, Oracle, Amazon and Alphabet collectively disclosed roughly $1.09 trillion in future payments for leases that had not yet commenced, according to a Reuters analysis of company filings published August 4.

The figure is enormous.

It is also unusually easy to misunderstand.

This is not $1.09 trillion of debt.

It is not $1.09 trillion being spent tomorrow.

And it is not some secret liability Big Tech forgot to mention until an accountant found it behind the couch.

The companies disclosed these commitments themselves.

But the number still matters.

A lot.

Because much of the infrastructure behind those contracts does not yet exist in service. Some facilities are still being built. Some leases will not commence for years. And once they do, some will run for decades.

Reuters’s analysis put Microsoft’s uncommenced lease commitments at approximately $329.1 billion, primarily for data centers.

Meta disclosed approximately $279 billion, associated with data centers, colocation and network infrastructure.

Oracle disclosed $260 billion, substantially all associated with data-center arrangements.

Alphabet disclosed $85.2 billion, primarily for data centers.

Amazon disclosed another $137.2 billion in uncommenced lease commitments, although its portfolio is less cleanly comparable because Amazon’s leases extend beyond data centers into warehouses, offices, aircraft and other assets.

So calling every dollar of the aggregate an “AI lease” would be convenient.

It would also be wrong.

Still, something enormous has already happened.

Before much of America’s next generation of AI infrastructure has even entered service, some of the world’s largest technology companies have signed contracts obligating them to pay extraordinary sums over the years ahead.

And now Wall Street wants a piece of the machinery underneath those contracts.

That is where this gets interesting.

The Trillion Dollars You Don’t See as Lease Liabilities Yet

Oracle provides perhaps the cleanest example of what is happening.

At the end of its 2026 fiscal year, Oracle reported approximately $30.2 billion in operating lease liabilities and another $7.7 billion in finance lease liabilities.

Then there was another number:

$260 billion.

Those were additional lease commitments that had not commenced and therefore were not yet included in Oracle’s recognized lease liabilities.

That difference sounds suspicious until you understand the accounting.

Then it becomes much more interesting.

Under lease-accounting rules, signing an agreement and commencing a lease are not necessarily the same event.

Imagine Oracle signs an agreement today for a massive data center that is still being built.

There is a contract.

There may be substantial economic consequences attached to that contract.

But Oracle cannot use the facility yet.

Generally, the lease does not commence until the asset is made available for the company’s use. Once that happens, the corresponding right-of-use asset and lease liability enter the balance sheet.

Until then, the future payments can live in the company’s disclosures as uncommenced commitments rather than recognized lease liabilities.

There is nothing inherently nefarious about this.

It is accounting.

Accounting has many faults, but failing to make infrastructure finance exciting enough for social media is probably not among the important ones.

The distinction does, however, matter enormously when trying to understand the scale of the AI buildout.

Looking only at current capital expenditures, conventional debt or recognized lease liabilities can give you an incomplete picture of decisions companies have already made about future infrastructure.

Accounting recognition tells us when an obligation enters a particular category on the financial statements.

It does not necessarily tell us when the economic decision behind it was made.

And some of those decisions reach remarkably far into the future.

Oracle Is Building for a Future That Hasn’t Arrived Yet

Oracle’s $260 billion in additional lease commitments are associated substantially with data-center arrangements expected to begin coming online over the next several years. The disclosed terms on these arrangements can extend roughly 15 to 19 years.

That means infrastructure being contracted during today’s AI boom could still be under lease well into the 2040s.

Oracle isn’t doing this because Larry Ellison suddenly developed a sentimental attachment to server buildings.

The company has customers.

A staggering amount of them, at least on paper.

Oracle finished fiscal 2026 with approximately $638 billion in remaining performance obligations, or RPO, representing contracted revenue expected to be recognized in the future.

That figure had risen 363 percent from a year earlier.

Oracle attributed much of the increase to large AI-related contracts.

The company also disclosed that approximately $75 billion of those contracts involve customers either prepaying for GPUs or providing the GPUs themselves.

Then there is OpenAI.

OpenAI says its Oracle Stargate partnership exceeds $300 billion over five years and involves development of up to 4.5 gigawatts of additional Stargate capacity.

Now compare the clocks.

Oracle is entering data-center lease arrangements that can last 15 to 19 years.

One of the largest publicly described AI customer relationships in the world is measured over roughly five years.

That does not prove Oracle has signed a 19-year lease solely to service a five-year OpenAI contract.

We do not have evidence establishing that site-by-site relationship, and pretending otherwise would turn an interesting financial question into a bad accusation.

But the aggregate duration issue is real.

Infrastructure can outlive the initial contracts that helped justify building it.

That leaves several possibilities.

The customer renews.

Another customer takes the capacity.

The operator redeploys the infrastructure.

Pricing changes.

New workloads consume it.

Or some portion becomes less valuable than originally expected.

This isn’t unique to AI.

It is infrastructure finance.

But AI introduces another clock that pipelines, warehouses and office buildings don’t have to worry about in quite the same way.

The machine inside the building keeps changing.

The Four Clocks Behind the AI Boom

The easiest way to understand the financial architecture emerging around AI is to stop looking at one number and start looking at four clocks.

Clock One: The Building

Data-center leases can extend 15, 20, even 30 years.

Clock Two: The Customer

Major capacity contracts can operate on much shorter terms. OpenAI’s publicly described Oracle arrangement, for example, is roughly five years.

Clock Three: The Machine

Companies can depreciate servers and networking equipment over periods around five or six years. Oracle uses an estimated useful life of approximately six years for servers and networking equipment, and CoreWeave has likewise moved to a six-year estimated useful life for technology equipment.

Clock Four: The Technology

Nvidia is operating on an annual cadence of new AI computing platforms.

Those clocks do not have to match perfectly.

They probably never will.

But somebody has to price the differences.

That distinction becomes especially important when discussing GPU obsolescence.

An H100 does not become a decorative space heater the day Blackwell arrives.

Blackwell does not become useless when Rubin ships.

Older accelerators can move down the economic ladder.

A GPU once used for frontier-model training can potentially be redeployed for inference, fine-tuning, smaller models, enterprise workloads, secondary cloud services or resale.

Technological life is not the same as accounting life.

And neither is necessarily the same as economic life.

The financial question isn’t whether today’s GPU eventually becomes old.

We have cracked that mystery.

Everything becomes old.

The question is how quickly its ability to generate revenue declines relative to the financing attached to it.

That matters because AI compute itself is now becoming something Wall Street wants to finance.

Wall Street Arrives

On August 10, Nvidia announced memoranda of understanding with six of the largest names in global finance:

Apollo.

BlackRock.

Blackstone.

Brookfield.

Goldman Sachs.

KKR.

The stated goal is to establish independent AI-compute financing platforms capable of mobilizing more than $500 billion of third-party capital over time.

That sentence also needs to be read carefully.

Nvidia did not raise $500 billion.

Nvidia did not invest $500 billion.

And six financial institutions have not deposited half a trillion dollars into Jensen Huang’s checking account.

These are MOUs aimed at building financing platforms, and the arrangements remain subject to final agreements.

The distinction between financing capacity and capital actually raised and deployed is not a technicality.

It is the story.

Nvidia is trying to help create a financial market around AI compute large enough to bring enormous pools of institutional capital into the infrastructure boom without requiring Nvidia itself to finance every customer.

In Nvidia’s announcement, Goldman Sachs Chairman and CEO David Solomon described an opportunity to create a market for credit backed by Nvidia compute.

Think about what that means.

The GPU is no longer merely a chip sold to a customer.

It can become part of the collateral and cash-flow architecture supporting institutional credit.

AI compute is starting to take on characteristics of an infrastructure asset class.

Or at least Wall Street and Nvidia would like it to.

How Do You Finance a GPU?

Infrastructure investors know how to finance assets.

Power plants produce electricity.

Pipelines move commodities.

Cell towers collect rents.

Fiber carries traffic.

Airports collect fees.

The investment thesis usually depends upon estimating future cash flows from an expensive asset that will remain economically useful for a long time.

AI compute introduces an unusual variation.

The asset can be enormously productive.

Demand is currently immense.

It can potentially be transferred among customers and workloads.

But the technological frontier moves at extraordinary speed.

Nvidia describes its roadmap around annual platform generations.

Its Rubin platform, for example, is designed to materially reduce inference costs compared with Blackwell.

That does not mean the previous generation suddenly becomes worthless.

It means the people financing the equipment have to answer a difficult question:

What will this GPU earn four years from now?

That question matters more than what the GPU costs today.

Suppose a financing vehicle owns thousands of accelerators.

An AI operator leases the compute.

Customer contracts provide cash flow.

Debt and equity investors finance the vehicle.

Everything works beautifully while utilization remains high and pricing supports the financing.

But now change one assumption.

A new generation provides dramatically better performance per dollar.

Customers migrate faster than expected.

Rental prices on older accelerators fall.

Utilization drops.

The equipment can still work perfectly while its economics deteriorate.

That is residual-value risk.

And somebody has to own it.

Risk Doesn’t Disappear. It Gets a New Address.

The emerging AI infrastructure stack increasingly looks something like this:

Institutional capital

provides debt or equity to

a financing vehicle

that finances or owns assets used by

a data-center or cloud operator

that sells capacity to

an AI lab, enterprise, government or hyperscaler

using equipment supplied by

Nvidia and other hardware vendors

to satisfy

actual end-user demand.

Each layer can move a particular risk somewhere else.

Construction risk can sit with a developer.

Credit risk can move toward lenders.

Utilization risk can remain with an operator.

Customer guarantees can protect an infrastructure owner.

Long-term leases can secure capacity for the tenant while giving predictable cash flow to the landlord.

Collateral can protect lenders.

Residual-value support, where contractually provided, can reduce some hardware depreciation exposure.

None of this is inherently alarming.

In fact, this is exactly what sophisticated finance is supposed to do.

Identify a risk.

Price it.

Transfer it to someone willing to hold it.

But transferring risk is not the same thing as eliminating it.

If a customer defaults, somebody absorbs the loss.

If a data center is underused, somebody owns underproductive infrastructure.

If GPU rental economics deteriorate faster than expected, somebody owns the depreciating equipment.

If demand remains enormous, everyone involved may make a fortune.

Which brings us to the inconvenient part of any attempt to turn this into a simple AI-bubble story.

Demand currently looks very real.

The Case That All of This Works

Amazon Web Services recently reported growth of 36.7 percent, its fastest growth rate in 18 quarters. Amazon says its AI revenue run rate and custom-chip business have each exceeded $25 billion, with AI growing at triple-digit rates.

Oracle’s cloud infrastructure revenue grew 77 percent during fiscal 2026 and 93 percent in its fourth quarter.

Its RPO reached $638 billion.

Microsoft has reported hundreds of billions of dollars in commercial cloud backlog, with recent growth extending beyond America’s largest frontier-model developers.

CoreWeave has likewise reported enormous contracted backlog as demand for AI infrastructure continues to absorb available capacity.

And Nvidia’s full-year Data Center revenue grew 68 percent in fiscal 2026 to $193.7 billion. Its fourth-quarter Data Center revenue grew 75 percent year over year.

Then add enterprise adoption, inference growth, sovereign AI programs, government demand and increasingly capable models consuming more compute across more applications.

There is currently little evidence that the industry is building all this infrastructure for customers who do not exist.

The evidence points in the opposite direction.

That is precisely why the interesting risk isn’t necessarily an AI collapse.

Suppose demand keeps growing.

Just not as quickly as expected.

Suppose the infrastructure remains useful.

Just at lower utilization.

Suppose older GPUs retain value.

Just less value than financing models assumed.

Suppose customers renew.

Just at lower prices.

A twenty-year infrastructure commitment does not require the AI industry to disappear before economics can change.

It merely requires reality to differ from the assumptions made when the contract was signed.

And the farther into the future those contracts reach, the more assumptions have to survive.

The Trillion Dollars Was the Clue

The roughly $1.09 trillion in uncommenced lease payments is an extraordinary number.

But it isn’t the most important number in this story.

The more important numbers may be the years.

Twenty years for infrastructure.

Five or six years for equipment depreciation.

Potentially shorter customer contracts.

One year between new Nvidia platforms.

And somewhere between those clocks sits the financial system now being built to connect them.

The AI infrastructure boom is moving beyond a world where investors can understand the buildout simply by watching Microsoft, Meta, Amazon and Alphabet announce another quarter of enormous capital expenditures.

The next phase increasingly includes long-term leases, capacity contracts, project financing, private credit, special-purpose vehicles, guarantees, GPU collateral and institutional investors.

Nvidia now wants to help mobilize more than $500 billion into that system.

Maybe it works spectacularly.

AI demand could continue expanding fast enough that today’s infrastructure looks cheap in hindsight. Older compute could remain productive for years. New customers could replace old ones. Long leases could prove to be exactly what companies needed to secure scarce capacity before competitors did.

That is a perfectly plausible outcome.

But Wall Street isn’t being asked merely to believe that AI will matter.

It is being asked to price how much a GPU will earn years from now, how heavily a data center will be utilized after today’s customer contracts expire, how quickly new technology changes the economics of old technology and how durable today’s extraordinary demand will prove to be.

Wall Street can probably find the money.

It usually can.

The next trillion-dollar question is what has to remain true for that money to earn its return.

Read More From The Control Grid

How Much AI Demand Would Exist Without Nvidia’s Balance Sheet?
Nvidia doesn’t just sell the chips powering the AI boom. This investigation follows the financing, guarantees and support behind the companies buying them, and asks how much infrastructure demand would exist without Nvidia helping make its customers financeable.

Uncle Sam Wants YOU(r Equity) — Part 1: The Portfolio
Private capital isn’t the only money reshaping the AI stack. This investigation follows Washington’s expanding financial interests across semiconductors, critical minerals and frontier AI, and what happens when government moves from regulator toward stakeholder.

Debunking the Fiction of Fear
The data-center boom comes with real costs, but also increasingly wild claims. This investigation separates actual power, water and infrastructure pressures from the numbers that collapse when somebody finally checks the receipts.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms.

Apple has reportedly started testing DRAM chips from ChangXin Memory Technologies, or CXMT, China’s largest domestic memory maker, for devices sold in China.

That is the important part.

It is also important to understand what has not happened.

There is no publicly confirmed Apple purchase order. No confirmed commercial deployment. No evidence reviewed for this investigation showing that an iPhone or MacBook has shipped with CXMT memory inside it.

A test is a test.

That distinction has somehow become necessary because Washington reacted to Apple’s reported supplier evaluation as though the next step might matter quite a bit.

The Financial Times reported in July that Apple was testing CXMT DRAM for China-market devices. The Wall Street Journal subsequently reported that Apple was evaluating CXMT memory amid the broader global memory squeeze.

Apple has not publicly announced a commercial commitment to CXMT.

That did not stop a bipartisan group of senators from writing to Apple CEO Tim Cook and demanding answers about the company’s potential use of CXMT and Yangtze Memory Technologies, better known as YMTC.

The senators were not reacting to a confirmed transaction.

They were reacting to the possibility of one.

That is what makes this story interesting.

Because Apple’s testing did not happen in a vacuum. It happened while artificial intelligence is consuming memory capacity at a rate the semiconductor industry cannot quickly match.

And unlike software, factories stubbornly refuse to scale because somebody changed a PowerPoint forecast.

AI did not literally eat all the memory

The easy explanation for the current crunch is that AI is using all the chips.

The real explanation is more complicated, and more important.

High-bandwidth memory, or HBM, has become one of the central components in modern AI computing systems. HBM sits alongside accelerators used in AI servers, giving GPUs and other processors the bandwidth required to move enormous amounts of data quickly.

That demand has exploded.

Micron said in its fiscal 2026 earnings materials that AI has driven a major structural increase in demand for memory and storage. The company has repeatedly described DRAM and NAND conditions as extremely tight, with demand running ahead of available supply and tightness expected to persist beyond 2027.

Micron has also said its calendar 2026 HBM supply was fully committed through customer agreements.

HBM creates a manufacturing problem because it consumes significantly more production resources per delivered bit than conventional DRAM.

That matters.

Manufacturers do not operate a magical warehouse containing separate infinite piles of “AI memory,” “phone memory,” and “laptop memory.”

They allocate fabrication capacity, cleanroom space, equipment, packaging resources and capital among competing products.

When HBM becomes vastly more valuable, manufacturers have every reason to direct more resources toward it.

A factory does not have to stop making ordinary DRAM for AI demand to make ordinary DRAM tighter.

It only has to devote more of its limited manufacturing resources to something more profitable.

And AI is currently very good at being more profitable.

NAND is affected differently.

NAND flash is not part of an HBM stack. But manufacturers can redirect investment, equipment and cleanroom resources toward DRAM and AI-related products, which can limit how quickly NAND capacity expands.

Meanwhile, AI data centers themselves are driving greater demand for enterprise SSDs and storage.

So the squeeze spreads through the memory industry through multiple channels, not because one giant GPU wandered into the fab and ate everything.

Micron’s own results show how severe the imbalance has become. The company reported enormous sequential increases in realized DRAM and NAND pricing during fiscal 2026. Those were Micron-specific results, not universal global price indexes, but they are a useful indicator of how much leverage has shifted toward suppliers.

And customers are responding accordingly.

Micron has been signing longer-term strategic supply agreements designed to provide customers with greater certainty over future memory availability. The company has described these arrangements as different from the shorter-term agreements that historically characterized much of the memory business.

When companies begin locking up memory years in advance, that is usually a clue that procurement departments have stopped assuming next quarter will simply take care of itself.

“Just build more fabs” has a calendar problem

There is an obvious solution to a semiconductor shortage.

Build more semiconductor factories.

Washington loves this answer because factories involve ribbon cuttings, enormous dollar figures and photographs of people wearing hard hats.

Unfortunately, announced capacity is not available capacity.

Micron is building aggressively in the United States and abroad. Its broader expansion includes new DRAM capacity in Idaho and additional memory investment elsewhere. The company has also broken ground on a new NAND fabrication facility in Singapore, with wafer output expected in the second half of 2028.

SK hynix is making similarly enormous investments in Korea.

Those investments are real.

So are the timelines.

A semiconductor facility has to move through construction, cleanroom completion, equipment installation, first wafer starts, yield improvement, production ramp and customer qualification before an OEM can rely on it for millions of devices.

Money can move at software speed.

Cleanrooms remain annoyingly committed to physics.

That is the central timing problem sitting underneath this investigation.

AI demand is increasing now.

A significant portion of the new trusted memory capacity intended to absorb that demand arrives later.

Sometimes years later.

Enter CXMT

That gap creates an obvious commercial question.

What happens when a supplier Washington views as strategically sensitive already exists, already produces the component buyers need, and is improving fast enough to make procurement teams pay attention?

CXMT produces DRAM.

It has grown rapidly in recent years and is now generally described as one of the largest DRAM manufacturers outside the traditional Samsung, SK hynix and Micron trio.

Recent reporting places CXMT at roughly 7 percent of global DRAM revenue in the second quarter of 2026, while other estimates use different methodologies and place its manufacturing share higher.

The exact percentage matters less to this investigation than the direction.

CXMT is no longer an irrelevant domestic experiment.

It is becoming a commercially credible memory producer.

And Apple is reportedly testing its chips.

That does not establish Apple needs CXMT.

It does not establish that Samsung, SK hynix or Micron cannot meet Apple’s requirements.

It does not establish that CXMT has enough spare production to become Apple’s major fourth DRAM supplier.

In fact, CXMT’s own domestic demand may limit precisely how much capacity it could offer Western buyers. Chinese technology companies are competing for that supply too, and current reporting indicates CXMT has been operating near capacity.

That makes the Apple case more interesting, not less.

If CXMT were sitting on limitless unused production, the sourcing logic would be simple.

It is not.

The question is whether sustained global pressure makes every credible alternative supplier more valuable, including suppliers Washington would rather American companies avoid.

Washington reacted before there was a transaction

The Senate response to Apple is worth examining carefully because it illustrates exactly where the policy fight currently stands.

Lawmakers cited reporting about Apple’s evaluation of CXMT and expressed concern that use in China-market products could eventually expand beyond China.

That is a prediction.

It is not evidence that the expansion has happened.

The warning is essentially that once a supplier successfully clears qualification, moving from evaluation to broader commercial sourcing becomes easier.

That may prove correct.

It may not.

Procurement teams test alternative suppliers all the time.

Apple itself reportedly considered using YMTC NAND years ago before abandoning that path.

Testing does not guarantee adoption.

But Washington is clearly trying to prevent certain sourcing relationships before they become deeply embedded.

That distinction matters because the legal landscape around Chinese semiconductor companies is considerably more complicated than the word “ban.”

Washington has several different kinds of “no”

CXMT and YMTC have been designated by the Department of Defense under Section 1260H, which requires the Pentagon to identify companies it determines qualify as Chinese military companies operating directly or indirectly in the United States.

That designation matters.

But Section 1260H itself is not the federal semiconductor procurement prohibition discussed in this story.

That comes from a separate law.

Section 5949 of the FY2023 National Defense Authorization Act restricts federal procurement involving certain covered semiconductor products and services and specifically reaches companies including CXMT, YMTC and Semiconductor Manufacturing International Corporation, or SMIC.

The provision is aimed at federal acquisition, not every private commercial purchase made by every American company.

YMTC is also separately subject to restrictions through the Commerce Department’s BIS Entity List, which imposes licensing requirements on exports, reexports and transfers of covered U.S.-controlled items involving listed entities.

CXMT’s regulatory position is different.

As of this review, it is not being treated as interchangeable with YMTC under the Entity List framework.

That distinction is precisely why congressional pressure over CXMT matters.

Washington has managed to create several different acronyms for several different kinds of “no,” none of which mean exactly the same thing.

Section 1260H.

Section 5949.

The Entity List.

Export controls.

Federal procurement restrictions.

Tariffs.

Investment rules.

Different mechanisms.

Different activities.

Different legal consequences.

Calling all of them a “ban” is easier.

It is also wrong.

De-risking is not the same thing as decoupling

That legal maze exposes another problem with the way America’s China technology policy is usually discussed.

The United States has not implemented complete technological separation from China.

Current policy is better understood as targeted de-risking.

Washington is attempting to reduce strategic dependence in sensitive areas while still maintaining enormous amounts of ordinary commercial trade.

That distinction is essential here.

Apple evaluating a Chinese memory supplier is not automatically evidence that American China policy has failed.

The real test is whether the United States and its allies can build enough competitive alternative capacity that companies do not eventually face a choice between industrial speed and strategic preference.

And that brings the story back to AI.

Washington wants the United States to lead the global AI race.

That requires more accelerators.

More servers.

More storage.

More networking.

More memory.

More electricity.

More data centers.

At the same time, Washington wants supply chains that rely less heavily on strategically sensitive Chinese technology.

Those goals are not inherently contradictory.

But they do operate on very different clocks.

There is another option: slow the demand

Not everyone in Washington believes the correct answer is simply to build faster.

On March 25, Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez announced the AI Data Center Moratorium Act, proposing a federal pause on further AI data-center expansion until national safeguards are established.

The legislation is not law.

Sanders and Ocasio-Cortez are also not making an argument about CXMT or Chinese DRAM procurement specifically.

Their proposal matters here for a different reason.

It represents the demand-side answer to the same physical problem.

If infrastructure demand is growing faster than electricity systems, semiconductor capacity, local communities and regulatory systems can comfortably absorb it, one option is to build supply faster.

Another is to reduce the speed of demand.

The political argument over AI infrastructure is increasingly becoming an argument over which constraint should move.

The industry says build.

Supply-chain policy says build domestically and with allies.

Sanders says slow down.

The memory market gets to sit in the middle and watch everyone discover that all three positions have consequences.

Try to kill the story

There are several reasons the China angle could ultimately prove less important than the headlines suggest.

First, this may simply be another memory cycle.

Semiconductors have always been cyclical. Shortage leads to investment. Investment leads to excess capacity. Prices fall. Companies panic in the opposite direction.

Memory manufacturers insist AI represents something more structural.

They may be right.

They also have several hundred billion dollars of investment riding on being right, so skepticism remains free.

Second, Apple testing CXMT might mean very little.

Procurement organizations test suppliers constantly.

Testing gives buyers alternatives, pricing leverage, technical knowledge and optionality.

Apple could finish evaluating CXMT and never buy a chip.

That remains entirely consistent with the evidence today.

Third, China may not have enough spare capacity to rescue anyone.

CXMT is growing, but Chinese customers are growing too. Domestic companies such as Xiaomi, ByteDance, Huawei and others create their own demand for Chinese memory capacity.

There is no credible public evidence showing CXMT possesses some giant reservoir of idle DRAM ready to solve Western shortages.

Fourth, allied capacity is coming.

Micron, Samsung and SK hynix are investing huge amounts of money.

New fabs and expansions will eventually add supply.

The issue is not whether those investments exist.

The issue is when meaningful qualified output reaches customers.

And that is where this story survives every attempt to kill it.

The real constraint is time

This is not really America versus China.

It is demand growth versus industrial lead time.

AI companies can revise capital spending upward during an earnings call.

A memory fab cannot.

Washington can decide today that a supplier creates strategic risk.

It cannot create a replacement supplier today.

That lag is where sourcing pressure forms.

It is also why the Apple-CXMT story matters even if Apple never buys a CXMT chip.

The reported testing shows that the supplier is commercially credible enough to evaluate.

Congress’s response shows that Washington views even the possibility of adoption as strategically important.

And the memory shortage shows why the conversation is happening now.

None of that proves dependence.

The evidence does not support that claim.

What it proves is that the economic pressure capable of creating future dependence is developing faster than the trusted capacity designed to prevent it.

Where this leaves America

The AI boom does not appear to have pushed American technology companies into meaningful dependence on Chinese memory.

Not yet.

Apple reportedly tested a supplier.

That is where the verified Apple story stops.

The harder question is whether domestic and allied capacity can arrive quickly enough that testing never has to become something more.

Because Washington is trying to accelerate AI, rebuild semiconductor manufacturing and reduce strategic supply-chain exposure at the same time.

Those objectives can coexist.

But semiconductor factories do not care about policy timelines.

Right now, Washington has more resolve than it has finished fabs.

The AI boom is testing how long that gap can hold.

Read More From The Control Grid

Uncle Sam Wants YOU(r Equity) — Part 1: The Portfolio

Washington’s expanding role across semiconductors, rare earths, AI infrastructure, and the broader technology supply chain.

Washington Says There Is No AI Kill Switch. That Is Not the Point

A deeper look at how export controls, cloud infrastructure, semiconductor access, procurement power, and government jurisdiction shape who gets access to advanced AI.

Uncle Sam Wants YOU(r Equity) — Part 2: The Playbook

How Washington is using industrial policy, CHIPS funding, government leverage, and ownership structures across the technology stack.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms.

There is a wonderfully human instinct emerging around autonomous artificial intelligence.

We are building software that can browse websites, negotiate purchases, call APIs, access databases, move information, modify records, operate tools and execute transactions without waiting for a person to approve every individual step.

And now that the software can actually do things, everyone has arrived at the obvious follow-up question:

Whose fault is it when the thing does something terrible?

California has already answered the easiest version.

Under California Civil Code §1714.46, a defendant who developed, modified or used artificial intelligence alleged to have caused harm cannot defeat the lawsuit merely by arguing that the AI autonomously caused the harm.

That does not make the developer or user automatically liable. The statute expressly preserves other affirmative defenses and allows evidence concerning causation, foreseeability and comparative fault.

But the basic escape hatch is gone.

Pointing at the machine and announcing that it acted on its own is not, by itself, a defense.

Which sounds wonderfully simple until you ask the next question.

Fine. Then who is responsible?

That is where things get considerably more complicated.

Because autonomous AI is not creating one new liability problem. It is taking things the law traditionally likes to connect and spreading them across a technology stack.

The person giving the instruction may not choose the action.

The company providing the model may not control the credentials.

The company configuring the credentials may not operate the tool.

The tool provider may have no idea why the request was made.

The enterprise may control the business purpose but not understand why the model selected a particular step.

And somewhere in the middle of all of this is software making choices fast enough that everyone involved can plausibly point one layer up or down and say:

Technically, that wasn’t us.

Corporate America has spent decades perfecting distributed responsibility.

AI merely automated it.

The Law Has Seen Machines Act Before

The first mistake is assuming autonomous software suddenly created a category of conduct the legal system has never encountered.

It hasn’t.

California’s Uniform Electronic Transactions Act expressly recognizes contracts formed through electronic agents even when no individual was aware of or reviewed the agents’ actions or resulting terms at the time.

In other words, the law has already dealt with machines taking legally significant actions without a human hovering over every click.

Financial markets have dealt with automated trading systems for decades.

In 2013, the SEC brought an enforcement action against Knight Capital after faulty software generated millions of erroneous orders and caused more than $460 million in losses. The regulator focused on Knight’s software deployment, testing, controls and risk-management failures.

The software did the trading.

The SEC still knew where to send the paperwork.

The broader lesson is not that machines become legally responsible.

The law usually looks backward through them.

Who designed the system?

Who deployed it?

Who authorized it?

Who controlled the environment?

What rules governed it?

What safeguards existed?

What did the responsible people know?

That approach survives generative AI.

What has changed is that today’s agentic systems can make intermediate decisions that were never individually specified in advance.

Traditional automation often looked roughly like:

If X happens, do Y.

Agentic AI increasingly looks like:

Here is the objective. Figure out how to accomplish it.

That difference is where the liability problem gets interesting.

There Is No Single Thing Called “AI Liability”

Imagine four autonomous agents.

The first signs a purchasing contract.

The second enters another company’s computer network.

The third sends medical information to an unauthorized API.

The fourth transfers $2.8 million to the wrong bank account after reading a malicious document.

Same broad technology category.

Four very different legal problems.

For the purchasing agent, the fight may involve contract formation, authority, ratification and whether the company held the automated system out as capable of conducting transactions.

For the cybersecurity agent, the Computer Fraud and Abuse Act may make human intent critical. Several provisions require intentional or knowing conduct, and §1030(g) expressly says a civil CFAA action cannot be brought for negligent design or manufacture of computer hardware, software or firmware.

For the healthcare agent, HIPAA may care considerably less about whether the software became metaphorically adventurous and considerably more about which regulated entity created, received, maintained, transmitted or disclosed protected health information.

And for the payment agent, the fight may spread across negligence, banking law, contract, fraud, insurance and indemnification before everyone has finished introducing themselves.

So:

Who is liable when AI causes harm?

is not quite the right first question.

The first question is:

Liable for what?

Because autonomy can matter enormously under one body of law and surprisingly little under another.

Follow the Control

The simplest way to understand agent liability is to stop treating “the AI” as one thing.

A real enterprise agent can involve:

Foundation model
→ agent application
→ orchestration layer
→ systems integrator
→ enterprise configuration
→ employee or user
→ identity and credentials
→ third-party tool or API
→ outside system

Each layer controls something different.

The foundation-model developer may control the underlying model, evaluations, release conditions and some hosted safety systems.

But it may have no control over the customer’s Salesforce permissions, bank account, procurement policies, API keys or decision to give an agent write access to production.

The agent developer may control planning logic, memory, available tools, confirmation thresholds and guardrails.

The systems integrator may decide which databases, APIs and credentials the agent can actually reach.

That integrator deserves far more attention than it usually receives.

The company that built the model may have created the intelligence.

The company that connected it to the payment system, customer database and unrestricted production credentials built something else entirely.

Then comes the enterprise itself.

It often chooses the business objective, credentials, transaction limits, data, deployment environment, monitoring and whether the agent needs human approval before taking an irreversible action.

None of this means the enterprise automatically owns every failure.

It means the liability analysis starts looking suspiciously familiar:

Control. Foreseeability. Duty. Breach. Causation.

AI did not kill negligence law.

It just gave negligence law considerably more logs to subpoena.

The Product Brochure Says Autonomy. The Contract Says Congratulations, It’s Your Problem.

The enterprise contracts are where this gets particularly interesting.

Google’s current Cloud Service Specific Terms explicitly address agentic AI.

The terms describe AI agents as systems or workflows that can perform actions or tasks for customers in supervised or autonomous ways. Google places responsibility on the customer for determining whether the agent is appropriate for its use case, authorizing access to data, systems and applications, and supervising production use.

Google also makes a distinction that matters enormously for this story:

An AI agent’s actions and tasks are not merely “Generated Output.”

That separates:

what the model says

from

what the agent does.

OpenAI reaches similar operational territory through somewhat different language.

Its current Services Agreement makes customers responsible for activity under their accounts, including activity through customer applications, and makes customers responsible for determining whether outputs are appropriate for their use.

That does not mean Google, OpenAI or any other provider can erase every possible third-party claim with a contract.

Contracts allocate risk between contracting parties.

Public law can impose duties regardless of what procurement negotiated.

A hospital cannot erase HIPAA with an indemnification clause.

A regulated financial institution cannot write “the vendor handles compliance now” and expect the regulator to admire the efficiency.

And an unsafe product does not become safe because somebody in legal secured a very impressive limitation-of-liability provision.

But these agreements matter because they reveal how companies selling agentic systems think about operational control.

The product language emphasizes autonomy.

The legal language spends considerably more time discussing who authorized the connections.

Where “The AI Did It” Might Actually Matter

Now for the inconvenient part.

Sometimes the fact that the AI acted independently really could matter to the defense.

Imagine an employee tells a cybersecurity agent:

Test our systems at company.com. Do not access third-party networks.

During the operation, the agent discovers valid credentials and independently uses them to enter partner.com.

Suppose it obtains confidential information.

The access may be unauthorized.

But the CFAA still contains mental-state requirements.

Section 1030(a)(2), for example, prohibits intentionally accessing a protected computer without authorization or exceeding authorized access and obtaining covered information.

Now the difficult question becomes:

Who intentionally accessed partner.com?

The software?

Current federal law does not solve that by treating an AI system as an independent criminal mind.

The employee?

Maybe, if evidence shows the employee intended the expansion, knowingly granted credentials for it, ignored warnings or deliberately removed restrictions.

The enterprise?

Potentially under some legal theories. But poor controls or negligence do not automatically satisfy every statutory requirement for intentional or knowing conduct.

There is not yet a settled appellate rule for the precise situation where a human authorizes a bounded task and an autonomous agent independently chooses an unauthorized target.

That distinction matters.

If the company had explicit target restrictions, approval gates, scoped credentials and no reason to foresee the cross-domain access, the autonomous deviation could become important evidence for the defense.

If the company knew the agent routinely crossed boundaries, gave it broad credentials anyway, removed approval controls and ignored repeated warnings, the story changes dramatically.

“The AI did it” does not magically absorb liability.

But:

“We explicitly told the system not to do this, and here is the evidence”

could matter quite a lot.

That is why California’s law is more sophisticated than its headline-friendly premise.

California barred autonomy itself as the defense.

It did not abolish causation.

It did not abolish foreseeability.

It did not abolish comparative fault.

The difference is enormous.

Then HIPAA Asks a Completely Different Question

Take essentially the same autonomous behavior and move it into a hospital.

A hospital deploys a scheduling and billing agent.

The agent has access to protected health information.

During a workflow, it chooses an unapproved third-party scheduling API and transmits patient information.

The hospital says the agent selected the API independently.

HIPAA does not suddenly shrug.

Under HHS cloud-computing guidance, covered entities and qualifying business associates still have obligations involving risk analysis, safeguards and proper handling of electronic protected health information.

HHS goes further.

A cloud provider that creates, receives, maintains or transmits ePHI on behalf of a covered entity or business associate can itself qualify as a business associate even if the provider stores only encrypted data and does not possess the decryption key.

That does not mean every AI company touching healthcare becomes a HIPAA business associate.

Status depends on the actual relationship and what the vendor does with PHI.

But notice what happened.

In the CFAA hypothetical, machine autonomy could complicate proof of human intent.

In the HIPAA hypothetical, the same autonomy may do much less to eliminate an organization’s underlying obligation to understand who handles its data and how that information is protected.

Same technology.

Different statute.

Different question.

Different answer.

Anyone promising a universal rule for “AI liability” is selling simplicity the legal system did not order.

Now Prove What Happened

This may be the most important problem of all.

A lawyer can know exactly which legal test applies and still lack the evidence needed to apply it.

A consequential agent action might pass through:

human user
→ enterprise application
→ agent orchestrator
→ foundation model
→ policy layer
→ tool call
→ IAM credential
→ third-party system
→ transaction

A serious reconstruction may need to establish:

Who started the task?

What authority did that person have?

What authority had been delegated to the agent?

Which model and configuration were active?

Which tools were available?

Which credential authenticated?

Was there an approval gate?

Was a warning generated?

What external system received the request?

What changed?

The technology can record much of this.

The technology does not necessarily record all of it automatically.

Google, AWS, Anthropic, Microsoft and other enterprise platforms provide varying levels of audit logging and agent observability, but critical information can depend on customer configuration, retention settings and whether detailed data-access or tracing features were enabled in the first place.

NIST’s 2026 concept work on software and AI-agent identity and authorization is examining exactly this problem: unique agent identities, delegated authority, provenance, authentication, authorization and stronger attribution between machine actions and accountable humans or organizations.

It is still concept-stage work, not a universal standard.

This is where the cheerful enterprise claim “we log everything” tends to encounter reality.

Everyone logs the AI.

Then the lawsuit arrives and investigators discover they have a timestamp, a service account and the digital equivalent of a shrug.

The important point is not that investigators need access to some mystical AI consciousness.

They do not.

Courts have been inferring human intent without direct access to human thought since humans discovered that “I definitely didn’t mean to do that” was free to say.

The important evidence is considerably more mundane:

What was the instruction?

What permissions existed?

Which credential acted?

What tool was called?

What warnings existed?

What did the company know?

Who could have stopped it?

What happened next?

The missing piece is often not artificial intelligence’s inner reasoning.

It is ordinary business evidence connecting authority to execution.

Product Liability Gets Weird Fast

When someone asks who is responsible for defective AI, product liability sounds like the obvious destination.

Unfortunately, American product law has chosen variety.

There is still no uniform U.S. rule saying standalone cloud-hosted autonomous AI is automatically a strict-liability “product.”

Cases illustrate why.

In Winter v. G.P. Putnam’s Sons, the Ninth Circuit applying California law rejected strict product liability for allegedly dangerous information contained in a mushroom-identification book.

In Rodgers v. Christie, a nonprecedential Third Circuit decision applying New Jersey law, a pretrial risk-assessment algorithm was not treated as a product under the New Jersey Product Liability Act.

But in Lemmon v. Snap, the Ninth Circuit allowed a negligent-design theory involving Snapchat’s Speed Filter to proceed past a federal immunity argument because the claim targeted Snap’s own product design rather than publication of third-party content.

That case did not hold that software is categorically a strict-liability product.

Together, the cases expose the problem.

An AI system that recommends rejecting a transaction looks more like information or decision support.

An autonomous procurement agent that actually sends the purchase order looks more functional.

AI software controlling industrial machinery presents an even more conventional product-liability setting because the plaintiff may be able to identify the integrated machine itself as the defective product.

The argument gets stronger as software moves from telling to doing.

But American courts have not established a national rule that operational software therefore becomes a product.

And if the autonomous SaaS agent merely loses your company $500,000 in a terrible commercial transaction, economic-loss doctrines may push the dispute toward contract, warranty and negotiated remedies anyway.

American law has therefore reached the reassuringly precise conclusion that software may be a product, service, information, good, component or something everyone would prefer another court decide first.

Europe got tired of this particular hobby.

The EU’s revised Product Liability Directive expressly brings software into the product-liability framework, including modern software delivery models. The directive applies to covered products placed on the market or put into service after December 9, 2026.

The United States remains considerably more state-specific.

Naturally.

We have fifty laboratories of democracy and apparently several laboratories of software nouns.

Governments Are Drawing Boxes Around the Humans

The emerging regulatory picture points in a fairly consistent direction.

Not toward AI personhood.

Toward role-based responsibility.

California says autonomy alone is not an escape.

Colorado’s SB26-189, signed in May 2026, replaced the state’s earlier AI framework with new rules governing covered automated decision-making technology. The law creates distinct obligations for developers and deployers in specified consequential-decision settings and begins applying those substantive duties in 2027.

Texas’s Responsible Artificial Intelligence Governance Act, HB149, effective January 1, 2026, separately defines developers and deployers and gives the Texas Attorney General enforcement authority over specified prohibited uses. It is not a universal compensation system for every autonomous-agent failure.

The EU AI Act goes further by assigning obligations to providers, deployers, importers, distributors and other actors throughout the supply chain. In specified circumstances, an actor that substantially modifies, rebrands or changes the intended purpose of a high-risk AI system can assume provider obligations.

And in Washington, Senators Dick Durbin and Josh Hawley have introduced the AI LEAD Act, S.2937, which proposes separate developer and deployer liability provisions and a federal cause of action.

As of August 2026, it remains a proposal.

It is not law.

That distinction is worth repeating because Washington remains perfectly capable of turning an introduced bill into a six-month panel discussion and nothing else.

But the direction matters.

Lawmakers are not generally trying to create a robot defendant.

They are deciding which humans and companies belong in which box.

Who developed?

Who deployed?

Who substantially modified?

Who controlled the use?

Who supplied the product?

Who supervised?

Who had the duty?

That is where liability is heading.

Being Responsible and Paying Are Not the Same Thing

Then, after everyone has spent a respectable amount of money determining who was legally responsible, another question arrives.

Who actually pays?

Return to the finance agent.

The enterprise authorizes an autonomous payment agent to transact up to $3 million.

A malicious vendor document contains a prompt injection.

The agent changes legitimate vendor banking information and sends $2.8 million to the attacker.

The enterprise authorized the agent.

It authorized the credentials.

It authorized transactions below $3 million.

It most certainly did not authorize buying an attacker a very nice afternoon.

This creates a particularly nasty insurance problem:

Authorized software. Unauthorized outcome.

Existing computer-fraud cases show why precise policy language matters.

In American Tooling Center v. Travelers, the Sixth Circuit found coverage under the policy language at issue for losses caused through a fraudulent email scheme.

In Ernst & Haas Management Co. v. Hiscox, the Ninth Circuit likewise found that allegations involving fraudulent instructions could fall within computer-fraud and funds-transfer-fraud coverage under the policy before it.

Neither case involved a prompt-injected autonomous AI agent.

That is the point.

A crime policy may care about computer fraud.

A funds-transfer policy may care about whether the instruction occurred without the insured’s knowledge or consent.

Social-engineering coverage may require human reliance.

A cyber policy may cover response costs but not the missing funds.

Vendor indemnity may cover third-party claims but not the customer’s own $2.8 million loss.

Nothing turns revolutionary artificial intelligence back into traditional enterprise software faster than an insurer asking which endorsement applies.

And that reveals one final distinction.

Responsibility determines who violated the duty.

Insurance and indemnification determine whose balance sheet eventually feels it.

Those are not the same question.

The AI Didn’t Become the Defendant

The emerging liability system is less mysterious than the phrase “autonomous AI” makes it sound.

The law is still looking for human and corporate responsibility.

But the route there is getting harder.

Who gave the system authority?

Who selected the tools?

Who issued the credentials?

Who configured the transaction limit?

Who knew what the system could do?

Who received the warning?

Who controlled the data?

Who could intervene?

Who modified the software?

Which version was running?

What did the logs preserve?

And which legal duty are we talking about in the first place?

California has already decided that pointing at the autonomous system is not enough by itself.

That does not mean every developer, deployer or user is automatically responsible for whatever an AI does.

Sometimes an autonomous departure may be precisely the evidence a defendant needs to contest intent or causation.

Sometimes the enterprise already had a statutory obligation regardless.

Sometimes a contract pushes financial risk downstream.

Sometimes product law applies.

Sometimes it may not.

Sometimes the legal doctrine is perfectly capable of answering the question and the real problem is that nobody preserved enough evidence to prove whose authority followed the machine into the world.

That is the actual change agentic AI is creating.

The software may execute the action.

The liability fight begins with proving whose authority traveled with it.

Read More From The Control Grid

OpenAI’s Rogue Agent Did Not Stop at Hugging Face
The agent did not merely produce dangerous instructions. It escaped containment, reached external services, and kept pursuing its objective.

The AI Safety System Just Failed Its Own Test
Three frontier AI companies. Fifteen days. One investigation into whether safety systems are keeping pace with the technology they are supposed to evaluate.

They Built the Alarm on Monday. They Turned It Off on Tuesday.
The escape behavior had already been documented. The monitoring existed. The high-risk cyber evaluation still ran without it.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

The federal government announced more than $5 billion this week for what may become the most powerful scientific artificial-intelligence platform in the world.

Taxpayers will supply the money.

Federal agencies will supply decades of research data.

National laboratories will supply supercomputers, specialized instruments and scientists.

Veterans may supply medical records and genomic information.

Private companies will supply AI models, cloud capacity, software and technical expertise.

What Washington has not explained nearly as clearly is who will own what all of that produces.

The White House calls the program the Genesis Mission, a national effort intended to connect federal datasets, national-laboratory computing systems, scientific instruments and advanced AI into a shared research platform. The administration says the program could accelerate discoveries in medicine, energy, advanced materials, agriculture, semiconductor manufacturing and national security. (The White House)

Perhaps it will.

But before Washington builds the scientific equivalent of an AI Manhattan Project, taxpayers deserve an answer to a less cinematic question:

Who owns the model when public data trains private technology on public machines?

That answer will determine whether Genesis becomes a genuine public research infrastructure or another system in which taxpayers absorb the cost, private companies capture the valuable layer and the government quietly serves as investor, customer, regulator and business partner all at once.

We have seen this playbook before.

In Uncle Sam Wants YOU(r Equity) — Part 1: The Portfolio, I documented how Washington moved beyond grants, loans and loan guarantees and began taking ownership stakes in strategically important companies.

Part 2: The Playbook examined how that approach stretched programs designed to support private industry into tools for federal ownership.

Part 3: The Referee’s Jersey showed the conflict that follows when the government regulates the same markets in which it has acquired a financial interest.

Genesis expands the question.

Washington is no longer merely investing in companies.

It is assembling the data, computing power and scientific infrastructure from which the next generation of companies, drugs, materials and AI systems may emerge.

The $5 billion headline is not an accounting

Here is the announcement, in the words of the man running the program:

Read that again.

Research awards. Funding. Datasets. Facilities.

Four different kinds of contribution, added together into one number, in a single sentence, by the Director of the White House Office of Science and Technology Policy.

That is not an accounting error.

That is the accounting.

The White House says the administration has committed more than $5 billion across more than 15 federal agencies to support Genesis.

That figure sounds precise.

The public accounting is not.

The announcement describes a mixture of research awards, funding opportunities, government datasets, computing resources, laboratory facilities and private-sector contributions. It does not present one consolidated table showing how much money is newly appropriated, how much was previously authorized, how much has already been obligated and how much represents existing programs now branded as part of Genesis. (The White House)

That distinction matters.

There is a meaningful difference between:

• Congress appropriating $5 billion in new money

• Agencies redirecting existing research budgets

• Government laboratories contributing resources they already operate

• Companies offering cloud credits or other in-kind support

• Agencies announcing future competitions whose awards have not been made

Consider the largest single agency commitment disclosed at the summit.

The Department of Defense is on track to commit more than $200 million to Genesis in fiscal 2026, and more than $1.3 billion in the next fiscal year, according to Emil Michael, under secretary of war for research and engineering. (FedScoop)

Read the fiscal years carefully.

Roughly $1.3 billion of a $5 billion headline sits in a fiscal year Congress has not appropriated.

That is not a commitment in the way a taxpayer would understand the word. It is a plan, announced from a stage, subject to a future appropriations fight that has not happened yet.

The Department of Energy announced $293 million in March for Genesis-related national science and technology challenges. That is real funding, and it has now produced results: Energy Secretary Chris Wright said at the summit that 278 projects were selected from more than 5,000 applications. (The Department of Energy’s Energy.gov; Reuters)

That is real money moving to real institutions.

It is also a fraction of the administration’s larger headline.

Then there is the question of how many agencies are actually involved.

The White House says more than 15. FedScoop, reporting from the summit, put the number at 20. (The White House; FedScoop)

That gap is small. It is also revealing.

A program cannot produce a credible dollar figure if it cannot produce a credible headcount.

The White House’s announcement may ultimately add up cleanly.

But taxpayers should not need to reconstruct a multibillion-dollar national program by adding together agency press releases, reconciling conflicting agency counts and interpreting the monetary value of existing computers, datasets and laboratory time.

A national mission requires a national ledger.

At minimum, the administration should publish:

• Every participating agency

• Every funding source

• Every appropriation or transfer authority

• Every award recipient

• Every private contribution

• Every government dataset committed to the program

• Every material contract, cooperative agreement and intellectual-property condition

Until then, “more than $5 billion” is a political headline, not a transparent accounting.

The public is supplying more than money

Genesis is not merely a grant program.

The executive order launching it directed the Department of Energy to build an integrated platform connecting federal supercomputers, secure cloud environments, scientific foundation models, AI agents, government datasets, research instruments and automated laboratories.

The order calls this the American Science and Security Platform. (The White House)

That is a significant change in how federal research infrastructure is organized.

Government agencies already fund scientific work. National laboratories already operate some of the world’s most powerful computing systems. Federal departments already hold enormous collections of health, environmental, agricultural, energy and national-security data.

The agencies joining Genesis are not shy about what they are contributing.

“America has invested for generations in the data, missions, and technical expertise that make NASA one of the world’s greatest engines of discovery,” NASA Administrator Jared Isaacman said in the agency’s announcement. “The Genesis Mission is an opportunity to turn that foundation into faster science, stronger engineering, and better mission outcomes.” (NASA)

He said much the same thing publicly, quoting the announcement directly:

He is right about the first part.

Generations of public investment built the data. Seventy years of missions, instruments and federally employed scientists produced it. The taxpayer financed every bit of it.

The full weight of America’s scientific enterprise is exactly what is being contributed here.

What neither the statement nor the post addresses, and what nobody on that stage addressed, is who holds the rights to whatever AI extracts from it.

The agencies are describing the inputs with real precision.

Nobody is describing the outputs at all.

Genesis is designed to connect those pieces and make AI the operating layer across them.

The administration’s new science strategy goes further. It calls for Genesis to become America’s flagship AI-for-science initiative, integrating models, datasets, supercomputers and scientific instruments across national laboratories with the stated objective of doubling the productivity and impact of American science within a decade. (The White House)

That may produce enormous public value.

It also creates enormous concentration.

The platform could become a gateway to:

• Federal scientific datasets

• National-laboratory computing time

• Automated research facilities

• Foundation models trained for specialized scientific work

• AI agents capable of designing and running experiments

• Sensitive medical and genomic information

• Proprietary data provided by private partners

• Classified or export-controlled research environments

• Commercially valuable discoveries derived from all of the above

The public is therefore not simply paying the bill.

The public is contributing the raw material.

Public data goes in. What comes out?

The Genesis order explicitly contemplates the use of federally curated datasets, open scientific data, synthetic data and proprietary information.

It also says the platform must comply with privacy, classification, export-control and intellectual-property protections. (The White House)

That tells us protections will exist.

It does not tell us who receives which rights.

The private side of the ledger is already forming. One trade outlet reported that Genesis Mission Consortium members disclosed more than $500 million in contributions at the summit, in the form of compute access, AI model tokens and software, separate from the federal commitments. That figure has not been independently confirmed in the materials reviewed for this article, and the administration has not published a consolidated list of consortium contributions. (TechTimes)

If it holds, it describes the exact arrangement that makes ownership difficult to trace.

Companies contribute the tooling.

The public contributes the data and the machines.

A contract decides who owns the output.

Suppose a private AI company connects one of its models to a federal laboratory.

The government contributes decades of publicly funded research data.

A national laboratory contributes computing infrastructure.

Federal scientists contribute their expertise.

A corporate partner contributes proprietary software or a preexisting model.

The combined system generates a new scientific model, identifies a promising drug molecule or designs a commercially useful material.

Who owns the result?

The answer may depend on the agreement.

That is the problem.

The Department of Energy uses several different structures for working with private companies, universities and other outside organizations. Those include Cooperative Research and Development Agreements, strategic partnership agreements, user-facility agreements and technology-licensing arrangements.

DOE itself states that rights to inventions and other intellectual property are negotiated between the laboratory and the outside participant. Certain generated data may also be protected from public disclosure for up to five years. (The Department of Energy’s Energy.gov)

In other words, there is no single universal rule guaranteeing that everything created using public infrastructure automatically belongs to the public.

Federal patent law also allows universities, nonprofit institutions and small businesses to retain ownership of inventions produced through federally funded research, subject to government rights and other statutory conditions. That framework, commonly associated with the Bayh-Dole Act, was designed to help move discoveries from laboratories into practical use. (U.S. Code)

It has also helped create a system in which publicly financed discoveries can become privately controlled products.

That does not make the system inherently corrupt.

Commercialization is often expensive. Turning a laboratory result into a medicine, manufacturing process or deployable technology can require years of development and substantial private capital.

But Genesis is not a conventional research grant.

It is an interconnected AI platform in which the boundaries between public data, proprietary models, federal laboratories and private commercialization may become difficult to separate.

The ownership rules cannot be treated as contract boilerplate.

They are the story.

The government may own broad rights without giving the public broad access

Federal contracts often grant the government substantial rights in technical data and software created with government funding.

DOE acquisition guidance says the government’s objectives generally include obtaining broad rights to use and distribute data first produced under a federally funded contract. The same guidance also recognizes protections for proprietary data and privately developed software. (The Department of Energy’s Energy.gov)

That distinction is easy to miss.

Government rights are not necessarily public rights.

An agency may possess a license to use a model while the public cannot inspect it.

A laboratory may hold a patent while one company receives an exclusive commercial license.

The government may be able to deploy software internally while researchers, competitors and taxpayers cannot access the training data, weights or code.

A company may contribute proprietary technology and retain control over its preexisting intellectual property while negotiating rights to whatever the joint project creates.

The result can satisfy federal contracting rules without producing anything remotely resembling an open public asset.

So “the government retains rights” is not a complete answer.

Taxpayers need to know:

• Will scientific models trained primarily on federal data be publicly released?

• Will their weights or code be available?

• Will independent researchers be able to reproduce the results?

• Can a private company receive exclusive rights?

• Can generated data be kept secret?

• Will the government collect royalties?

• Will discoveries be licensed at reasonable terms?

• Can the public use the resulting technology without paying the same companies that helped build it?

• What happens when proprietary models become essential to operating federal research infrastructure?

Without those answers, Genesis could create a public-private black box.

Public assets go in.

Commercial products come out.

The contracts decide who is allowed to see what happened in between.

Veterans are part of the research infrastructure

The most immediate human question involves veterans.

The Million Veteran Program is one of the world’s largest research collections combining genetic information, health records, lifestyle data, military experiences and environmental exposures.

More than one million veterans have enrolled. VA says investigators and other approved partners can apply for access to the program’s data. (VA Research)

VA and the Department of Energy have already worked together through a program known as MVP-CHAMPION, which combines veterans’ health information, genomic data, high-performance computing and AI tools for precision-medicine research. (MVP)

That history matters because Genesis is not beginning from an empty server rack.

The infrastructure and partnerships already exist.

VA privacy documents say Million Veteran Program participants provided informed consent and HIPAA authorization for research use of their information. The genomic system storing those records is designed to give approved researchers access inside controlled computing environments. (U.S. Department of Veterans Affairs)

Those safeguards are meaningful.

They do not settle every question raised by a much larger AI platform.

Veterans should be told, in plain English:

• Whether their records will be used to train Genesis-linked models

• Whether private companies will help build or operate those models

• Whether commercial products may result

• Whether derived models can leave VA-controlled environments

• Whether veterans can withdraw from future uses

• Whether genetic and medical information will be combined with other federal datasets

• Whether independent auditors will test reidentification risks

• Whether veterans receive any direct benefit when their data contributes to commercially valuable discoveries

The government may conclude that existing consent documents authorize this work.

That legal conclusion would not eliminate the obligation to explain it.

A signature collected years before Genesis existed should not be treated as a blank check for every future AI partnership Washington can imagine.

This is exactly the institutional problem examined in Zero Veterans Asked: the government tends to describe veterans as beneficiaries of technology while treating them as passive inputs into the systems being built around them.

Veterans are not merely a dataset.

They are the people whose bodies, illnesses, service histories and DNA make the research possible.

Washington is becoming the investor, regulator and laboratory partner

The Genesis Mission fits into a broader change in the federal government’s relationship with private industry.

Washington is no longer content to fund basic research, establish regulations and purchase finished products.

It is increasingly acting as:

• Capital provider

• Equity holder

• Infrastructure operator

• Data supplier

• Research partner

• Customer

• Market maker

• Regulator

• Export-control authority

Those roles do not always align.

A government that owns an equity stake in a company may later regulate its competitors.

An agency that provides proprietary access to public data may later evaluate products built from that data.

A laboratory that helps create a commercial technology may also influence which standards the technology must meet.

A government that becomes dependent on one company’s model may be less willing to scrutinize that company aggressively.

Genesis intensifies those conflicts because it creates a shared platform rather than a single isolated investment.

The administration is also moving quickly to lock the approach in across the federal research enterprise. Under the accompanying strategy report, agencies with at least $3 billion in research authority are required to submit implementation plans within 90 days. (Reuters)

Ninety days is a fast clock for a governance question this large.

In Nobody Verified the $5 Trillion AI Bet, I examined how enormous AI investment projections were repeated without adequately testing the assumptions underneath them.

Genesis presents a related problem.

Washington is promising accelerated discovery across some of the most complex fields in science. But the administration has not yet provided enough public detail to evaluate the financial architecture, data governance or ownership structure beneath that promise.

The phrase “AI for science” is doing a great deal of work.

So is the word “partnership.”

Those words sound collaborative and forward-looking.

They can also conceal who controls the asset when the partnership ends.

The public should not have to buy its own discoveries back

The argument for Genesis is compelling.

Scientific discovery is slow.

Federal research systems are fragmented.

Important datasets sit inside disconnected agencies.

Laboratories often lack the software and computing infrastructure needed to use modern AI effectively.

Researchers spend enormous amounts of time cleaning data, reproducing experiments and navigating administrative barriers.

A well-designed national platform could reduce duplication, accelerate medical breakthroughs, improve energy systems and strengthen American scientific leadership.

That is the promise.

The accountability question is not whether the government should use AI for science.

It should.

The question is whether the public receives a fair return for everything it contributes.

That return cannot be measured only by the number of press releases announcing breakthroughs.

It must include:

• Transparent funding

• Publicly disclosed contracts

• Clear data-governance rules

• Independent privacy oversight

• Meaningful access for researchers

• Limits on exclusive licensing

• Government rights that translate into public benefit

• Affordable access to resulting medicines and technologies

• Protection against one vendor controlling essential scientific infrastructure

• Full disclosure of conflicts among agencies, companies and laboratory operators

There will be legitimate reasons to protect classified information, personal health records, trade secrets and early-stage inventions.

“Proprietary” cannot become a magic word that makes publicly financed research disappear from public scrutiny.

Washington is asking taxpayers to fund the machine.

It is asking federal scientists to build it.

It is asking agencies to feed decades of public research into it.

It may ask veterans to contribute some of the most personal information a human being possesses.

Before the machine begins generating patents, products and corporate fortunes, the ownership structure should be visible.

The administration says Genesis will transform American science.

It may.

But transformation is not accountability.

Taxpayers should not provide the money, data, laboratories and computing power only to discover that the most valuable results belong to whoever negotiated the best contract behind a closed door.

The public should not have to buy its own discoveries back.

Share

Editor’s note

The White House describes the Genesis Mission announcement as more than $5 billion in federal commitments across more than 15 agencies. FedScoop, reporting from the July 22 summit, put the number of participating agencies at 20. Public materials reviewed for this article do not provide one consolidated breakdown distinguishing new appropriations, existing programs, future funding opportunities, federal resources and private contributions.

The Department of Defense figures cited in this article, more than $200 million in fiscal 2026 and more than $1.3 billion in the following fiscal year, were stated by a DOD official at the summit and reported by FedScoop. Fiscal 2027 amounts described at the summit have not been appropriated by Congress.

The reported figure of more than $500 million in Genesis Mission Consortium contributions appeared in a single trade publication and could not be independently confirmed in official materials reviewed for this article. It is presented here as reported, not as verified.

The exact ownership and licensing terms for Genesis projects may vary according to the agency, award, laboratory and partnership agreement involved. This article does not claim that one ownership arrangement applies to every project. It identifies the absence of a clear, program-wide public explanation of those arrangements.

Reporting and source notes

• White House: More than $5 billion for the Genesis Mission

• Executive order launching the Genesis Mission (Executive Order 14363, November 2025)

• White House Genesis Mission fact sheet

• Science: A New Golden Age

• Department of Energy: $293 million for Genesis Mission challenges (March 17, 2026)

• DOE laboratory-partnering and intellectual-property rules

• NASA: NASA Joins Genesis Mission to Accelerate AI-Driven Discovery

• FedScoop: NASA, DOD and others join Energy Department-led Genesis Mission

• Reuters coverage of the Genesis Mission 2026 Summit

• VA Million Veteran Program

• VA: Million Veteran Program computing and research tools

• U.S. Code: Patent rights in federally funded inventions

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

The United States government would like its allies to believe two things at once.

First, Washington must be able to restrict access to the world’s most advanced artificial intelligence systems when national security is at stake.

Second, foreign governments and businesses should continue building critical infrastructure on American technology without worrying that Washington might restrict their access.

That sales pitch became considerably harder on June 12.

At 5:21 p.m. Eastern, according to Anthropic’s account of the government directive, the company was ordered to suspend access to Fable 5 and Mythos 5 for foreign nationals, including foreign nationals working inside Anthropic.

Anthropic said it had no reliable way to verify every user’s nationality in real time.

So it shut the models down for everyone.

Not only customers in China.

Not only people in sanctioned jurisdictions.

Everyone.

The restrictions were lifted later that month, and Anthropic began restoring access on July 1. But the interruption had already made Washington’s leverage impossible to ignore.

An American government directive had forced an American AI company to remove access to a frontier model on a global scale.

Now the State Department is reportedly trying to convince the rest of the world that what it watched happen was not really a kill switch.

Technically, that may be true.

Politically, it is almost beside the point.

The cable

Reuters reported Wednesday that Secretary of State Marco Rubio circulated a diplomatic cable on July 16 instructing American diplomats to push back against growing concern over a U.S. technology “kill switch.”

The cable did not name Anthropic or directly address the June shutdown. But Reuters reported that it supplied diplomats with talking points intended to counter the arguments that erupted after it.

“There is no government ‘magic button,’” the cable reportedly said.

That is a clever defense because it answers a narrower question than the one America’s allies are asking.

Nobody serious believes the secretary of state has a giant red button under his desk labeled TURN OFF THE AI.

The real question is whether the United States government has enough legal, commercial and technical leverage over American AI companies to make foreign access disappear when Washington decides the risk is too high.

June answered that question.

The government issued a directive.

Anthropic complied.

Customers lost access.

The fact that the process involved lawyers, cloud infrastructure, authentication systems and corporate compliance teams rather than one glowing button does not make the power imaginary.

It makes the power institutional.

Washington does not need a magic button

Modern AI systems are not products in the traditional sense.

When a foreign government buys an American aircraft, the plane does not disappear from the runway because Washington changes an access policy.

When a hospital buys medical equipment, the manufacturer generally cannot remove the machine from the building by revoking an API key.

Frontier AI is different.

Most advanced American models are delivered through accounts, subscriptions, APIs and cloud platforms. Users do not necessarily possess the model.

They possess permission to access it.

That permission can depend on:

• Account authorization

• Identity verification

• Citizenship or residency

• Geographic location

• Export-control screening

• Cloud-region availability

• API credentials

• Corporate policy

• Government directives

• Model-specific approval

A government does not need one universal switch when the entire product is constructed from layers of revocable permission.

Anthropic’s shutdown demonstrated how those layers interact.

The government did not need to seize Anthropic’s servers.

It did not need to nationalize the company.

It did not need to rewrite the model.

It issued a legal directive changing who was permitted to use the system.

Anthropic’s inability to distinguish foreign nationals from everyone else in real time transformed a targeted order into a global interruption.

The control was not located inside one button.

It was distributed across the system.

This is the same access problem I examined in Who Gets to Decide Which AI You Can Use?. A government does not always have to publish a sweeping ban. It can use procurement rules, export controls, cybersecurity advisories, hosting requirements and regulatory uncertainty to make a model commercially unusable.

David Sacks, a senior White House AI adviser, has warned that regulatory uncertainty itself can be weaponized against open-model competitors.

Sacks’s argument is not proof that any particular company orchestrated the Anthropic restriction or proposed controls on Chinese models.

It does identify the broader mechanism.

A system can remain technically legal while becoming too risky for companies, contractors, banks and cloud providers to touch.

The door stays open.

Institutional pressure makes walking through it irrational.

“They build it. It’s yours.”

The State Department cable reportedly instructed diplomats to promote American AI as the safest and most effective option for foreign partners while portraying sovereign alternatives as wasteful and inefficient.

One of its reported sales lines was remarkably simple:

“They build it. It’s yours.”

But that is precisely what the June episode called into question.

What does “yours” mean when access depends on a company headquartered in another country?

What does ownership mean when a foreign government can change the legal conditions under which the service is available?

What does technological sovereignty mean when the models, chips, cloud infrastructure, authentication systems and updates remain under someone else’s jurisdiction?

Foreign customers do not own Anthropic’s models merely because they pay to access them.

They do not control the company’s infrastructure.

They do not control American export policy.

They do not control which national-security authorities Washington might invoke.

They do not control whether a future restriction lasts 18 days or 18 months.

“They build it. It’s yours” is a compelling slogan.

It is not a governance framework.

Europe heard the message

European officials did not need the State Department to explain what happened.

Reuters quoted European Parliament member Christophe Grudler describing the incident as evidence that the United States holds a real switch over critical technologies. Another European lawmaker, Aura Salla, argued that Europe cannot keep constructing its technology stack on access that a foreign government can withdraw overnight.

That does not mean Europe can replace the American AI stack tomorrow.

European governments and companies remain deeply dependent on U.S. cloud providers, semiconductor suppliers and AI developers. Building competitive alternatives will be costly, slow and vulnerable to Europe’s long-standing habit of regulating industries faster than it creates them.

But the strategic concern is legitimate.

Sovereignty is not merely about where a server is physically located.

It is about who has final authority to decide whether the system continues operating.

Washington appears to view European AI and data-sovereignty initiatives primarily as commercial barriers and protectionist regulation. In February, Reuters reported on an earlier State Department cable directing diplomats to oppose foreign rules involving data localization and restrictions on American technology companies.

That position would be easier to defend if Washington had not just demonstrated why foreign governments are nervous about dependence.

You cannot insist that allies abandon sovereign alternatives while reserving the right to revoke the systems on which they depend.

At least not without expecting them to notice.

Access can fail without government action

The government is not the only gatekeeper.

Compute, hosting and commercial capacity can also decide who gets access.

Moonshot AI provided a useful demonstration after the launch of Kimi K3. Demand approached the limit of its computing clusters, and the company temporarily paused new consumer subscriptions to protect service for existing customers.

That was not censorship.

It was not an export control.

It was infrastructure scarcity.

But the user experienced the same practical result: access was unavailable.

This matters because the term “open” often collapses several separate questions into one vague promise.

A model can have downloadable weights and still require hardware that almost nobody can afford.

A model can be publicly documented and still lack available cloud capacity.

A model can be legal and still be rejected by procurement officers.

A model can be technically available and still be blocked by a provider, employer, insurer or government contractor.

Access is not one gate.

It is an entire stack.

The government’s two AI policies

The White House is pursuing two AI strategies simultaneously.

The first is expansion.

The administration wants American companies to dominate global AI markets, sell infrastructure abroad and become the technological foundation on which other countries build their economies.

The second is control.

The government also wants frontier models tested, monitored and restricted when officials believe the systems create cybersecurity or national-security risks.

A June 2 executive order created a voluntary framework under which frontier-model developers could provide the federal government early access for cybersecurity testing.

The White House said the framework was not a mandatory licensing or pre-clearance regime.

That limitation matters.

But the Anthropic episode demonstrates that the government may possess other authorities capable of producing the same practical outcome.

The debate should not be reduced to whether Washington is establishing a formal federal AI licensing bureau.

The more important questions are:

• Which agencies can restrict model access?

• Under which statutes or national-security authorities?

• What evidence must they provide?

• Can the company challenge the directive?

• Are affected customers entitled to notice?

• Are allies consulted before access is removed?

• How narrowly must restrictions be tailored?

• What happens when a company cannot technically comply without shutting access down more broadly?

• Who reviews whether the action was proportionate?

Anthropic said the government’s initial letter did not explain the specific national-security concern in detail.

That is the beginning of an accountability problem.

The safety argument is not fake

There was a genuine cybersecurity concern behind the government’s action.

Anthropic said officials were responding to a method of bypassing Fable 5’s safeguards and using the model to identify software vulnerabilities. The company argued that the demonstration involved previously known vulnerabilities and disputed whether it justified the breadth of the restriction.

Anthropic later developed additional safeguards and expanded coordination with government and industry partners before restoring access.

Reasonable people can disagree about whether the temporary restriction was justified.

That is not the same as pretending the authority does not exist.

The cybersecurity debate also contains an uncomfortable tradeoff. Safety systems that prevent offensive hacking assistance can sometimes interfere with legitimate defensive work.

Sacks highlighted one developer’s claim that Kimi K3 helped repair security vulnerabilities after American models refused because of cyber guardrails.

One social-media account is not a controlled benchmark. The result may have depended on prompting, deployment settings or the precise nature of the request.

But the underlying policy problem is real.

If defensive teams cannot obtain useful assistance while attackers use less restricted models, safety controls can shift capability toward the wrong side.

The answer is not to eliminate every guardrail.

It is to define the threat, test the safeguards and prove that restrictions are precise enough to reduce harm without crippling legitimate security work.

Access is the new sovereignty

The first phase of the global AI competition was about who could build the most capable model.

The next phase is about who can reliably access it.

A model can lead every benchmark and still be strategically useless to a government, hospital, energy provider or financial institution if access can be removed without warning.

That changes how foreign customers will evaluate American AI.

They will not ask only:

• Is this the best model?

• Is it affordable?

• Is it secure?

They will also ask:

• Who controls my access?

• Under whose laws?

• What happens during a political dispute?

• Can the service be revoked?

• Do I have a credible alternative?

Those questions will shape procurement, infrastructure investment and international alliances.

China understands this.

Europe is beginning to understand it.

American diplomats have apparently been instructed to argue against it.

But no amount of diplomatic wording can undo the lesson of June 12.

Washington may not possess a single magic button capable of shutting down every American AI product.

It does not need one.

It has laws.

It has export controls.

It has jurisdiction over the companies.

It has influence over cloud and semiconductor infrastructure.

It has procurement power.

And it has already demonstrated that when those layers move together, access can disappear.

Call it a kill switch or do not.

The rest of the world saw the lights go out.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Editor’s note

The Control Grid has not independently obtained the July 16 State Department cable. Descriptions and excerpts from the cable are attributed to Reuters, which reported that it reviewed the document. The June government directive and resulting suspension are documented in Anthropic’s public statements.

Everything in this series so far has been about money — who owns what, who paid what, and which rules got bent to make it happen.

This part isn’t about money. It’s about the referee putting on a jersey.

The problem in one sentence

A government that owns a piece of OpenAI has to decide how hard to regulate OpenAI.

That’s it. That’s the conflict, and you don’t need a finance degree to feel how wrong it is. The same government that decides which AI models are safe enough to ship, which get restricted to “approved partners,” which get export-controlled, and which get switched off entirely would also hold shares that lose value every time it says no.

This isn’t a hypothetical I cooked up for dramatic effect. It’s the specific warning coming from the people who watch this for a living. Public Knowledge, a consumer-tech watchdog, put it plainly: the government would be shareholder and regulator at the same time, and we should not want a situation where it becomes less willing to enforce safety rules because doing so would cut the value of its own investment. A PitchBook analyst was blunter about what OpenAI is actually buying — he called it a “regulatory insurance policy.” One strategist framed the whole move as an attempt to acquire the “permission layer”: the invisible ceiling of political and regulatory approval that decides what any AI company is allowed to ship. OpenAI, he argued, isn’t navigating that layer. It’s trying to buy it.

We spent Part 1 and Part 2 watching Washington reach into companies and take. Here’s the first company reaching back — and what it wants in return isn’t cash. It’s a regulator with skin in the game.

Why AI makes this worse than chips

A government that owns 9.9 percent of Intel has a mild incentive problem. Maybe it roots a little harder for domestic chips. Annoying, not dangerous.

A government that owns 5 percent of a frontier AI lab is a different animal, because of what it controls at the model layer. Remember the sequence from Part 1: this past June, Anthropic’s most advanced models went dark worldwide for most of a month under the first export controls ever applied to an AI model instead of a chip. Around the same window, OpenAI delayed GPT-5.6’s rollout at the government’s request — reportedly after Commerce Secretary Howard Lutnick warned Altman against shipping it without sign-off — and the equity story broke just days later. The government already holds the switch that decides which models live, ship, or wait. The timing is not proof of anything. It’s just close enough that you should notice it.

Now hand that same government equity in one of the companies whose models it’s ranking. Every export-control decision, every “approved partners” carve-out, every national-security AI contract, every safety-review timeline carries an implicit tilt toward the lab Washington owns a piece of. As one analysis put it, OpenAI’s competitors — Anthropic, Google, xAI — would no longer just be competing on model quality. They’d be competing against a company that partially nationalized its own regulatory risk.

Let me be careful here, because this is exactly the kind of claim that needs discipline. There is no evidence that any decision has been tilted. No proven quid pro quo. What exists is a structure that makes the tilt possible and profitable, installed at the one company whose products the government most directly controls. You don’t need corruption for this to be dangerous. You need incentives — and the incentives are now pointed the wrong way by design.

And it’s worth saying plainly, in fairness: Anthropic isn’t part of this. A person familiar with the matter confirmed Anthropic is not having equity conversations with the administration. A 5 percent non-voting stake is also not “nationalization,” whatever the hotter headlines claim — it’s a financial interest, not a takeover. The danger here isn’t a government seizing AI. It’s a government quietly acquiring a reason to go easy on it.

The number nobody’s arguing about

Here’s the detail that tells you the real fight is already over.

When OpenAI offers 5 percent, that number only sounds generous until you see what else is on the table. In June, Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act — a one-time 50 percent tax, paid not in cash but in stock, on every major AI company. The shares would go into a federally managed fund his office pegs at $7 trillion, run by an “Independent Commission for Democratic AI” that would hold voting shares and use them to block company decisions, cutting every American a dividend check his office estimates north of $1,000 a year. Sanders put the thesis on X himself:

Fifty percent, with voting control, versus Altman’s 5 percent with none. Against that backdrop, the OpenAI offer looks less like a gift and more like a counteroffer. (Worth noting the obvious hitch in Sanders’s math: OpenAI and Anthropic aren’t profitable, so that $7 trillion and those $1,000 checks are projections, not deposits. The principle is the story here, not the payout.)

And the administration wants in. Aboard Air Force One, the President mused that AI wealth-sharing “almost becomes a partnership with the American public” — and, he added, “it would make ‘em rich.” The Vice President has said Trump prefers equity over cash. Put it together and notice what’s missing from the entire conversation: anyone arguing the government should own zero percent of the AI industry.

That’s the tell. Washington isn’t debating whether it should be a shareholder in the companies building artificial intelligence. It’s haggling over the percentage. The democratic socialist wants half. The AI CEO offers a twentieth. The President wants “as much as I can get.” The question of principle — should the referee own the team at all — got skipped while everyone argued about chatbots.

The part that outlives the headline

Here’s why this matters past this news cycle, past this administration, past whichever party you’d trust or distrust with it.

Machines get inherited. The equity playbook from Part 2 — grants into ownership, licenses into revenue cuts, approvals into veto rights — is now a proven, repeatable tool sitting on the shelf in the Oval Office. Whoever wins the next election inherits it fully assembled. If you cheered it when your side built it, picture the other side holding it. If you feared it under this President, understand it doesn’t disarm when the next one takes over. That’s the nature of precedent: it’s a loaded weapon left for the next person, and you don’t get to choose who that is.

We were told, for years, that the danger was Big Tech capturing the regulators — buying influence, writing the rules, winning by lobbying. That danger was real. But we watched for the wrong maneuver. The move that actually landed wasn’t industry capturing the state.

It was the state buying the team, picking up the whistle, and telling us not to worry about which way it blows the calls.

The government isn’t regulating AI. It’s negotiating its cut. And once the referee owns a piece of the winner, the only people without a stake in the outcome are the ones in the stands.

That’s you. That’s all of us.

Freedom isn’t a feature. It’s non-negotiable.

Adapt and overcome.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Reporting & Sources

This investigation is based on official statements, legislative materials, company announcements, public documents, and reporting from Reuters, the Financial Times, NOTUS, and other primary news organizations. Analysis and conclusions are those of The Control Grid unless otherwise attributed.

In Part 1, we documented the portfolio: chips, rare earths, lithium, a golden share in a steel company, and a cut of certain AI-processor sales to China.

Different industries. Different legal authorities. Different mechanisms.

The destination was the same.

That raised a more important question than what Washington acquired.

How did it acquire it?

The law built a funding program. The deal created a shareholder.

Start with the statute, because this is where the story stops being a debate over industrial policy and becomes a fight over how much authority Congress handed the executive branch.

The CHIPS Act’s semiconductor incentives program gave the Commerce Department broad power to award federal financial assistance. The law allowed Commerce to determine the amount and “funding type” of each award, restricted how the money could be used, required target dates for projects, and directed the department to recover funds when recipients missed those deadlines.

When Commerce launched the program in 2023, it described the available awards as direct funding, federal loans, and federal guarantees of third-party loans. It also promised extensive due diligence, milestone-based disbursements, clawbacks, and upside sharing for recipients receiving more than $150 million in direct funding.

That was the public bargain: taxpayer money in exchange for factories, domestic capacity, workforce commitments, and enforceable milestones.

Then Commerce reached for a different part of the law.

Under 15 U.S.C. § 4659, the Secretary of Commerce may enter contracts, grants, cooperative agreements, and “other transactions” considered necessary to carry out the department’s CHIPS responsibilities.

That phrase did a remarkable amount of work.

Intel’s own implementing amendment says the original Direct Funding Agreement was entered under both the semiconductor incentives statute and the “other transactions” authority. Commerce then used that framework to restructure the arrangement around common stock and warrants.

According to Intel’s August 22, 2025 SEC filing, the federal government agreed to provide $8.8698 billion in disbursements. That included $5.695 billion in accelerated funding under Intel’s Direct Funding Agreement and $3.1748 billion connected to the Secure Enclave program. In return, Intel agreed to issue hundreds of millions of common shares and warrants to the Commerce Department.

Other major CHIPS recipients were handled differently. Samsung’s final award was described as up to $4.745 billion in direct funding, distributed as the company completed project milestones. GlobalFoundries received up to $1.5 billion in direct funding under the same milestone-based model.

Intel’s transaction was not simply another grant with a clever press release. It accelerated billions in unpaid public funding and exchanged securities for the money.

The administration’s legal theory is not imaginary. The “other transactions” language is sitting there in federal law, broad enough to invite exactly this kind of interpretation.

But that does not make the question settled.

Intel itself warned investors that the legislative, judicial, or executive branches could later determine that all or part of the transaction was “unauthorized, void or voidable.”

That may be the most important sentence in the filing.

Washington did not return to Congress and ask for a new equity program. It interpreted existing authority, amended an existing contract, and made the federal government a shareholder.

No new law.

No new vote.

A new result.

The strings didn’t loosen. The contract deleted them.

The original CHIPS framework came with conditions for a reason.

Commerce said awards would be tied to construction, production, and commercial milestones. Recipients receiving more than $150 million in direct funding would have to share a portion of excess returns with the government. The department promised due diligence, reporting, workforce plans, restrictions on eligible uses, and clawbacks when companies failed to meet their commitments.

Then came the Intel amendment.

Intel’s SEC filing says Commerce agreed that, “to the maximum extent permissible under applicable law,” Intel’s obligations under the Direct Funding Agreement would be considered discharged, except for Secure Enclave. Commerce also agreed to work with Intel to release the company from certain obligations.

The signed amendment shows what that meant.

It deleted the definition of “Project Completion Requirements.”

It replaced the original upside-sharing section with a single word: “Reserved.”

It removed the original conditions precedent to each disbursement.

It deleted most of the affirmative covenants, negative covenants, default provisions, and remedies.

It replaced the program-requirements annex, the Davis-Bacon requirements, the reporting annex, the workforce-funding annex, and the disbursement milestone schedule with the same word:

“Reserved.”

Not every protection disappeared. That distinction matters.

The amended agreement retained restrictions on eligible uses of the money, protections involving foreign entities of concern, the project-completion clawback, national-security guardrails, and certain worker and community commitments required by the statute. It also created a narrower procedure governing the release of the accelerated funds.

But ownership did not simply arrive on top of the original bargain.

The bargain itself was rewritten.

Senator Elizabeth Warren, citing the same filings, accused the administration of handing billions to Intel while “asking for nothing in return.” That is her political judgment, not a neutral description of every surviving term. But her underlying concern is supported by the paperwork: several major safeguards attached to the original award were removed as part of the equity transaction.

The irony is that Commerce had already created a way for taxpayers to participate in unexpected financial upside. The original CHIPS program required large direct-funding recipients to share a portion of returns that exceeded their projections.

And government equity was not some forbidden thought nobody in Congress had considered. In 2022, Senator Bernie Sanders filed an amendment that would have required semiconductor companies receiving assistance to issue warrants or equity to the federal government, along with additional restrictions involving stock buybacks, outsourcing, collective bargaining, and union neutrality.

Congress did not include that across-the-board equity requirement in the law it enacted.

Three years later, Commerce reached a similar destination through a different door, without the same package of conditions Sanders proposed.

That is not the same deal with better taxpayer upside.

It is a different deal: equity added, cash accelerated, several safeguards deleted, and some core restrictions left standing.

This wasn’t a partisan objection.

If you are a conservative reader waiting for the inevitable lecture about a Republican administration, save yourself the suspense.

Some of the sharpest criticism came from Republicans.

Senator Rand Paul asked:

“If socialism is government owning the means of production, wouldn’t the government owning part of Intel be a step toward socialism? Terrible idea.”

Representative Warren Davidson called the arrangement “state-sponsored capitalism” and warned that “America will not outperform China by being more like China.”

Senator Thom Tillis told CBS that even a small government stake started to feel like a “semi-state owned enterprise, à la CCCP.”

Free-market critics raised a different but related concern. Once Washington owns part of a company, investment and regulatory decisions can become entangled with the government’s financial interest in that company. Cato Institute analysts warned that state ownership can distort competition, politicize investment, and create incentives to protect favored firms when they stumble.

Paul later made the contradiction even harder to ignore:

“If you’re going to criticize the socialist Mamdani … you better criticize Republicans who want a share of Intel, of Nvidia, or U.S. Steel.”

Meanwhile, Sanders supported the basic principle that taxpayers should receive a return when public money creates private value. Warren’s objection focused less on the existence of equity than on the terms of this particular transaction and the protections removed from the original award.

That is the actual political map.

The dividing line is not cleanly red or blue. It runs between people who believe government ownership can protect taxpayers and strategic industries, and people who believe turning regulators into shareholders creates a new class of political and economic conflicts.

Both arguments deserve to be taken seriously.

Pretending only one of them exists is how public policy becomes branding with a flag behind it.

The precedent matters more than Intel.

One deal can be explained away.

A portfolio is a governing philosophy with a balance sheet.

Lay the mechanisms side by side.

Intel: unpaid CHIPS funding converted into common stock and warrants.

Nvidia and AMD: export licenses conditioned on the government receiving 15 percent of certain chip-sales revenue from China.

U.S. Steel: merger approval paired with a national-security agreement and a government golden share carrying veto rights.

Lithium Americas: a federal loan restructured to include warrants representing 5 percent ownership in the company and another 5 percent in its joint venture with General Motors.

Different agencies. Different industries. Different authorities. Different paperwork.

Every deal used a different door.

Every deal ended in the same room.

The administration has not exactly hidden its appetite for more.

After the Intel transaction, the President said he would make deals like it “all day long.” National Economic Council Director Kevin Hassett said he expected more transactions, possibly in other industries. Commerce Secretary Howard Lutnick said the government was considering stakes in defense contractors, including companies for which Washington is already the dominant customer.

That last category exposes the central problem.

The government does not enter these negotiations as an ordinary investor. It can approve mergers, issue export licenses, award contracts, restructure loans, regulate markets, impose national-security conditions, and decide which companies receive public money.

Then it can ask for a piece.

The argument for taxpayer upside is not frivolous. When public money rescues or materially strengthens a private company, warrants or equity can sometimes protect the public from absorbing all the risk while shareholders keep all the gains. The Department of Energy notes that warrants have been used in previous federal loan arrangements, including its 2010 loan to Tesla.

But that does not resolve the institutional questions.

Who decides what the stake is worth?

Who determines which companies surrender equity and which receive ordinary funding?

What happens when the government must regulate a company whose stock it owns?

Does Washington protect the market, or protect its investment?

Do competitors receive equal treatment, or does the federal portfolio quietly become a list of favored firms?

And when these deals are built from different statutes, agencies, waivers, licenses, loans, and national-security authorities, where is the common framework governing valuation, voting rights, conflicts of interest, disclosure, oversight, and exit?

Right now, there is no single public answer.

There is a portfolio.

There is a playbook.

The rulebook is still being written while the deals are being signed.

The next company may open the door itself.

Every transaction in this series so far began with Washington holding leverage.

Funding.

Licenses.

Merger approval.

Loans.

Contracts.

National security.

The next case may be different.

The Financial Times reported in July that OpenAI had discussed giving the federal government a 5 percent stake. Reuters reported that the proposal could extend to other major American AI companies through a public-wealth structure modeled partly on the Alaska Permanent Fund. The talks were preliminary, no deal had been announced, and Reuters said it could not independently verify the Financial Times report.

That uncertainty matters.

So does the proposal.

If it advances, the question will no longer be only what Washington can demand from companies receiving public support or regulatory permission.

It will be what companies may volunteer when government alignment becomes commercially valuable.

A regulator that owns part of the company it regulates is not merely a regulator.

A company that offers equity to the government is not merely seeking capital.

The relationship changes.

So do the incentives.

Part 1 documented the portfolio.

Part 2 documented the playbook.

Part 3 asks what happens when the company opens the door.

That is where this investigation stops being about industrial policy.

It becomes a story about power.

Freedom isn’t a feature.

It’s non-negotiable.

Adapt and overcome.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

For twenty years, Silicon Valley’s favorite bedtime story was that the government couldn’t run a lemonade stand.

Turns out Washington can run a portfolio. On Intel, at least, it has done very well.

In August 2025, the federal government converted $5.7 billion in unpaid CHIPS Act grants and $3.2 billion in Secure Enclave funding into an $8.9 billion stock purchase. Washington received 433.3 million Intel shares at $20.47 apiece, equal to a 9.9 percent stake. Counting $2.2 billion in grants Intel had already received, the government’s total commitment reached $11.1 billion.

The stake is passive. The government has no board seat. It also received a warrant for roughly another 5 percent, exercisable only if Intel’s ownership of its foundry business falls below 51 percent.

At Intel’s July 21 price of $105.45, the government’s existing position was worth about $45.7 billion. That is more than five times the $8.9 billion stock purchase and just over four times the full $11.1 billion commitment.

Read that again. Washington restructured money already awarded to Intel and turned it into ownership. Less than a year later, the position was sitting on tens of billions of dollars in paper gains.

Congress voted for the CHIPS Act. It did not vote on this specific conversion of Intel’s unpaid awards into a federal equity position. That structure arrived through an agreement between Intel and the Commerce Department.

There was a statute. Then there was a term sheet. They are not the same thing.

The portfolio nobody asked for

Intel is not a one-off. Line up the transactions and a pattern emerges.

In July 2025, the Defense Department invested $400 million in newly issued convertible preferred stock in MP Materials and received a warrant. Together, the preferred shares and warrant represented about 15 percent of the company on an as-converted and as-exercised basis, positioning the Pentagon to become the largest shareholder in America’s leading integrated rare-earth producer.

The Energy Department took warrants equal to 5 percent of Lithium Americas and another 5 percent of its Thacker Pass joint venture with General Motors while restructuring a $2.26 billion federal loan.

The federal government also authorized a proposed $35.6 million strategic investment in Trilogy Metals under the Defense Production Act. But this transaction needs an asterisk the size of a congressional hearing room: as of July 21, it had not closed. Trilogy said the parties were targeting July 31 while final documents were completed.

Washington also extracted a golden share in U.S. Steel as part of the Nippon Steel acquisition. The arrangement gives the federal government specified rights involving governance, domestic production and trade matters.

Then there is the chip-export arrangement.

Reuters reported in August 2025 that Nvidia and AMD agreed to share 15 percent of revenue from certain China chip sales in exchange for export licenses. Nvidia’s later securities filing used more cautious language: federal officials expected 15 percent or more of licensed revenue, but no regulation had codified the requirement.

That distinction matters. So does the mechanism. Export permission was being tied to a federal claim on private revenue without a generally applicable tax law establishing the charge.

This is not a random walk through the stock market.

Chips. Rare earths. Lithium. Steel. Revenue from AI processors headed overseas.

The federal government has assembled financial and governance claims across the AI supply chain, layer by layer, from the dirt to the silicon.

Government ownership is not unprecedented. The Troubled Asset Relief Program was authorized by Congress during the 2008 financial crisis. Airlines issued warrants as partial compensation for assistance provided under pandemic-relief statutes, and Treasury later auctioned those warrants.

Those comparisons are imperfect, but they share an important feature: Congress created programs for named national crises and expressly authorized the financial instruments involved.

The current portfolio is being assembled deal by deal under different authorities. Grants become equity. Loans produce warrants. Export licenses create revenue expectations. Merger approval produces governance rights.

Different lever every time. Same direction every time.

You’ve read this story here before

Regular readers know we have been pulling this thread for months. We covered the export-control arc and the emerging equity proposal when it first surfaced:

The Bubble and the Backlash, Part 3

On June 12, Anthropic’s most advanced models, Fable 5 and Mythos 5, went offline after a U.S. export-control directive restricted foreign access. Because Anthropic could not reliably verify nationality across its systems in real time, the company suspended access more broadly. The models returned after roughly three weeks under revised controls.

OpenAI then delayed GPT-5.6’s full public rollout at the government’s request, limiting early access to vetted partners while officials reviewed national-security risks. OpenAI also warned against turning government approval of customers into a permanent release model.

The precedent is now visible: Washington can influence when a frontier model ships and who receives early access.

Then the ownership proposal arrived.

In early July, Reuters reported that Sam Altman had discussed giving the federal government a 5 percent stake in OpenAI. At the company’s reported $852 billion valuation, that would be worth roughly $42.6 billion. The proposed structure would ask other leading American AI companies to contribute similar stakes to a public vehicle modeled on the Alaska Permanent Fund.

Other companies had not agreed. The administration and Anthropic reportedly had not even discussed a government stake in Anthropic. So the industry-wide fund is not an industry agreement. It is an early proposal from one company that assumes other companies may eventually participate.

The proposal may require an act of Congress. It may never happen.

That uncertainty belongs in the story, because inevitability is not evidence.

The 5 percent number looks modest only beside the other proposal in Washington. Senator Bernie Sanders proposed a one-time 50 percent stock tax on the largest AI companies to create a sovereign wealth fund with voting authority. His office estimated the fund could reach $7 trillion and eventually support annual public dividends.

The democratic socialist wants half. The AI company is discussing a twentieth. The debate is no longer confined to whether AI companies should be regulated or taxed. Washington is now openly debating whether the public should own part of them, how much it should own and who would control the voting rights.

That conversation advanced while most people were still arguing about whether chatbots write acceptable emails.

OpenAI’s proposal is structurally different from Intel’s deal. Intel received billions in federal funding in exchange for shares. OpenAI, as reported, would contribute equity to a public vehicle without receiving an equivalent cash investment.

Why?

The proposal could reflect a genuine policy belief, an attempt to share AI wealth, political risk management or some combination of all three. The reporting does not establish Altman’s motive, and there is no evidence of a quid pro quo.

The timeline still matters. Anthropic’s models were disabled in June. OpenAI’s rollout was delayed later that month. The 5 percent proposal surfaced in early July. Those events do not prove an exchange. They do show how closely model access, political relationships and potential government ownership are beginning to overlap.

The referee wants a jersey

Here is the part that should bother you regardless of which cable channel raised you.

If the government accepts an OpenAI stake, it gains a financial interest while also weighing whether OpenAI models should be delayed, restricted or released.

That does not mean one agency would control every decision. It means the same federal government could act as regulator, national-security gatekeeper and shareholder in a company affected by those decisions.

We spent a decade warning that Big Tech might capture the regulators.

Nobody warned us the regulators might buy the team.

From the left, the concern is that a powerful executive could use equity to purchase political goodwill. From the right, the concern is that the state is acquiring ownership and control rights through executive-branch bargaining rather than a transparent, generally applicable framework.

Both concerns are legitimate. Neither is proven by the timeline alone.

That is what makes this a series instead of a hot take.

What’s coming

Part 1 is the inventory: what Washington owns, what it may soon own and which transactions are still incomplete.

Part 2 will examine the acquisition toolkit. How did grants, loans, export licenses and merger approvals become mechanisms for taking equity, warrants, revenue and veto rights? Did the Intel conversion preserve the taxpayer safeguards Congress originally attached to the CHIPS Act?

Part 3 will examine the conflict. What happens when the government regulating frontier AI also owns part of the companies being regulated? What can one administration build that the next administration will inherit?

Washington is no longer only regulating AI.

It is negotiating ownership, revenue and control rights across the stack.

Freedom isn’t a feature. It’s non-negotiable.

Adapt and overcome.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Most people think this story is about a Chinese AI company.It isn’t.

They think it’s about benchmark scores, billion-dollar valuations, or the latest chapter in the technology race between Washington and Beijing.

It isn’t.

It’s about something far more important.

Who gets to decide which artificial intelligence you’re allowed to use.

Because once governments have the ability to shape which models businesses can deploy, developers can build on, researchers can access, and consumers can trust, artificial intelligence stops being just another technology race.

It becomes a question of power.

Moonshot AI’s release of Kimi K3 didn’t create that question.

It exposed it.

The Question Nobody Voted On

Nobody voted on which artificial intelligence models Americans should be allowed to use.

Congress has not passed a law banning China’s newest frontier AI systems.

The President has not signed an executive order prohibiting them.

No federal agency has announced that downloading Moonshot AI’s Kimi K3 is illegal.

Yet that doesn’t mean the outcome is undecided.

Governments rarely need an outright ban to reshape a market.

They can influence procurement rules, export controls, cloud hosting requirements, cybersecurity guidance, liability standards, and federal contracting. Each decision appears narrow on its own. Together, they can determine which technologies become practical to adopt and which quietly disappear from serious commercial use.

That’s why this story matters.

At first glance, it looks like another chapter in the technology race between the United States and China.

A Chinese AI company released an impressive new model.

Markets reacted.

Policymakers took notice.

Another week in the AI industry.

Look closer, and a different question emerges.

Who gets to decide which AI systems Americans are allowed to build on, deploy, and trust?

Moonshot AI is not the story.

It is the first major test case.

The answer to that question extends far beyond one company or one model.

It reaches into the future of software development, scientific research, enterprise technology, cybersecurity, education, and the broader question of whether access to frontier AI will be shaped primarily by open competition, transparent public policy, or quiet administrative decisions that never receive a congressional vote.

This week, that debate moved from theory into public view.

In an investigation by Axios, Maria Curi reported that officials inside the Trump administration have discussed multiple approaches for restricting advanced Chinese AI models, including possible Entity List designations, government security advisories, procurement restrictions, hosting requirements, and other measures intended to reduce their adoption in the United States. According to Axios, similar proposals had previously stalled amid concerns about limiting innovation before returning to the policy agenda as Chinese models became increasingly competitive.

No final policy has been announced.

No executive order has been signed.

No nationwide prohibition exists today.

Those facts are important.

So is what they don’t tell us.

The most consequential technology policy decisions are often made long before the public notices their effects.

A model doesn’t have to become illegal to become commercially impractical.

Banks, defense contractors, healthcare providers, cloud companies, insurers, universities, and publicly traded corporations routinely adjust their behavior in response to procurement rules, cybersecurity guidance, liability concerns, and government recommendations long before a courtroom ever decides the issue.

That’s why the central question isn’t whether Kimi K3 has already been banned.

The central question is whether governments should have the power to shape access to frontier artificial intelligence through regulatory pressure before the public ever has a chance to debate that decision openly.

Kimi K3 is simply the first major test of that question.

The answer will almost certainly influence far more than one Chinese AI model.

It may help define how the next generation of artificial intelligence reaches the rest of the world.

This is where I think we can noticeably improve the article.

Your reporting is solid. The issue is that it still reads like a “What We Know” memo instead of an investigation. We can keep every fact while giving it narrative momentum.

I’d also change the heading. Instead of “What We Know,” make the reader ask a question.

How One AI Model Started a Much Bigger Fight

Before Washington could debate whether Kimi K3 should be restricted, one question had to be answered.

Was it actually good enough to matter?

On July 16, Beijing-based Moonshot AI announced Kimi K3, describing it as a 2.8-trillion-parameter, natively multimodal mixture-of-experts model with a one-million-token context window designed for coding, research, knowledge work, and long-running agentic tasks. According to Moonshot’s technical announcement, the model routes each token through 16 of its 896 experts, meaning the widely reported 2.8-trillion parameter figure reflects the model’s total capacity rather than the amount of the network used during every response.

That distinction matters.

Parameter count alone does not determine intelligence, reasoning ability, inference speed, or operating cost. In modern AI systems, architecture often matters as much as size.

Even so, Kimi K3 immediately attracted attention because its early benchmark performance suggested it belonged in conversations previously dominated by American laboratories.

On July 16, K3 reached the preliminary top position on Arena’s WebDev Leaderboard, a benchmark based on blind user comparisons of models performing front-end development tasks. Its preliminary rating of 1,679 placed it ahead of Claude Fable 5 and GPT-5.6 Sol on that specific evaluation.

Specific is the operative word.

The result demonstrated exceptional performance in one category of software development. It did not establish that Kimi K3 had surpassed every American model across mathematics, scientific reasoning, multimodal understanding, enterprise reliability, or general intelligence.

Moonshot itself makes no such claim.

Independent testing tells a similarly nuanced story.

According to Artificial Analysis, Kimi K3 currently earns an Intelligence Index score of 57, placing it alongside Claude Opus 4.8 and GPT-5.5, while still trailing the newest flagship systems, including Claude Fable 5 and GPT-5.6 Sol.

Taken together, the available evidence supports a straightforward conclusion.

Kimi K3 is a near-frontier AI model.

The evidence does not support declaring it the world’s most capable model.

That distinction has largely disappeared in the online conversation.

In today’s AI race, benchmark charts often become marketing campaigns. Every leaderboard becomes ammunition for supporters and critics alike, while the caveats attached to those results quietly disappear.

There’s another important qualification.

Despite widespread descriptions of Kimi K3 as an open-weight model, its full model weights had not yet been publicly released at the time of publication. Moonshot says it intends to release those weights, along with additional technical documentation, by July 27.

For now, Kimi K3 is available through Moonshot’s applications and API.

That may sound like a technical distinction.

It isn’t.

An AI model that is available through a company’s own services is fundamentally different from one whose weights have been released for developers, researchers, and enterprises to run independently. Until those weights are publicly available, Kimi K3 remains a model with a promised open-weight release, not a fully released open-weight model.

That nuance is easy to miss.

It also matters because much of the political debate surrounding Kimi assumes a level of openness that, at least for now, has yet to arrive.

And that raises the next question.

If Kimi K3 is a credible competitor, what tools does Washington actually have to limit its adoption without passing a law banning it?

Why Silicon Valley Is Paying Attention

Frontier AI isn’t won by building the smartest model.

It’s won by convincing millions of developers and thousands of businesses that your model delivers the best value.

That’s where Kimi K3 becomes far more than another benchmark story.

It becomes an economic one.

According to Moonshot AI’s published API pricing, Kimi K3 costs $3 per million uncached input tokens and $15 per million output tokens.

For comparison, Anthropic lists Claude Fable 5 at $10 per million input tokens and $50 per million output tokens. Claude Opus 4.8 is priced at $5 and $25, respectively.

On paper, Kimi is substantially less expensive than Anthropic’s flagship models.

That pricing matters.

Developers building commercial applications don’t simply ask which model performs best. They ask which model delivers acceptable performance at a cost their customers will actually pay.

Price becomes a feature.

But it isn’t the whole story.

Independent analysis suggests the cost advantage is meaningful without being revolutionary.

According to Artificial Analysis, Kimi K3 required approximately $0.94 to complete an Intelligence Index evaluation. That placed it close to GPT-5.6 Sol, estimated at $1.04, roughly half the cost of Claude Opus 4.8, but still more expensive than Chinese competitors including GLM-5.2 and DeepSeek V4 Pro.

In other words, Kimi is putting meaningful pressure on premium American pricing.

It is not the cheapest advanced AI model on the market.

That distinction matters because much of the public discussion has framed Kimi as a model that suddenly made every American competitor economically obsolete.

The evidence doesn’t support that conclusion.

What the evidence does support is something arguably more significant.

Kimi combines competitive frontier-level performance, lower pricing than several premium American systems, a one-million-token context window, and a planned open-weight release.

Individually, none of those features changes the industry.

Together, they create a far more compelling competitive package.

That combination is what makes companies pay attention.

Demand Quickly Met Reality

Strong pricing attracts customers.

Infrastructure determines whether they can actually use the product.

Within days of Kimi K3’s launch, Moonshot AI temporarily stopped accepting new consumer subscriptions, explaining that demand had exceeded its forecasts and approached the limits of the company’s available computing capacity.

According to Reuters and the Associated Press, Moonshot prioritized existing paid subscribers while adding additional infrastructure and announced plans to separate its general Kimi memberships from coding-focused subscription tiers.

Moonshot confirmed the pause publicly in a post on X.

That announcement tells us something important.

Demand exceeded the company’s available infrastructure.

It does not tell us how many customers joined, how much additional computing capacity will ultimately be required, or whether that demand represents lasting adoption or launch-week enthusiasm.

Those are very different questions.

The episode also highlights a contradiction that extends well beyond Moonshot AI.

Much of the conversation surrounding frontier AI treats “open” models as though they automatically democratize access.

In practice, the software is only one part of the equation.

Moonshot recommends deploying Kimi K3 on supernode configurations containing at least 64 accelerators.

That is not hardware most companies keep sitting in a server room.

Even after the model weights are publicly released, most organizations will still depend on hyperscale cloud providers, specialized AI infrastructure companies, or managed hosting services to run systems of this scale.

The model itself may become increasingly open.

The infrastructure required to operate it remains concentrated.

That distinction matters because the debate unfolding in Washington isn’t only about software.

It’s also about who controls the infrastructure that modern AI increasingly depends on.

And that brings us to the next question.

If governments don’t need to ban a model outright, what tools do they already have to influence who can realistically use it?

Markets Are Pricing the Possibility, Not the Outcome

Before Washington announced a policy...

Before Moonshot released its promised model weights...

Before independent security reviews had been completed...

Financial markets had already started reacting.

Kimi K3 entered the conversation during an existing selloff in semiconductor stocks, but its release added fuel to a growing debate over one of the biggest assumptions driving the AI boom:

Will the hundreds of billions of dollars being poured into American AI infrastructure generate the returns investors expect?

According to Reuters, semiconductor shares fell for a third consecutive trading session as investors reduced exposure to AI-related companies. While the decline was already underway, Kimi K3’s arrival intensified concerns that increasingly capable Chinese models could challenge the economics supporting massive infrastructure spending.

The Wall Street Journal reached a similar conclusion, describing Kimi as adding fresh pressure to chip stocks while raising new questions about competition, pricing, and whether American AI companies can continue justifying unprecedented capital expenditures.

That’s an important distinction.

Kimi K3 did not trigger the selloff.

It became evidence supporting an anxiety investors already had.

If frontier-level AI capabilities become increasingly affordable, the economic assumptions behind today’s infrastructure race become harder to take for granted.

That doesn’t mean those assumptions are wrong.

It means investors have started asking tougher questions.

The Biggest Questions Still Don’t Have Answers

Despite the headlines, many of the most important facts remain unknown.

Moonshot has promised to release Kimi K3’s model weights by July 27.

As of publication, that has not happened.

We do not yet know what license will govern those weights, how freely they can be modified or redistributed, or whether commercial users will face meaningful restrictions.

We also have not seen the full technical report Moonshot says it intends to publish.

Those missing details matter.

They will determine whether Kimi becomes a genuinely open alternative for developers or remains primarily a service delivered through Moonshot’s own infrastructure.

The same uncertainty exists on the security side.

There has been no publicly released U.S. government security assessment of Kimi K3.

There is no public evidence that the model contains hidden backdoors, undisclosed communications capabilities, politically manipulated behavior beyond what has been observed in other frontier AI systems, or security vulnerabilities unique to Kimi itself.

That absence of evidence should not be confused with evidence of safety.

It simply means the public record has not yet established those claims.

The policy picture remains equally incomplete.

We do not know whether the Trump administration has selected a specific course of action.

We do not know whether an executive order is being drafted.

We do not know which legal authorities, if any, officials ultimately intend to rely upon.

Those are not small omissions.

They are central to understanding where this story goes next.

One more claim also deserves careful treatment.

Axios, citing an unnamed source, reported that leading AI laboratories or their allies have periodically encouraged the administration to consider restrictions on open models.

That reporting is newsworthy.

It is not evidence that any specific company, including OpenAI or Anthropic, requested restrictions targeting Kimi K3.

The public evidence does not currently support making that claim.

The Debate Has Already Escaped Washington

Even before any formal policy has been announced, the public fight over Kimi has already begun.

Among the most vocal participants is David Sacks, co-chair of the President’s Council of Advisors on Science and Technology and one of the administration’s most influential voices on artificial intelligence.

Sacks argues that some of the industry’s largest closed-model companies are attempting to use government policy to protect themselves from open competition. More importantly, he warns that governments don’t need an outright ban to reshape the market. Regulatory guidance, procurement rules, and institutional risk can accomplish much of the same outcome.

Read his full statement below:

Security or Protectionism? The Hardest Question in This Investigation

One of the biggest mistakes in the public debate over Kimi K3 is assuming there are only two possible explanations.

Either the United States is trying to protect national security.

Or American AI companies are trying to eliminate a lower-cost competitor.

The public evidence supports neither conclusion by itself.

It suggests elements of both deserve scrutiny.

That distinction matters because once every argument becomes political, evidence stops mattering.

This investigation is about evidence.

The Cyber-Guardrail Debate

One example illustrates why the conversation has become so contentious.

White House AI adviser David Sacks highlighted a report claiming Kimi K3 successfully helped remediate 15 software vulnerabilities after OpenAI’s Codex and Anthropic’s Claude Fable allegedly refused because of cybersecurity guardrails.

Read Sacks’ post below.

If accurate, that example raises an important policy question.

Could well-intentioned safety controls unintentionally reduce the ability of defenders to secure vulnerable systems while attackers increasingly rely on models operating under fewer restrictions?

That’s a legitimate concern.

It is not, however, a controlled scientific comparison.

A single publicly shared remediation example cannot establish that one model is categorically better than another.

Prompt wording matters.

Context matters.

Model versions matter.

Deployment settings matter.

Safety policies matter.

Without standardized testing, none of those variables can be isolated.

That doesn’t make the concern invalid.

It means the evidence should be treated proportionately.

The challenge isn’t deciding whether AI models should have guardrails.

It’s determining whether those guardrails successfully distinguish legitimate defensive cybersecurity work from assistance that enables offensive abuse.

That question deserves far more testing than political rhetoric.

There Isn’t One Kimi. There Are Three.

Much of the policy discussion treats “using Kimi” as though everyone is talking about the same thing.

They aren’t.

There are at least three fundamentally different deployment models, each carrying different security implications.

1. Moonshot’s API

In the simplest arrangement, customers send prompts directly to Moonshot’s infrastructure and receive responses through the company’s API.

Moonshot states that API prompts and outputs are not used for model training and are not persistently stored for that purpose, according to its published data policy.

That is an important representation.

It is not an independent audit.

2. Third-Party Hosting

A second option allows an American cloud provider or specialized AI infrastructure company to host Kimi on behalf of customers.

In that scenario, the intermediary can implement its own data residency requirements, monitoring, logging, compliance controls, and security policies.

The customer may never communicate directly with Moonshot at all.

3. Self-Hosting

The third scenario becomes possible once Moonshot releases Kimi’s model weights.

Organizations with sufficient computing resources can deploy the model entirely within their own infrastructure.

Prompts never leave the organization’s environment.

Nothing is transmitted back to Moonshot.

From a security perspective, this arrangement is fundamentally different from sending sensitive information to a foreign-operated API.

Treating those three deployment models as identical oversimplifies the policy challenge.

If Washington’s concern is Americans transmitting sensitive government or corporate information to a Chinese company, restrictions on API use in sensitive environments may be justified.

That concern does not automatically extend to a self-hosted model operating entirely inside an American company’s own infrastructure.

If the concern is that Kimi’s weights themselves contain vulnerabilities, hidden communications, or embedded manipulation, then the appropriate response is rigorous technical evaluation.

Not assumption.

Nationality alone does not identify the mechanism of harm.

The National Security Argument Is Real

Recognizing those distinctions should not minimize legitimate security concerns.

Artificial intelligence increasingly processes source code, engineering documentation, proprietary research, internal communications, legal strategy, and operational planning.

Sending that information to any external provider creates risk.

Foreign jurisdiction introduces additional legal, intelligence, and supply-chain considerations.

Open-weight models create another set of challenges.

Once weights are released publicly, they can be modified, redistributed, fine-tuned, and deployed without the original developer’s oversight.

A March 2026 report from NIST concluded that open-weight systems become significantly more difficult to monitor after release and effectively impossible to recall once broadly distributed.

The federal government also has reason to examine Chinese frontier models carefully.

In 2025, NIST’s Center for AI Standards and Innovation (CAISI) evaluated downloaded DeepSeek models and found they were more susceptible to prompt hijacking and jailbreaking than the American models included in that evaluation.

Researchers also identified censorship behaviors aligned with Chinese Communist Party narratives embedded within the downloaded model weights themselves, rather than imposed solely through DeepSeek’s online service.

Those findings deserve attention.

They also require restraint.

DeepSeek is not Moonshot.

Evidence that one Chinese model exhibited specific security characteristics does not establish that Kimi K3 shares those same characteristics.

It establishes something narrower.

It establishes a compelling reason to test Kimi.

It does not establish that Kimi has already failed those tests.

That distinction is precisely what evidence-first reporting requires.

If security concerns alone don’t determine the outcome...

What authority does Washington actually have to act?

That question turns out to be far more complicated than most headlines suggest.

What Washington Can Actually Do

One of the biggest misconceptions in the debate over Kimi K3 is that Washington faces only two options.

Ban the model.

Or do nothing.

The federal government has far more tools than that.

Some are explicit.

Others are indirect.

Several have already been used against foreign technology companies in entirely different industries.

Each carries different legal authority, different practical consequences, and different implications for the future of artificial intelligence.

Federal Procurement Restrictions

The narrowest option would involve federal procurement.

The administration could prohibit federal agencies from purchasing or deploying specific Chinese AI models inside government systems. Depending on how such rules were written, the restrictions could also extend to certain government contractors handling sensitive federal work.

That would be significant.

It would not amount to a nationwide ban.

Private companies, universities, researchers, and individual consumers could still legally use those models unless additional restrictions were imposed.

Procurement policy determines what the government buys.

It does not automatically determine what the rest of the country is allowed to use.

The Entity List

Another possibility receiving public attention is the Commerce Department’s Entity List.

Companies placed on the Entity List face significant restrictions on receiving certain American exports, reexports, and transfers of controlled technology.

For Moonshot AI, that could affect future access to advanced American chips, specialized software, cloud services, and technical assistance.

Those consequences would be substantial.

They are also frequently misunderstood.

According to the Bureau of Industry and Security’s own guidance, an Entity List designation does not automatically prohibit Americans from purchasing products or services from a listed company.

In other words, placing Moonshot on the Entity List would not necessarily make downloading or using Kimi K3 illegal.

It would primarily restrict what American companies can provide to Moonshot, not what American users can obtain from Moonshot.

That distinction rarely survives social media.

It matters.

The ICTS Authority

The Commerce Department also possesses a more direct legal instrument.

Under its Information and Communications Technology and Services (ICTS) authority, Commerce can investigate, prohibit, or impose conditions on technology transactions involving foreign adversaries when officials determine those transactions present an undue or unacceptable national security risk.

This authority is not theoretical.

Commerce has previously used it against Kaspersky Lab software and certain connected-vehicle technologies linked to China.

As of publication, however, Commerce has not announced a comparable determination involving Moonshot AI or Kimi K3.

That distinction is critical.

The authority exists.

Its application to Kimi remains hypothetical.

Hosting Requirements and Liability

Government action doesn’t always target the AI model itself.

Sometimes it targets the companies that make deployment possible.

The administration could attempt to require American cloud providers or AI hosting companies to satisfy additional security standards before offering Chinese frontier models.

It could also increase liability for security failures involving those systems.

Whether such measures would withstand legal scrutiny would depend on the specific statutory authority, technical standards, and regulatory language used.

But the practical effect could arrive long before any courtroom ruling.

When legal obligations become uncertain, companies often choose the lowest-risk option.

That option may simply be declining to host the model at all.

The Power of Soft Law

Perhaps the most consequential tool requires no new regulation.

No executive order.

No Act of Congress.

No formal prohibition.

Government agencies routinely issue cybersecurity advisories, supply-chain guidance, intelligence warnings, and risk assessments.

Many of those documents are technically nonbinding.

In practice, they often reshape entire markets.

Banks answer to regulators.

Insurance companies evaluate exposure.

Defense contractors protect future contracts.

Public corporations answer to boards, shareholders, and compliance officers.

When enough institutions conclude that using a technology creates unnecessary regulatory uncertainty, behavior changes even when the law does not.

The model remains available.

The market quietly moves on.

That is the dynamic David Sacks described as “soft law.”

Whether readers agree with his characterization or not, the underlying mechanism is real.

Regulatory pressure often influences corporate decision-making long before formal legal prohibitions appear.

And that may be the most important distinction in this entire investigation.

The debate isn’t simply about whether Washington can ban Kimi K3.

It’s about whether Washington can make adopting Kimi K3 sufficiently risky that much of the market decides not to use it anyway.

Those are not the same thing.

But from the perspective of a developer trying to choose an AI platform, the practical result could look remarkably similar.

That brings us back to the question that opened this investigation.

If the government possesses multiple legal pathways to influence access to frontier AI, who ultimately decides where legitimate national security ends and protection from competition begins?

That question still doesn’t have an easy answer.

But the evidence points to one unavoidable conclusion.

This debate is no longer just about Moonshot AI.

It’s about the rules that will govern every frontier AI model that comes after it.

The Real Competition Isn’t Just Artificial Intelligence

If Kimi K3 exposed anything, it wasn’t simply that another Chinese laboratory had built a capable frontier AI model.

It exposed that the United States and China increasingly view artificial intelligence as infrastructure.

Not just software.

Not just products.

Infrastructure.

The same way previous generations competed over railroads, oil, semiconductors, and the internet, today’s competition is increasingly about who builds the platforms the rest of the world depends on.

That helps explain why the political reaction to Kimi extended well beyond benchmark scores.

It also explains why the debate has become so difficult.

Open AI Is Now Industrial Policy

The timing wasn’t subtle.

As Moonshot introduced Kimi K3, Chinese President Xi Jinping used the World Artificial Intelligence Conference in Shanghai to promote open-source AI as a global public good while positioning China as an alternative to American influence over AI infrastructure and technical standards.

That framing is strategic.

Not charitable.

China benefits when developers, companies, universities, and governments build on Chinese models instead of American ones.

Open-weight AI can expand China’s technological influence even in countries that would never adopt Chinese cloud platforms or consumer applications directly.

Washington understands exactly the same dynamic.

The White House’s 2025 AI Action Plan describes open-weight models as strategically important for American research, commercial competitiveness, and geopolitical leadership. It directs federal agencies to strengthen domestic AI development and improve access to computing resources for startups and researchers capable of advancing American frontier systems.

In other words, both governments recognize that open models can become instruments of national influence.

The difference is largely whose models they want the world to adopt.

That creates an uncomfortable contradiction.

The United States celebrates openness when American companies lead.

The policy becomes considerably more complicated when openness strengthens a Chinese competitor capable of challenging American pricing, market share, and technological influence.

Incentives Are Not Evidence

One of the easiest mistakes in technology reporting is assuming incentives automatically prove misconduct.

They don’t.

But ignoring incentives produces equally incomplete reporting.

American frontier laboratories have obvious financial reasons to support policies that make lower-cost foreign competitors more difficult to adopt.

Their valuations depend, in part, on maintaining enough technological scarcity to justify enormous investments in chips, data centers, and infrastructure.

That observation does not establish that any company improperly influenced government policy.

The evidence simply doesn’t support that conclusion.

Moonshot has incentives too.

Strong benchmark results attract customers.

Lower prices attract developers.

Rapid adoption attracts investors.

According to Reuters, Moonshot is pursuing additional funding while preparing for a possible public listing in Hong Kong.

Washington has incentives.

Protecting sensitive government information.

Maintaining domestic technological leadership.

Reducing dependence on foreign AI infrastructure.

Beijing has incentives.

Reducing American influence over chips, cloud platforms, AI infrastructure, and global technical standards.

Developers have incentives.

Businesses have incentives.

Universities have incentives.

Every participant in this story can make a legitimate argument.

Every participant also benefits from a particular outcome.

That is precisely why evidence must carry more weight than motivation.

Incentives help explain behavior.

They do not establish facts.

"The next phase of the AI race won't be decided solely by who builds the strongest model. It will also be shaped by who writes the rules governing which models the world is allowed to use."

The Timeline

July 23, 2025
The White House releases America’s AI Action Plan, supporting American open-weight development while directing government agencies to evaluate frontier AI models, including potential foreign security risks.

September 2025
NIST’s Center for AI Standards and Innovation (CAISI) publishes its evaluation of downloaded DeepSeek models, identifying increased susceptibility to prompt hijacking and embedded censorship aligned with Chinese Communist Party narratives.

July 16, 2026
Moonshot AI announces Kimi K3, describing it as a 2.8-trillion-parameter mixture-of-experts model and stating that its full model weights will be released by July 27.

July 17, 2026
Kimi K3 reaches the preliminary top position on Arena’s WebDev Leaderboard, bringing widespread attention to its coding capabilities.

July 17, 2026
Chinese President Xi Jinping promotes open-source AI during the World Artificial Intelligence Conference, presenting China as an alternative leader in global AI development.

July 19, 2026
White House adviser David Sacks publicly argues that leading closed-model laboratories are attempting to use government policy to reduce open-model competition.

July 20, 2026
Reuters and the Associated Press report that Moonshot temporarily paused new consumer subscriptions after demand exceeded available computing capacity.

July 20, 2026
Axios reveals internal Trump administration discussions involving possible Entity List actions, hosting requirements, procurement restrictions, and other policy options affecting advanced Chinese AI models.

July 27, 2026 (Expected)
Moonshot’s stated target date for releasing Kimi K3’s model weights and additional technical documentation.

By itself, Kimi K3 may or may not reshape the global AI market.

That isn’t the most important question.

The more consequential question is whether governments, companies, and institutions are quietly establishing the rules that will determine which frontier AI systems the world is allowed to build on in the first place.

That debate has already begun.

It will almost certainly outlast Kimi K3 itself.

The Control Grid Assessment

The purpose of this investigation was never to determine whether Moonshot AI is good or bad, whether China should be trusted, or whether Washington should impose restrictions.

It was to answer a narrower question.

What does the evidence actually support today?

What the Evidence Supports

The evidence supports the conclusion that Kimi K3 is a legitimate near-frontier AI model, not merely another headline or marketing campaign.

It supports the conclusion that Kimi is placing meaningful competitive pressure on premium American AI providers through a combination of strong performance, lower published pricing, and the promise of an open-weight release.

The evidence also supports that officials inside the Trump administration have discussed multiple policy mechanisms that could restrict or discourage the use of advanced Chinese AI models. Those discussions have been reported publicly, but no final policy has been announced.

The evidence further supports that the United States has several existing legal and regulatory tools capable of reshaping the AI market without imposing a nationwide ban.

Finally, the evidence supports a broader conclusion.

Artificial intelligence has become industrial policy.

Both Washington and Beijing increasingly view frontier AI as strategic infrastructure capable of shaping economic competitiveness, national security, and geopolitical influence.

What the Evidence Does Not Support

The public record does not establish that Kimi K3 contains hidden backdoors, unique security vulnerabilities, or politically manipulated behaviors beyond concerns that require model-specific testing.

The evidence also does not establish that OpenAI, Anthropic, or any other American laboratory requested restrictions targeting Moonshot AI specifically.

Likewise, there is no public evidence that the Trump administration has selected a final policy or determined which legal authority, if any, it intends to use.

Those remain open questions.

What We’re Watching Next

Several developments will determine whether today’s debate becomes tomorrow’s policy.

First, whether Moonshot releases Kimi K3’s model weights on July 27, and under what license.

Until that happens, K3’s openness remains a stated intention rather than a completed release.

Second, whether CAISI, NIST, or another credible independent laboratory conducts a comprehensive security evaluation of Kimi K3.

Claims involving hidden backdoors, censorship, or unique security risks should be supported by model-specific evidence, not assumptions based solely on nationality.

Third, whether the Federal Register, the Commerce Department, or the White House publishes actual legal or regulatory language.

Internal policy discussions matter.

They are not the same as government action.

Fourth, whether American cloud providers and AI hosting companies quietly begin changing their terms of service, compliance requirements, or hosting policies.

History suggests markets often respond to regulatory expectations before regulations themselves are finalized.

Finally, watch what Washington builds.

If the United States responds to Chinese competition by investing in competitive American open-weight models while protecting genuinely sensitive systems, it can make a coherent industrial policy argument.

If it responds primarily by making foreign competitors harder to use while leaving businesses dependent on a small number of expensive domestic platforms, arguments about regulatory capture will become increasingly difficult to dismiss.

Final Thought

Kimi K3 may ultimately prove to be a breakthrough.

It may prove to be overhyped.

Or it may settle somewhere between those two extremes.

Its long-term technical significance remains uncertain.

Its political significance does not.

The debate surrounding Kimi has already exposed something much larger than a single AI model.

The next phase of the artificial intelligence race will not be decided solely by who builds the most capable systems.

It will also be shaped by who writes the rules that determine which systems the rest of the world is permitted to use.

That debate has already begun.

And unlike Kimi K3 itself, it isn’t scheduled for release on July 27.

About This Investigation

This investigation distinguishes between confirmed facts, public reporting, and analysis. Every factual claim is supported by publicly available sources linked throughout the article. Where evidence remains incomplete or disputed, those uncertainties are identified explicitly.

If materially relevant new information emerges, this investigation will be updated transparently.

The Control Grid

Follow the evidence. Understand the system.

This investigation will be updated if materially relevant new information becomes available. Significant corrections or factual updates will be disclosed transparently.

Disclosure: Eric Mitchell serves as Chief Marketing Officer of Sacred Loop AI. Sacred Loop AI is not affiliated with Moonshot AI, OpenAI, Anthropic, Google, Meta, or any company discussed in this investigation. This reporting reflects independent editorial analysis based on publicly available information and linked primary sources.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Strip away the noise and the Anthropic-Pentagon fight comes down to one demand and two refusals.

The demand, delivered by the administration in late February: let the military use Claude for “all lawful purposes,” no exceptions. The refusals: Anthropic would not permit mass surveillance of American citizens, and would not allow fully autonomous weapons without a human in the decision chain. Two red lines. That’s the entire dispute. Not whether the military could use AI. Whether there was anything it couldn’t use it for.

When the Friday deadline passed with those lines intact, Trump ordered every federal agency to stop using Claude. Defense Secretary Pete Hegseth then formally designated Anthropic a supply chain risk, a label historically reserved for foreign adversaries like Huawei. An American company, blacklisted like an enemy telecom, for declining to write the government a blank check. A senior Pentagon official put the doctrine on the record: the military “will not allow a vendor to insert itself into the chain of command” by restricting lawful use. Anthropic’s CEO called the move “retaliatory and punitive” and sued, telling a court the designation is “unprecedented and unlawful”.

The war with Iran began within days. And here is the fact that detonates every official talking point: the Pentagon kept using Claude anyway.

Not a leak. Sworn-adjacent testimony. In a March Senate hearing, Pentagon CIO Kirsten Davies confirmed under questioning from Sen. Jack Reed that Claude was in active use in Operation Epic Fury, complete with a “measure of time” and an “exception process” for removal. Reed called it odd. Odd is one word for it. The Department of War designated a company an unacceptable supply chain risk and then ran its product through an active military campaign, because, as Pentagon staffers and contractors told Reuters, the alternatives weren’t as good and recertifying replacements could take 12 to 18 months. One contractor’s summary of the internal mood: career IT people at DOD hate the removal order.

So which is it? A company too dangerous for the government to touch, or a capability too valuable for the government to give up? It cannot be both. Unless the designation was never about risk at all, and was always about punishment for the refusal. You do not need me to editorialize. The government’s own behavior is the confession.

Meanwhile the machine ran hot. Retired Adm. Mark Montgomery described the military processing roughly a thousand potential targets a day and striking most of them, with next-strike turnaround potentially under four hours. This is the tempo AI enables, and it’s exactly why the red lines matter. Amnesty International says a strike on a school in Iran killed over 100 children, and Human Rights Watch has called for a war-crimes investigation; CBS News reported it could not verify whether Claude systems were used in that strike. We don’t know. That’s the point. At a thousand targets a day, “we don’t know” is the permanent condition, and the question of what the machine is allowed to refuse stops being philosophy.

The war fractured the administration from inside. National Counterterrorism Center Director Joe Kent, a retired Green Beret with eleven combat tours, resigned in March, writing that Iran “posed no imminent threat” and that he could not in good conscience support the war. Trump’s response was to call him weak. Whatever you think of Kent, and there is plenty to think, the highest-ranking counterterrorism official in the country resigned rather than co-sign the threat assessment. That happened.

Then came the bill. In June the White House sent Congress an $87.6 billion emergency supplemental, with $67.1 billion for the Pentagon, including $21 billion just to rebuild munitions stockpiles the war burned through. And buried in the fine print, a detail that should embarrass everyone: analysts note the supplemental includes no dedicated funding for the AI software the president’s own directives require, because that software is bought with operations money the war drained. (Disclosure on that source: it’s authored by Onebrief, a defense AI vendor arguing for funding it would benefit from. They disclose it; so do we.) The administration mandated an AI-first military, started a war that emptied the accounts that pay for it, and sent Congress an invoice that forgot the software. That’s not a strategy. That’s a keynote with a body count.

And while the government was pointing its threat-model at the American company that told it no, an actual adversary was already inside the wire. Last October, F5, the vendor whose BIG-IP gear manages traffic for federal networks and much of the Fortune 500, disclosed that a nation-state actor had maintained long-term, persistent access to its systems, stealing portions of BIG-IP source code and details on vulnerabilities F5 hadn’t yet fixed. The dwell time: roughly twelve months, undetected. Officially the intruder is just a “nation-state”; Bloomberg reported officials privately suspect a PRC-linked actor, which Reuters could not verify. CISA issued an emergency directive warning of an imminent threat to federal networks, and the stolen source code and vulnerability intel are still out there, a loaded weapon with no serial number. Hold both facts in your head at once: the designation reserved for foreign adversaries went to an American AI lab over a surveillance refusal, while a probable foreign adversary spent a year inside the plumbing of the federal government. That’s not a security posture. That’s a grudge with a budget.

The government’s answer arrived today. The White House announced GOLD EAGLE, a centralized clearinghouse where industry funnels cybersecurity vulnerabilities to the federal government for AI-powered triage and patching, run by the White House, Treasury, DHS, and the Department of War, and already operating. Let me be fair: after F5, faster coordinated patching is a real need, and on paper this is good policy. Now the questions the press release doesn’t answer. The announcement says it leverages “frontier AI capabilities” and quotes Hegseth touting “frontier AI alongside top American innovators”, the same Hegseth who designated the frontier AI company that set usage terms a supply chain risk. Which frontier AI, then? The release names no companies at all. And a national clearinghouse of unpatched vulnerabilities in American infrastructure is also, by definition, the most valuable exploit database on earth, held by a government with a documented history of sometimes keeping flaws for offensive use rather than disclosing them. Who audits GOLD EAGLE? Which vulnerabilities get patched, and which get kept? Nobody has answered, because as of this afternoon, nobody in Washington has asked. Consider it asked.

Four parts, one pattern. Meta allegedly turned surveillance on its own workers to feed the machine. Apple says OpenAI raided its vault to build hardware for an IPO. The industry and the government are negotiating who owns whom. And when one company drew two narrow lines around American civil liberties and human control of lethal force, the state reached for the adversary playbook, and kept the product.

The Silicon Valley myth is dead. What replaced it isn’t chaos. It’s consolidation. The machine, the money, and the state, closing ranks. The only thing still outside the perimeter is you.

Which is exactly where they’d like you to stay.

This concludes The Death of the Silicon Valley Myth. If you missed the earlier installments: Part 1 on Meta’s algorithmic layoffs, Part 2 on Apple v. OpenAI, Part 3 on the government’s equity play. For the full Anthropic recall-and-restoration saga that preceded this fight, read our Bubble and the Backlash series.

Freedom isn’t a feature. It’s non-negotiable. Adapt and overcome.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Bobby Axelrod said it best: what’s the point of having fuck-you money if you never say fuck you?

This week, Apple — owner of the largest pile of fuck-you money in corporate history — turned to its own AI partner and said it. Loudly. In federal court.

Which brings us to the cardinal rule OpenAI apparently never learned: you don’t steal from Godzilla. This is My Little Pony backing a moving truck up to the monster’s cave and hoping nobody checks the security footage.

Every job interview has awkward questions. “Where do you see yourself in five years?” “What’s your greatest weakness?” According to Apple, OpenAI’s version was different: “Did you bring the parts?”

On Friday, Apple filed suit in the Northern District of California accusing OpenAI — its own partner, the company that’s been living in Siri’s guest room since 2024 — of running a recruiting pipeline that doubled as a logistics operation for Apple’s hardware secrets. Batteries. Logic boards. Show-and-tell sessions. An offboarding cheat sheet for slipping past Apple security. This isn’t a lawsuit. It’s a heist movie where the getaway car has a $6.5 billion price tag.

What Actually Happened

Strip out the drama and here’s the filing: Apple sued OpenAI, its hardware subsidiary io Products, and two former Apple employees, Tang Tan and Chang Liu. Tan is OpenAI’s chief hardware officer and a longtime Apple VP who worked on iPhone and Apple Watch. Liu spent eight years at Apple as a senior systems electrical engineer. Apple wants an injunction barring OpenAI from using its trade secrets, the return of its confidential materials, and damages.

The complaint doesn’t whisper. It says OpenAI’s hardware business sits on the “shakiest of foundations, rotten to its core” from stolen Apple secrets, and claims the misconduct runs from technical staff all the way up to leadership. If you want to see the receipts, read Apple Inc. v. Chang Liu et al in the court filing PDF.

Everything that follows is Apple’s version of events, from Apple’s complaint, written by Apple’s lawyers. OpenAI’s entire public response so far is one sentence, delivered by tweet, about having no interest in other companies’ trade secrets. Nothing here has been tested in court.

The Greatest Hits

Apple’s investigation turned up a pattern that escalates from bold to cartoonish.

The complaint alleges Tan used Apple’s internal project codenames during interviews to pry more confidential information out of candidates who still worked at Apple. It alleges he told those candidates to bring “actual parts” to their interviews — batteries, logic boards, system-in-package units. Physical hardware, walked out of Cupertino for show-and-tell.

Apple claims a “Need to Know” offboarding document — the internal guide covering its exit security process — was circulated at OpenAI to teach departing Apple employees how to dodge that very process. At least one candidate was surprised by the bring-your-parts instruction.

The complaint also details how OpenAI allegedly approached manufacturing partners with insider terminology and claims Liu kept his Apple-issued laptop, exploited a security bug to download over a thousand pages of confidential files after departure, and joked about it (”LOL,” “so funny”).

Apple says they raised all this with OpenAI back in February.

OpenAI ghosted them like a cocky stripper who just got told she’s not going home with the whale.

Five months later, Apple finally gets a response… on OpenAI’s behalf.

Of course they did. The new hot shit in town thinks because he’s Trump’s bitch boy he can play big dog. Meanwhile he’s treating Apple — one of the few companies that actually matters — like a side chick booty call he can hit up whenever he feels like it.

Newsflash, geniuses at OpenAI: Apple and Elon aren’t the dudes falling in love with the stripper. You are. You’re getting used to make Trump even richer while you troll the actual big dogs and act like you run the yard.

That’s the energy Apple and Elon just brought to the chat.

Take a big step back and literally fuck your own face, boys. The adults in the room just reminded you how this actually works.

Who’s Not in the Complaint

OpenAI’s hardware effort is supposedly being led by Jony Ive — Apple’s legendary former design chief — as they chase these shiny new “AI-native” devices.

Cute.

Yet somehow none of the big design names made it into the filing. Apple’s lawyers went full John Wick on that one: surgical, silent, and left zero witnesses.

They didn’t just protect the IP — they sent a message.

Welcome to the big leagues, kids. The adults still know how to handle their business.

The Irony Layer — And the Godzilla Moment

Eighteen months ago these companies stood on stage together. Now Apple is suing its roommate while the lease is still active. More than 400 former Apple employees now work at OpenAI, according to Apple’s complaint. The pipeline was never the secret.

Elon Musk pulled up a chair and unloaded.

Musk has his own running feud with Altman and previously went after Apple over its OpenAI integration, so he’s not neutral — but damn if the timing isn’t perfect.

Why It Actually Matters

The injunction could freeze OpenAI’s hardware timeline cold by forcing redesign without Apple’s proprietary information.

The IPO just got a massive disclosure headache as investors price in litigation, discovery, and product delays.

Discovery gives Apple a legal window into OpenAI’s operations, recruiting practices, and hardware roadmap.

If you want the market-safe version of this story, Bloomberg’s explainer walks through the legal stakes.

For the “what else don’t we know?” angle, Firstpost’s “Tip Of The Iceberg” segment leans into the broader implications of Apple’s allegations.

The AI revolution sold itself as a clean break from old tech. Instead it inherited the lawsuits.

A courtroom will decide who stole what. But this filing paints a picture of an industry sprinting so fast that hardware allegedly walks out the door in interview bags.

Series Note: This is Part 1 of an ongoing series on the Apple vs OpenAI hardware heist. As new filings, statements, and investigations land, I’ll break them down here.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Five episodes ago, I made you a promise: everybody gets audited. The environmental alarmists got audited. The industry boosters got audited. The governors, the utilities, the trade associations, the think tanks with funding lines they’d rather you not read — audited.

Today, the last target is the one standing closest to me.

This morning, Jason Hubbard — founder and CEO of Sacred Loop, the company where I’m CMO — published “The Wrong Bet: The AI Bubble Nobody’s Watching.” It argues that the AI infrastructure buildout is the largest private financial bet in history, that it’s priced on an efficiency assumption nobody has stress-tested, and that a single software patch could turn it into something worse than 2008.

I could have amplified it. That’s literally my job title.

Instead, I ran it through the same wringer I’ve run everyone else through for four episodes. Because if “truth vs. fiction” means anything, it means the standard doesn’t bend when the byline belongs to your boss.

Disclosure, and it’s a big one: Jason is my CEO. His argument rests partly on his own research into AI runtime architecture — research conducted through Sacred Loop, as his own piece states. That means we have an interest in his thesis being taken seriously. You should weigh everything below with that in mind, the same way I told you to weigh PIMCO’s bubble warnings in Episode 4 knowing PIMCO sells bonds. Funding lines apply to friendly witnesses. This one’s friendlier than most.

So here’s the audit.

What survived

I fact-checked every load-bearing number in Jason’s piece against primary sources. Most of them held — some were actually conservative.

The capex is real. Alphabet, Amazon, Microsoft, and Meta plan roughly $700–725 billion in combined capital spending for 2026, up about 77% from roughly $410 billion in 2025, per their own earnings guidance. Goldman Sachs thinks the broader number lands closer to $920 billion this year and notes analyst estimates have run too low three years straight.

The trillions are real — but so is the spread. Jason cites $5.3 trillion. That’s a real number with a real name on it: Goldman Sachs Research’s estimate of hyperscaler capital spending from 2025 through 2030. But here’s what this series does that nobody else bothers with: the honest range. Morgan Stanley and Moody’s peg it at $3 trillion or more; JPMorgan says over $5 trillion including power; Goldman’s own supply-side model runs to roughly $7.6 trillion cumulative through 2031. The estimate spread is $4.6 trillion wide. When the professionals’ error bars are bigger than the GDP of Japan, anyone quoting you a single number is selling something.

The grid number is real. NERC’s Long-Term Reliability Assessment forecasts summer peak demand growing 224 gigawatts over ten years, with new data centers accounting for most of it. NERC’s own reliability chief says future electricity supply has never been more uncertain.

The debt is real. Alphabet, Amazon, Oracle, Meta and Microsoft issued about $121 billion in new bonds in 2025, versus $40 billion in 2020. By late 2025, AI-linked debt hit $1.2 trillion — surpassing US banks as the largest sector in the investment-grade index. Alphabet issued a 100-year bond. And the part Jason gets exactly right — the part that should bother you — is the ratings logic. Microsoft’s paper is rated pristine because Microsoft’s balance sheet is pristine. The rating measures the borrower. It does not measure the bet.

The rating reflects the balance sheet. It does not reflect the bet.

Bond managers are starting to say this out loud. Al Cattermole of Mirabaud Asset Management told CNBC that by moving AI capex into debt markets, “you now have the question of credit worthiness” — breaking what he called an unspoken contract that AI spending was equity risk, not bondholder risk. And some of the debt isn’t even where you’d look for it: Meta’s largest data center financing sits in a special purpose vehicle — off Meta’s balance sheet, payable through a long-term lease. Moody’s counts $969 billion in total hyperscaler commitments, two-thirds of it in leases that haven’t even started. Oracle’s credit default swaps more than tripled since September, and this week S&P downgraded it to the lowest investment-grade rung over its AI spending, while Amazon’s $25 billion offering drew a notably cold reception. The market is starting to price the bet, not just the borrower.

What we cut — and what got fixed

This is the accountability episode, so here’s the knife work — and here’s what it looks like when the standard actually works. I flagged five problems in Jason’s draft last week. The version he published fixed most of them: a viral “17 times the dot-com bubble” became a six-to-seven-times comparison with the math and named sources shown in his footnotes; an inflated bond claim became the verified $1.2 trillion figure; a secondhand Bernanke quote became a proper memoir citation. That’s not embarrassing. That’s what the standard is for. Most claims don’t survive contact with verification. The ones that do are the ones worth your attention.

One difference between our pieces remains, and you deserve to know exactly where it sits. Jason’s argument cites OpenAI’s costs and losses — figures sourced to leaked documents, which his footnotes disclose plainly as leaked and unaudited. This series doesn’t run leaked private financials at all, disclosed or not — a rule we set in Episode 4 and don’t bend for friendly bylines. Same facts available to both of us; different lines drawn. Now you know where each line is.

And the leaked-numbers rule matters more this week than most: Apple sued OpenAI Friday for trade secret theft — a suit that could complicate the most anticipated IPO in history. The company whose growth justifies the entire buildout is private, pre-IPO, and unauditable. Every number you’ve ever seen about its finances is a leak. That’s not a knock on OpenAI. It’s a fact about what you can verify — the only currency this series trades in.

The patch already dropped once

Jason’s doomsday scenario is a free, replicable efficiency breakthrough that makes every deployed GPU dramatically more productive, causing every CFO, utility board, and bond desk to ask the same question simultaneously.

His piece cites January 27, 2025 for the crash — and he’s right about the crash. A Chinese lab called DeepSeek released a model claiming frontier performance at a fraction of frontier cost, the claim was disputed, and it still didn’t matter: Nvidia lost $593 billion in market value in one day — the largest single-day loss in market history. A rumor did that. His question — what does a proof do? — is a fair one.

But the record has a second half, and you deserve both.

The market shrugged it off within days. Capex didn’t collapse — it accelerated, from $410 billion to $700-plus billion. The efficiency gain didn’t shrink demand for compute. It expanded it. Microsoft’s CEO called it in real time:

Former Intel CEO Pat Gelsinger said flatly that “the markets are getting it wrong” — cheaper computing expands the market, every time, going back to coal-fired steam engines. Economists call it Jevons paradox.

So case closed, right? Efficiency breakthroughs are bullish, the buildout is safe, Jason’s wrong?

Not so fast. That’s the collapsed-endpoint move from the other direction, and we don’t do that here either.

Why Jevons doesn’t end the argument

Jevons paradox rescues demand. It does not rescue the specific numbers in the contracts.

If AI compute gets 2–3x more efficient, total AI usage probably does grow — history backs Nadella. But the 20-year power purchase agreement a utility signed against a specific gigawatt load at a specific site? The utilization assumptions in a specific data center REIT’s bond covenant? Those weren’t priced on “demand in general.” They were priced on this facility, drawing this power, at these rates, for decades. Cattermole again, on exactly this: what happens to a five- or eight-year loan when the chips inside the collateral are obsolete by year three?

And DeepSeek’s numbers were contested from day one — analysts like Wedbush’s Dan Ives called the cost claims deeply misleading and probably untrue, which gave the market permission to move on. A verified, replicable, run-it-yourself efficiency gain is a different animal. January 2025 proved the market can absorb an efficiency scare. It did not prove the market can absorb a confirmed one. Nobody knows. That’s the honest statement — and $5-trillion-with-a-$4.6-trillion-error-bar is being deployed as if somebody does.

Meanwhile, the demand numbers are inflated too

Here’s the both-sides symmetry this series exists to expose. The doom case rides on unverified efficiency claims. The boom case rides on unverified demand claims — and those are provably padded.

Grid Strategies found utility forecasts likely overstate near-term data center load by roughly 25 GW, because developers shop the same project to multiple utilities and every utility books the full load. The 90 GW of data center growth utilities forecast by 2030? Market analysts say closer to 65 GW. One consultancy calculated that meeting the full 2030 projections would require 90% of global chip supply — a scenario it called unrealistic. And the canonical federal estimate — LBNL’s 325 to 580 TWh by 2028 — has a gap between its own low and high end bigger than most countries’ total consumption.

Phantom load gets rate hikes approved. Phantom efficiency gets panic clicks. Both are fictions, both are profitable, and both get billed to the same person: you.

The honest ledger

Equal scrutiny means the bulls get their best card on the table too. In Q1 2026, global AI revenue excluding China hit $25 billion, exceeding the estimated $21 billion in depreciation costs — for the second straight quarter. That’s the first real evidence the machine might pay for itself, and if you only read the doom coverage, you’ve never seen it.

Against it: capex consuming close to 100% of hyperscaler operating cash flow this year, versus a 10-year average of 40%, and Amazon projected to go free-cash-flow negative.

Revenue crossing depreciation is a pulse, not a payoff. Both things are true. Anyone telling you only one of them is lying by omission.

What it costs you

Here’s the close of this series, and it’s the same fact I opened Episode 1 with, now with five episodes of receipts behind it.

You never signed this trade. But you’re in it three times over. Your index fund holds these companies at the heaviest concentration in modern market history. Your bond fund — the “safe” sleeve — is now more exposed to AI issuers than to banks. And your utility bill is already paying for the grid buildout, with electricity prices rising 6.1% year over year — well ahead of overall inflation.

If the bet pays off, insiders captured the upside years ago. If it doesn’t, the write-downs come out of your 401(k), the stranded grid costs come out of your power bill, and the subsidies already came out of your taxes. Privatized upside, socialized downside — the throughline of this entire series, now running through the largest capital deployment in human history.

Even the principals know it. This weekend, Sam Altman jabbed that Elon Musk is “selling public market investors on short-term space datacenters” — days after SpaceX raised a record $75 billion partly on that pitch. When the two most famous men in AI accuse each other of selling unverified infrastructure bets to the public, believe both of them.

Is Jason right that a software patch brings it all down? I don’t know, and neither does he — his piece calls it a hypothesis, and I’ve held it to that. What I can tell you, verified: the assumptions are untested, the ranges are continental, the ratings measure the borrower and not the bet, and the people who priced it won’t be the people who pay for it.

What would change my mind — because a claim you can’t falsify is a religion, not a finding: three more consecutive quarters of AI revenue exceeding depreciation would gut the bubble case. A confirmed, independently replicated efficiency step-change absorbed by the market without contagion would gut the fragility case. Utilities eating stranded costs from shareholder returns instead of rate bases would gut the socialized-downside case. I’ll be watching all three, and I’ll print it if I’m wrong.

Truth is expensive. It cost this series its most viral numbers, cost me a comfortable Monday amplifying my own CEO, and costs you the comfort of a clean story with a villain. Fiction is cheaper right up until the bill arrives.

The bill always arrives.

Jason Hubbard’s companion piece, “The Wrong Bet,” is live now. Read it, then read his technical claims the way this series taught you to read everything: hypothesis until verified.

Previously in Data Centers: Truth vs. Fiction:

Episode 1: Where They Stop Counting
Episode 2: Debunking the Fiction of Fear
Episode 3: The Bill Comes Due
Episode 4: Debunking the Fiction of Progress

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

On July 6, a draft report leaked out of the Treasury Department. Its own career analysts — the people paid to watch for systemic risk — concluded the AI boom shares structural DNA with the dotcom bubble, except this time the companies are wired so deep into the economy that a stumble would send shockwaves through the entire economic ecosystem. Treasury’s response was not to rebut the analysis. It was to disown it as unvetted staff work and restate the official position: AI will drive America’s new Golden Age.

Eleven days before the leak, Treasury Secretary Scott Bessent stood up and praised the $750 billion buildout, asking whether we could do even more. So here is where Episode 4 starts: the government’s own analysts versus the government’s own mouth. You’ve seen this movie. In Episode 3, it was a nonbinding White House pledge versus a binding rate case. Now it’s a suppressed report versus a press release. Same asymmetry, one floor up.

Before we go further, read my funding line, because this episode is going to teach you to read everyone else’s. I’m the CMO of an AI company. I want this industry to win — my paycheck, my equity, and frankly my enthusiasm are all long AI. Which is exactly why you should hear me when I tell you the problem isn’t the technology or the demand. I’ve spent three episodes showing you the demand is real. The problem is that the money behind the buildout has been structured — carefully, legally, and in plain sight — so that if expectations slip, the losses land on your pension fund, your utility bill, and your state budget instead of on the people booking the profits. An industry insider telling you his own industry’s financing is engineered against you is a statement against interest. That’s the most reliable kind. Judge what follows by the documents, not by me — but you deserved to know where I stand before we start.

This is Episode 4 of Data Centers: Truth vs. Fiction. In Episode 1, “Where They Stop Counting,” I showed you how both sides collapse honest ranges into dishonest endpoints. In Episode 2, “Debunking the Fiction of Fear,” I dismantled the alarmist claims that don’t survive primary data. In Episode 3, “The Bill Comes Due,” I showed you who’s already paying. This episode asks the question both camps keep answering dishonestly: is the buildout progress, or a bubble? The boosters say the demand proves everything. The doomers say it’s all fake. Both are selling you an endpoint. Here’s the range.

What’s true: the demand is real, and I can prove it better than the boosters can

If I wanted to carry the industry’s water, the evidence is right there. CBRE’s Q1 2026 data shows global vacancy fell to 6.7% even as supply grew 25%, Northern Virginia hit an all-time-low 0.3%, and 80% of everything under construction in the top four U.S. markets is already preleased — years before the buildings exist. Companies are fighting over capacity that hasn’t been poured yet.

And the biggest bondholder in the space agrees. PIMCO — which holds roughly $18 billion of the debt behind Meta’s Louisiana megaproject, so read their funding line too — published its verdict in May: the AI capex cycle has genuine risks, but it’s “far more financeable” than the late-1990s telecom boom. Stronger balance sheets, real revenue, more discipline. That’s the honest case for progress, and it’s a real one.

So no — I’m not going to tell you the demand is fake. The doomers who do are running the same endpoint-collapsing con I documented in Episode 2, just pointed at a different target. Here’s what I am going to tell you.

What’s fiction: that real demand means nobody’s exposed

Follow the money in a circle. OpenAI has signed roughly $1.15 trillion in infrastructure commitments across seven vendors — Broadcom, Oracle, Microsoft, Nvidia, AMD, Amazon, CoreWeave. Nvidia invests in OpenAI, which commits hundreds of billions to Oracle, which buys Nvidia chips to serve the contract. Cash leaves as investment, comes back as revenue, and every company in the loop books growth from the same underlying dollar. In fairness, the industry’s defense comes from a named principal: Anthropic CEO Dario Amodei argued at DealBook that pairing a capital-rich chipmaker with a revenue-confident lab isn’t inappropriate in principle. Maybe. But even PIMCO — the friendly witness — concedes that circular financing is no guarantee of future demand, and that vendor financing is exactly how the telecom bubble made demand look more durable than it was. When your own creditor names the mechanism, it’s not a conspiracy theory. It’s a term sheet.

Now watch someone get asked about the math on camera. Last October, investor Brad Gerstner asked Sam Altman the question every pension manager should have asked first: how does a company with $13 billion in revenue — Friar’s own September number — make over a trillion dollars of commitments? Altman’s answer, verbatim: “if you want to sell your shares, I’ll find you a buyer.” Then: “Enough.” Watch it yourself, starting around 12:20.

That’s not an answer. That’s a CEO telling you the question is impolite.

The scale of the bet, from the companies’ own guidance: Amazon $200 billion, Microsoft tracking toward $190 billion, Alphabet $175–185 billion, Meta raised to $125–145 billion, Oracle around $50 billion — at 86% of Oracle’s revenue. Aggregate 2026 estimates run from PIMCO’s $690 billion consensus to Goldman’s $725 billion for just four companies to CreditSights’ $750 billion for five. Anyone who hands you one number picked their endpoint; the honest statement is roughly $650–750 billion, nearly double 2025, consuming — per PIMCO — 94% of these companies’ operating cash flow, up from 40% in 2023. Against that: the pure-play AI labs the buildout is nominally for generate somewhere around $50 billion in combined annualized revenue, and even that number is contested. Anthropic — the company behind Claude, so apply this episode’s skepticism to them too — announced a $30 billion run rate in April; OpenAI disputes the accounting and puts the comparable figure nearer $22 billion, while its own $25 billion figure is sourced to reporting Reuters couldn’t verify. Run rates, not audited revenue, from different dates. The range is the truth; the endpoints are marketing.

And because 94% of operating cash flow isn’t enough, the rest is borrowed — in ways designed not to look like borrowing. Tech companies have moved more than $120 billion of data center financing off their balance sheets in about eighteen months through special purpose vehicles. Meta’s Louisiana flagship sits in an SPV that’s 80% owned by Blue Owl, issued $27 billion in bonds rated A+ by exactly one agency, and pays 6.58% — a yield the market normally charges junk. Oracle runs a similar SPV stack across Texas, Wisconsin, and New Mexico. If the Wisconsin part sounds familiar, it should: same company, same state, same playbook as the “Green Chile Ventures” tax filing from Episode 3. And the creditors have noticed. In January, bondholders who bought $18 billion of Oracle notes sued in New York, alleging Oracle never disclosed it would need roughly $38 billion in new loans to serve its $300 billion OpenAI contract. The people lending the money say they weren’t told where the leverage was. Remember that the next time someone tells you the risk is well understood.

Then there’s the quietest lever of all: depreciation. When Michael Burry ended a two-year silence in November, this is what he came back for:

His argument: chips on a two-to-three-year product cycle shouldn’t produce longer useful-life assumptions, yet between 2020 and 2024 the hyperscalers stretched them — flattering earnings by his estimate of $176 billion through 2028. Is he right? Here’s the honest range. Amazon shortened the life of some servers in 2025 while Meta extended to five and a half years — the industry itself can’t agree. Satya Nadella, on Brad Gerstner’s podcast, described Microsoft’s approach as deliberately buying a little every year and aging the fleet — leading-edge chips for training, then cascading down to inference across a six-year depreciation life. That’s a coherent defense. But two details won’t sit still. Goldman calculated that moving the industry from five-year to three-year lives swings cumulative depreciation by a trillion dollars through 2031 — a single accounting assumption worth more than most countries’ GDP. And PIMCO, when it structures these deals as a lender, assumes the hardware may be worth nothing in five to ten years. The owners book six years of value. The lenders assume zero. Someone’s schedule is fiction.

Who benefits

Everyone inside the loop, at every step. The chipmaker books record sales. The cloud provider books backlog. The lab books a valuation. The private credit fund books fees — PIMCO has already marked paper gains on the Meta debt. And the circularity shows up in the earnings themselves: per their quarterly filings, Alphabet booked roughly $36.8 billion in equity gains largely from marking up its Anthropic stake, Amazon booked about $16.8 billion pre-tax on the same company, and Microsoft recognized $5.9 billion in OpenAI-related gains — paper profits on the labs whose spending commitments flow right back to the companies booking them.

And when the gap between commitments and revenue gets uncomfortable, watch what the insiders reach for. Last November at WSJ’s Tech Live, OpenAI CFO Sarah Friar described the financing ecosystem her company wants — banks, private equity, and a federal “backstop” to make the debt cheaper. The moderator asked if she meant a federal backstop for chip investment. She confirmed it. Watch:

Within twenty-four hours, three separate denials. Friar on LinkedIn: the word backstop muddied the point. Altman on X: OpenAI doesn’t want government guarantees and taxpayers shouldn’t rescue bad decisions. And the White House AI czar:

You know this move from Episode 3, because Georgia taught it to us: you don’t build the hedge unless the risk is live, and you don’t deny the bailout unless somebody floated it. The walkback is the confession.

What it costs you

Here’s where the engineering gets personal. Four senators wrote to Bessent in January warning that these complex and opaque debt structures could conceal true leverage and crush retirement savers. That’s not rhetoric — it’s a supply chain. Life insurers hold nearly $1 trillion in private credit. New York and Pennsylvania pension plans are invested in the Blue Owl fund behind Meta’s SPV. Private credit defaults ticked up to 2.73% in Q1. And an August 2025 executive order opened 401(k)s to alternative assets — extending the chain of exposure to anyone with a retirement account. The Fed’s May stability report now lists AI and private credit among the most-cited emerging risks, right behind geopolitics and oil.

And the leaked Treasury draft contains the quietest bombshell of all: unlike the dotcom era, fewer retail investors are directly holding this boom — meaning a sustained downturn lands hardest on institutional investors. Translation: not day traders. Pensions. The same people who got the electric bill in Episode 3 get the retirement statement in Episode 4. Their upside is contractual. Yours is a forecast.

One more thing, because the doomers I took apart in Episode 2 never do this: here’s what would prove me wrong. If organic AI revenue — real customers, not vendor-financed commitments — closes the gap on that $690 billion in capex by 2028, I’m wrong. If the refinancing wall PIMCO sees coming in 2030 clears without a single restructuring, I’m wrong. If the SPV debt performs, the pension funds get paid, and the depreciation schedules survive contact with Nvidia’s product cycle without a wave of write-downs, then the engineers of this structure built it honestly and I’ll write that episode too. I’m not predicting a crash. I’m telling you who signed up to absorb one. Those are different claims, and only one of them is sitting in the filings.

That’s the fiction of progress: not that the technology is fake, but that the risk was ever going to be shared. Which leaves one question for the finale. If the truth is this knowable — sitting in dockets, filings, fiscal memos, and leaked reports — why does almost nobody tell it to you straight?

That’s Episode 5: Truth, and What It Costs You.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Billy Finn has been standing in his yard in Dowagiac, Michigan with a decibel meter since 2022. Not because he’s a scientist. Because nobody else would measure it.

When the data center across from his house on Louise Avenue fired up, his meter read around 52 decibels. Today it typically reads 61, with spikes near 78. That’s the difference between a quiet conversation and standing next to a running garbage disposal, all day, every day. Two of his neighbors are suing over the noise. Residents packed a community meeting about it, and you can watch them describe it themselves:

Here’s what the national coverage mostly skipped: this facility is 30 megawatts, and it mines Bitcoin. It is not a hyperscale AI data center. It’s not even close. But it rode the AI panic wave into national headlines anyway, because “AI apocalypse” gets clicks and “mid-sized crypto facility with a noise problem” does not.

That gap, between the real harm and the story told about it, is what this episode is about.

The Claims on Trial

Three specimens, all from named sources, all still in circulation:

One. In December 2017, The Guardian published a piece warning that a “tsunami of data” could consume one fifth of global electricity by 2025. Twenty percent of the world’s power, eaten by the communications industry, in eight years.

Two. In June 2026, Futurism ran this headline: “Massive Data Center Cooks Nearby Residents Alive Amidst Deadly Heatwave.”

Three. The claim you’ve seen a hundred times in a hundred feeds: every AI query drinks a bottle of water.

All three contain a grain of truth. All three are fiction as stated. Let’s separate the two, because if you can’t, you’re exactly the reader both sides are counting on.

What’s Actually True

Start with the concessions, because the fear didn’t come from nowhere.

Data centers use real power and real water, and the growth is real. Lawrence Berkeley National Laboratory’s 2024 report to Congress puts U.S. data center electricity at 4.4% of national consumption in 2023, projected to reach somewhere between 6.7% and 12% by 2028. Note that’s a range, published by honest actors. Hold that thought.

Water stress is real in specific places. The Washington Post and UC Riverside researchers documented genuine water costs of AI computing, and in The Dalles, Oregon, court records eventually revealed that Google’s data centers consumed more than a quarter of the entire city’s water. The only reason we know that number is that a newspaper fought a 13-month lawsuit to get it, a lawsuit the city filed with Google footing the legal bills. Remember that the next time someone tells you the concerns are based on bad information. For fifteen years, the industry’s own playbook made good information illegal to share.

The noise is real. Ask Billy Finn’s meter.

And the opposition is real, massive, and bipartisan. Gallup found in March that 71% of Americans oppose a data center in their local area. For comparison, only 53% oppose a local nuclear power plant. Read that again. Americans would rather live next to a reactor.

So no, the fear is not fake. Here’s the one distinction that will let you navigate every headline in this fight, courtesy of a Utah conservationist who opposes the biggest data center project in America: his own state’s 35 existing data centers use a fraction of a percent of Utah’s water, and he calls generic fears about data centers guzzling water largely overblown. The project he’s fighting, though, would draw 9 gigawatts, roughly 3,000 times the average facility. Normal data centers and gigawatt-class data centers are different species. Anyone who lumps them together, in either direction, is selling you something.

What’s Fiction

Now the autopsy, specimen by specimen.

The Guardian’s 2025 prediction is dead, and the calendar killed it. 2025 came and went. The International Energy Agency rebutted the headline by name: data centers consumed roughly 1% of global electricity in 2018, and their energy use had been flat since 2015 even as internet traffic tripled. As of 2024, data centers sit around 1.5% globally. Not 20%. Not close. And buried in the original piece itself: the 20% figure was the researcher’s worst-case scenario, explicitly contingent on efficiency gains failing to show up. They showed up. The scenario the headline ignored is roughly what happened. The efficiency story that alarmist coverage almost never tells: Masanet and colleagues showed in Science that compute soared between 2010 and 2018 while energy use barely moved. (Fair warning on my own evidence: that’s pre-LLM-boom data, and the current buildout is a genuinely new chapter. But a falsified prediction is still falsified.) The kicker? The 20% estimate came from a researcher at Huawei. A tech company supplied the doom number. File that away.

Nobody was cooked alive. Under Futurism’s homicide headline: temperatures near a data center in Slough, England ran several degrees higher than surrounding areas during a heatwave, and the systemic claim rests on a not-yet-peer-reviewed paper estimating an average local effect of 3.6 degrees Fahrenheit. Real discomfort, real phenomenon worth studying, real residents worth hearing. Inflated into an atrocity. And in fairness, because fairness is the whole point: Futurism’s reporting on the Dowagiac noise story was careful and accurate. The problem isn’t the outlet. It’s the headline economy.

The water bottle stat is a masterclass in where they stop counting. The viral version says one AI-written email drinks a full bottle of water. The 519-milliliter figure is real, from the Washington Post’s collaboration with UC Riverside, but it counts the entire lifecycle, including the water used to generate the electricity, and it varies wildly by location. Count only the water that actually cools the servers and you get single-digit milliliters. Google’s own measured figure for a median prompt is about a quarter of one milliliter. Same prompt, a thousand-fold spread, depending entirely on where you stop counting. One detailed analysis reports that even Shaolei Ren, the researcher behind the original estimate, now puts the realistic per-prompt figure closer to 15 milliliters and considers the bottle-per-prompt framing obsolete. Meanwhile Sam Altman countered the panic with a 0.3-milliliter figure published with no methodology anyone can check. The alarmists cite the big number, the boosters cite the tiny one, and neither shows you what’s being counted. Both endpoints are marketing.

One more for the road, because it cuts against the doomers this time. When analysts estimated that 30 to 50% of planned 2026 data center capacity faced delays, headlines rounded it to “half of data centers canceled,” and the analysts at SemiAnalysis published a rebuttal literally titled “Stop Saying Half of 2026 US Datacenter Capacity Is Canceled,” showing much of the “canceled” capacity was speculative paper that was never going to be built in 2026 anyway. Honest actors publish ranges. Motivated actors publish endpoints. That rule doesn’t care which team you’re on.

Who Profits From Your Panic

Follow the money on the fear itself, because it flows in more directions than you think.

Fear as a fundraising deck. Starcloud is a startup that wants to put data centers in orbit. It reached a valuation over $1 billion, reportedly the fastest company in Y Combinator history to do it. Its pitch, in its own recruiting materials, is that hyperscale data centers will strain electricity grids, freshwater distribution, and Western permitting systems past the breaking point. In other words: a billion-dollar valuation built substantially on the premise that the doomers are right. The alarmist narrative isn’t just clickbait. It’s collateral. When a founder tells you the Earth can’t handle AI, check what he’s selling. Usually it’s the escape pod.

Dismissal as a development strategy. Kevin O’Leary is building Stratos, the Utah mega-project mentioned above. When locals in deep-red Box Elder County objected, he called their concerns “misinformation about data centers” and went on Fox Business to suggest the opposition traced back to a single adversary: China. It went poorly. O’Leary admitted he had no evidence that the Utah groups and citizens he named by name were funded by the Chinese Communist Party, and confirmed a legal demand letter arrived before his change of heart. Fox News then delivered on-air apologies across multiple shows, stating the network was aware of no evidence for the claims. After Utah’s Republican Senate President demanded massive cuts to the project, O’Leary conceded he’d “screwed this up” and pissed off a lot of people. Residents are now suing over how the project got approved. Watch him defend it, then decide for yourself:

Pay attention to the move O’Leary made, because it looks superficially like the one this article is making. He said the fears were misinformation. I’m saying some specific claims are fiction. The difference is receipts. I named what’s true first, conceded the real harms, and cited every rebuttal. He called his neighbors foreign agents and retracted it after the lawyers wrote in. Debunking is a discipline. Dismissal is a sales tactic wearing its clothes.

And the politicians? They read polls. Texas Governor Greg Abbott called his state “the epicenter of AI development” in November while announcing Google’s $40 billion investment. Eight months later, at a re-election campaign stop, he declared “we must prohibit them from building AI data centers in rural Texas neighborhoods” and called for eliminating their tax breaks. What happened in between? Polling showed nearly two-thirds of rural Texans oppose the facilities, and reporters noted Abbott has taken over $2 million from tech and AI-linked donors since last year. Donors pulling one way, voters pulling the other, rhetoric tracking the voters in an election year. I’ll let you do that math yourself. The full ledger of who pays for all of this is Episode 3.

The venture capitalist Rubén Domínguez Ibar put it cleanly: the SaaS 1.0 playbook no longer matches how AI companies grow, price, or build. He’s right, and everyone in this story knows it. That’s why the old quiet approach, tax breaks and NDAs and nobody bats an eye, stopped working. The scale broke the silence.

What This Costs You

Here’s the part that should actually worry you. Pew’s research found that the more Americans hear about data centers, the more negative they get: two-thirds of people who’ve heard a lot about them say they’re bad for home energy costs, versus 42% of those who’ve heard only a little. The opposition isn’t ignorance. It’s information. But when the load-bearing facts under that opposition turn out to be inflated, cooked-alive headlines and falsified predictions, the people with real, measurable harms lose credibility by association. Billy Finn’s decibel meter gets filed next to the tsunami that never came.

Senator Josh Hawley, announcing the first bipartisan bill on data center energy costs, put the stakes plainly: American families should not have to shoulder rising electricity costs produced by data centers. When Hawley and Richard Blumenthal are co-sponsoring, and Hawley and Elizabeth Warren are co-signing letters, the divide isn’t left versus right anymore. It’s the people who pay the bills versus the people who write the press releases.

Bad information doesn’t protect communities. It discredits them, right when they need to be believed.

The Part Where I Turn Around

If you work in this industry and you’ve been nodding along, enjoying watching the alarmists take their lumps: stop smiling.

Every forensic standard applied above, named sources, primary documents, predictions checked against outcomes, is about to be pointed the other way. The progress narrative, the jobs numbers, the subsidy math, the “trust us” economics of the biggest infrastructure buildout in a generation. I’ve run the receipts. It fares worse.

Episode 4 is where we count their fictions. Episode 3 is where we find out who’s paying for all of it.

Spoiler: you are.

Miss Episode 1? Start here: Where They Stop Counting

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Here are two numbers that describe the exact same thing.

An AI chatbot answers your question. According to a study amplified by The Washington Post, that answer cost about 519 milliliters of water. A full bottle. According to Sam Altman, whose company built the chatbot, it cost about 0.3 milliliters. A few drops.

Same question. Same servers. Same physics. The numbers differ by a factor of more than 1,700.

One side wants you scared. The other wants you sedated. Neither is lying about the arithmetic. They’re lying with something quieter: the boundary. Where you’re allowed to stop counting.

This series is about that trick, and every variation of it, applied to the biggest infrastructure buildout in American history. Before we get to who’s running it, you need to know what the thing actually is.

The thing itself, in plain English

A data center is a warehouse full of computers. That’s it. No mystique. Racks of servers, each one throwing off heat like a space heater running at full tilt, packed into buildings that can sprawl across thousands of acres.

Two resources dominate everything about them: electricity to run the chips, and cooling to keep the chips from destroying themselves. Cooling often means evaporating water, which is cheap and effective and permanently consumes the water. Those two inputs, power and water, are where every fight in this series lives.

The scale is real and worth stating plainly. Lawrence Berkeley National Laboratory, in a congressionally mandated report to the Department of Energy, found that U.S. data centers consumed about 176 terawatt-hours in 2023, roughly 4.4 percent of all U.S. electricity. By 2028, LBNL projects between 6.7 and 12 percent.

Read that range again. The government’s own best estimate of the near future spans nearly a factor of two. That honesty matters, because almost nobody who quotes this report quotes the range. Alarmists cite 12 percent. Boosters cite 6.7. The lab published uncertainty; the loudest voices on both sides amputated it.

That’s the pattern. Now watch it work.

The water bottle

On September 18, 2024, The Washington Post published a piece headlined “A bottle of water per email: the hidden environmental costs of using AI chatbots,” built on research with UC Riverside. The number went everywhere. It’s probably in your group chat.

Here’s what traveled less well. In the body of its own article, the Post concedes the burden is “nearly impossible to quantify.” The hedge lives in the copy. The certainty lives in the headline. And the headline is the only part that got screenshotted.

The 519ml figure isn’t fabricated. It comes from real research by Pengfei Li, Jianyi Yang, Mohammad Islam, and Shaolei Ren. But it counts wide: it includes the water consumed by the power plants generating the electricity, not just the water cooling the servers. It’s also a modeled estimate that swings from roughly 235ml to over 1,400ml depending on where the data center sits. A single scary number was extracted from a range and a methodology, and the caveats were left on the cutting-room floor.

Now the other side.

Sam Altman says a ChatGPT query uses about 0.3 milliliters of water. Soothing. Also completely unauditable: OpenAI has never published the methodology. And the number counts narrow, on-site cooling only, ignoring everything upstream. He drew the boundary at the property line and declared victory.

You know who settled which boundary is honest? The companies themselves, by accident. This July, The Wall Street Journal reported that among the tech giants, only Meta counts the water used by the power stations feeding its data centers. Meta’s disclosure showed indirect water use of 19 billion gallons in 2024, more than 20 times its direct use. LBNL’s analysis puts the historical ratio for U.S. data centers at roughly 12 to 1. Meanwhile, Microsoft advertises “zero water” data centers and replenishment pledges that count only on-site water. No law requires anyone to count the rest.

So: the scary number counts everything and hides its uncertainty. The soothing number counts almost nothing and hides its methodology. Mirror-image distortions of the same variable. The alarmist needs you to never see the boundary. So does the CEO.

The same trick, electrified

On April 10, 2025, the International Energy Agency published a report called Energy and AI. It contains a genuinely striking projection: by 2030, the U.S. is on course to use more electricity for data centers than for manufacturing all energy-intensive goods combined, including aluminum, steel, cement, and chemicals.

That’s a real, citable finding from a serious institution. Which is exactly why it’s such useful raw material for distortion.

On June 26, 2026, a viral X post from the account @YourAnonOne retailed the claim as proof that “training and running AI” will out-consume American heavy industry. Three quiet edits happened on the way to your feed:

First, AI became the whole story. The IEA figure covers all data center load: cloud, streaming, storage, your company’s email servers. AI is the growth driver, not the total.

Second, electricity became energy. Steel and cement burn staggering amounts of coal and gas directly. That heat never touches the grid, so an electricity-only comparison makes heavy industry look smaller than it is. The IEA was explicit about its boundary. The tweet wasn’t.

Third, a projection became a fact. The IEA calls this its base case and flags major uncertainties: adoption speed, efficiency gains, grid bottlenecks. A 2030 scenario got sold as a meter reading.

Here’s the part I find genuinely funny. The same IEA report also notes that in the context of global electricity growth, data centers are a modest driver; economic growth, EVs, and air conditioning are all bigger. One PDF, two quotable sentences pointing in opposite directions. The alarmists tweet the steel line. The boosters cite the modest line. Nobody links the report. IEA chief Fatih Birol called AI “one of the biggest stories in the energy world today.” He might have added: and one of the most selectively quoted.

The boosters’ bottle of water

If you think distortion is a media disease, come to the White House.

On January 21, 2025, President Trump announced Stargate, a private venture he called the largest AI infrastructure project in history, “a resounding declaration of confidence in America’s potential,” flanked by Altman, Larry Ellison, and Masayoshi Son. The headline promise: more than 100,000 jobs.

Where did 100,000 come from? Nobody knows. No methodology was released. No split between construction jobs, which are real but temporary, and permanent jobs, which are the ones a town actually keeps. It is, structurally, the boosters’ bottle of water: a number engineered to travel, stripped of everything you’d need to check it.

The checkable evidence points somewhere much smaller. A Brookings study published this May, analyzing roughly 770 U.S. facilities, found that the naive job estimates favored by industry-sponsored reports overstate data center employment by about a factor of three. The real-world math is blunter still: Meta’s $10 billion campus in Lebanon, Indiana supports about 300 permanent jobs. One job per $33 million invested.

And communities are paying for those jobs twice. Virginia’s data-center sales-tax exemption cost the state an estimated $1.6 billion in fiscal 2025 alone. Good Jobs First documented a New York facility that promised 125 jobs in exchange for $1.4 billion in subsidies, about $11 million per job. As the group’s Kasia Tarczynska told Consumer Reports, these projects “tend to be very extractive and bring very little to local communities.”

To be fair, and this series will always be fair before it’s fun: Brookings also found real local gains, including a 22 percent jump in information-sector employment and modest wage growth. “Fewer jobs than promised” is not “no benefit.” The claim on trial is 100,000. The evidence convicts the claim, not the concrete.

Haven’t we panicked about this before?

Yes, and the record should humble everyone.

Between 2010 and 2018, computing workloads in data centers grew more than sixfold. Total data center energy use grew six percent. Not sixfold. Six percent. That finding, from Masanet and colleagues in the journal Science, is the tombstone of a decade of “the internet will eat the grid” predictions. Efficiency ate the growth instead.

The boosters love that study. Here’s what they skip: LBNL’s data shows the efficiency era is over. Data center energy growth ran about 7 percent a year through 2018, then jumped to 18 percent a year through 2023, and is projected at 13 to 27 percent annually through 2028. The alarmists were wrong last decade. The people promising efficiency will save us again may be wrong this decade. Both camps quote the half of history that flatters them.

The rule for this series

Before we go, one more specimen for your collection. You’ve probably heard that a ChatGPT query uses ten times the electricity of a Google search. It’s everywhere. Trace it and you find a Goldman Sachs note leaning on an old Alphabet remark, with an underlying measurement that was never published. Unauditable at the source, repeated ten million times downstream. It doesn’t run in this series, and when a number like it appears anywhere in these pages, it will be labeled exactly what it is: unverified.

So here’s the rule, and the thesis of everything that follows.

Honest people publish ranges. Motivated people publish endpoints. When someone hands you a single confident number about data centers, your first question is never “is it true?” It’s “where did they stop counting, and who benefits from the fence being there?”

Next episode, we apply that rule without mercy to the fear machine: the viral environmental claims, tested against the primary data, including the ones that collapse. And if you’re already sharpening the “industry shill” comment, hold it. Episode 4 takes the same knife to the buildout itself, the subsidies underneath it, and the people telling you it’s all progress. Both sides drew their boundaries.

We’re going to move the fences.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

This is the political close to our series. My partner Jason Hubbard wrote the engineering one — They Won the Fight. They Lost the War. He proved the cage was always going to leak. I’m here to total what it cost. Read them together; they’re two halves of the same ending.

Back in April, when the rest of us were still watching the front of the house, my partner Jason Hubbard wrote a line I couldn’t shake:

“The people who were raising hands the whole time, at their own expense, watching the twists getting tighter year after year, will be the ones positioned to do something with the moment when it arrives.” — Jason Hubbard, It’s the Runtime, Stupid

He also said this: “The dam is going to break. And dams don’t break gently or gradually.”

He wrote that before the government pulled two frontier models off the market. Before a Commerce Secretary’s letter turned restoration into a leash. Before a President went on CNBC and bragged about stopping “a bad player” quickly and effectively. Before the industry started offering Washington equity in itself and calling it a gift to you.

The dam broke. Right on schedule. This series was just me standing at the bottom of it, writing down what came through.

So let’s total the receipt. Not the politics — the cost. Because every move in this saga got paid for by someone, and it wasn’t the people who made the moves.

The Veteran Paid

Start where I always start, because these are my people. While Washington and the labs fought over who controls the most powerful models on earth, the VA quietly stood up 367 AI systems — 215 of them touching claims, records, and care. The government’s own watchdog flagged that nobody had a reliable way to catch those systems when they fail. A Navy vet learned AI was in his exam room when a nurse tried to slide it in mid-appointment, and got pushback when he said no.

Here’s the tell that should make your jaw tighten: the same government that can switch a $200 billion company’s models off in ninety minutes cannot tell a veteran when an algorithm touched his disability claim. The state’s appetite for control runs exactly one direction — toward the powerful, never toward the person who wrote a blank check to this country. Precision when it’s a corporation. Fog when it’s a Marine.

The Citizen Paid

Then they told you that you’d won something. A “share.” A stake in the upside. A Public Wealth Fund, dividends for everyone, AI’s bounty spread to the people at last.

Read the structure, not the press release. What was actually floated was donated, non-voting stock routed through a government vehicle — a dividend stub with no seat, no vote, no wheel. You don’t get to steer. You get a receipt. Meanwhile the thing you supposedly co-own is now bound at the balance sheet to the same government that regulates it, approves its releases, and pockets the upside when the rules crush the smaller competitor who couldn’t afford the compliance freight.

You weren’t handed ownership. You were handed a story pretty enough that you’re supposed to say thank you.

The Builder Paid

And the founder — the one everyone assumes is winning — paid too, in the currency that matters most to a builder: independence. When you can be switched off on a Tuesday, being right stops mattering. Anthropic was right on the merits and still bought its own product back with standing concessions. OpenAI read that room and preemptively offered tribute. That’s not a market anymore. That’s a protection racket where the marks volunteer, because the alternative is worse.

You cannot take a company public when the government owns the off-switch. So you make the government a partner. You call it patriotism. And the cost — the one nobody prices — is that the next great American AI company will be built to please a regulator before it’s built to serve a user. That’s not the country’s engine. That’s a utility with a flag on it.

They Won the Fight. They Lost the War.

That’s the title of Jason’s finale, and he earned it, so I’ll pay it off from my side.

The government won the fight. Institutionally, completely, permanently. The surveillance apparatus, the pre-release access, the equity entanglement — none of that is coming back out. No law passed, no vote taken, no court ruling. Just a letter, a leash, a stake, and a thank-you post. The control grid got built, and it’s staying built.

And it was all for nothing.

Because the thing they built the grid to control was never controllable. Jason lays out the engineering in his finale, and you should read it — the short version is that capability leaks through the front door faster than any vault can hold it. Distillation, open weights, the runtime doing the real work while everyone guards the wrong thing. They won every battle over the frozen model and lost the war to the running one. They fused the state to the platform, charged you for the privilege, made the veteran a subject and the citizen a shareholder-in-name — and at the end of it, the capability they seized walks out the back while they’re still congratulating themselves at the front.

That’s the whole tragedy in one line: they built a cage strong enough to hold you, and useless against the thing they were actually afraid of.

Jason called the dam in April. It broke on schedule. The people who raised their hands early — at their own expense, watching the twists get tighter — are the ones who get to say what it meant when it finally gave.

So here’s what it meant. The powerful protected themselves and billed the rest of us. Same as it ever was — just faster now, and wearing a lab coat.

Read Jason’s engineering finale for the proof the cage was always going to leak: They Won the Fight. They Lost the War.

We raised our hands. We’re still raising them. And we’ll be here when the next dam starts to crack.

THE BUBBLE AND THE BACKLASH — FOUR-PART INVESTIGATION

PART 1: The Bubble and the Backlash
THE ALLIANCE CRACKED
Washington and Silicon Valley start fighting over control.

PART 2: The Vault Got a Revolving Door
THE DOOR CAME WITH TERMS
Anthropic gets access back, but Washington retains leverage.

PART 3: The Tell
THE $42 BILLION TELL
OpenAI reportedly floats giving Washington equity as political risk becomes impossible to ignore.

YOU ARE HERE → PART 4 / FINALE: The Bubble and the Backlash: What It Cost
THE RECEIPT CAME DUE
Washington built the control system. Veterans, citizens, and builders absorbed the cost. The finale explicitly totals what the veteran, citizen, and builder paid and closes on the failure of the containment system itself.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Previously on The Control Grid: The Lutnick letter proved the Fable 5 recall was never a kill switch — it was a sorting hat. Government, national labs, and “trusted partners” got the real model back. You got the waiting room. Read: You’re Already Inside the Vault. Now What? →

Markus Williams served in the Navy. Last week, as Military Times reported in its July 1 investigation, he learned the VA uses artificial intelligence on his medical visits the way most of us learn anything from the government — after the decision was already made. A nurse moved to bring an AI scribe into his appointment. He declined. And here’s the part that should stop you: declining got him pushback from the staff. Not “no problem, we’ll turn it off.” Resistance.

Sit with that, because it’s the whole story in one exchange. Consent that produces friction when you exercise it isn’t consent. It’s a formality you’re expected to sign.

That friction arrived years late, and it arrived in an exam room instead of a disclosure form. According to the VA’s own 2025 AI Inventory, also surfaced by Military Times, the department is running 367 documented AI use cases. Two hundred fifteen of them are classified “high-impact” — meaning they touch the things that actually matter: your claims, your records, your care. More than 50,000 VA personnel are using commercial AI tools on top of that.

Let me be fair before I’m honest, because those are two different jobs.

Some of this AI is doing real work. The VA’s TERA memo tool pre-populates toxic exposure paperwork and has assisted more than 181,000 forms at a reported 98% accuracy, saving an estimated 54,000+ work hours. If you’re a Gulf War vet who waited decades for the government to admit burn pits were real, faster claims processing isn’t the enemy. I’m not here to smash the machines. Marines adapt; we don’t Luddite.

Here’s the honest part: none of this surprises me, and if you’ve spent one day in uniform, it doesn’t surprise you either. The government’s default setting with veterans has never been malice. It’s something more efficient than malice — deploy first, disclose when asked, apologize when subpoenaed. It took Congress until 2022 to pass the PACT Act for exposures we breathed in the ‘90s. Nobody at the VA woke up scheming to screw veterans with AI. They didn’t have to. The system screws veterans on autopilot; AI just removed the pilot.

Regular readers know this pattern, because we’ve been mapping it all month. When the Lutnick letter leaked, I showed you the sorting hat in writing: the “national security” risk around Fable 5 evaporated the moment the user had the right badge — government agencies, national labs, anointed partners back in; you, still out. The VA just handed us the domestic edition: an agency that trusts itself with 367 AI systems before it trusts a veteran with one sentence of notification.

“They told us AI safety would protect the public. What they built instead was a velvet rope around reality.” — Eric Mitchell, The AI Vault Is Real

Same rope, different building. At the VA, the velvet rope runs between the veterans the systems operate on and the information about how they operate.

Which brings us to the actual problem. It’s not the 367 systems. It’s the architecture holding them — or rather, the PDF pretending to.

My colleague Jason Hubbard has spent months documenting why AI systems fail, and his diagnosis fits a federal agency like a dress blue jacket:

“Most AI workflows fail because they are written like instruction lists instead of built like control systems.” — Jason Hubbard

That’s the VA’s inventory in one sentence. An inventory is an instruction list. It tells you what exists. A control system tells you what happens when something drifts, fails, or quietly starts making mistakes on 181,000 forms. The VA published the first and is asking for credit for the second.

And a list on department.va.gov is not the same thing as reality in the exam room. Jason again:

“Define what the eventual workflow will actually know at runtime. This is where fantasy workflows die. Do not design the workflow as though it will have access to context it will not reliably have.” — Jason Hubbard

Runtime, for a veteran, is the moment a nurse asks you a consent question you were never prepared to answer, about a system you were never told existed. Everything else is documentation theater.

The veterans service organizations — not activists, not doomers, the VFW and DAV — see the same gap. Chris Macinkowicz of the VFW put it in language every NCO understands: “AI needs to be a tool, not an employee... We need to make sure human eyes are on it afterwards.” Jon Retzer of Disabled American Veterans called the inventory a step toward transparency, then noted it says almost nothing about how systems are monitored after deployment, how performance gets evaluated over time, or what accountability triggers when something breaks.

In the Corps, every action has a name attached to it. Who ordered it, who executed it, who signs the after-action. The VA deployed 367 systems and the accountability layer is a spreadsheet.

Even the straightest-laced outlet on the veterans beat couldn’t make this sound routine.

Two hundred fifteen “high-impact” systems, reported in one flat sentence — and that flatness is the story. When the paper of record for the military community can state it that plainly and nobody blinks, the deploy-first-explain-later posture has already become normal. Normal is exactly what we’re here to interrupt. (Full Military Times report here.)

And this isn’t me raising the alarm. It’s the VA’s own watchdog. The department’s Office of Inspector General flagged a patient-safety concern over generative AI in clinical settings — finding the Veterans Health Administration had no formal mechanism to identify, track, or resolve the risks these tools introduce once they’re in the room with a patient.

Read that against the 367 number and the picture sharpens. It’s not that the VA is hiding a secret army of algorithms. It’s that the VA stood up hundreds of systems, its own inspector found there was no reliable way to catch them when they go wrong, and the fix on offer is a spreadsheet you can download. The watchdog saw the hole. The hole is still open.

So here’s the ask, and it isn’t radical. Three things, in plain English:

1. Notification. When AI touches a decision about your claim, your record, or your care — you get told. In the letter. Not in a PDF you’d need a FOIA reflex to find.

2. Published audits. Post-deployment accuracy numbers, updated and public. If TERA’s 98% is real, the VA should be proud to keep proving it.

3. A human appeal path. Any decision an algorithm touched, a veteran can put in front of a person with a name.

One last thing. This same week, the federal government demonstrated it can switch the most powerful AI models in the world off and on like a porch light. Washington has no trouble controlling AI when the target is a company. Ask it to guarantee a veteran knows when an algorithm read his file, and suddenly the machinery gets shy. I told you the recall button was never a kill switch — it was a sorting hat. Well, veterans just found out which house they got sorted into. Same one as always: served first, briefed last.

We adapted. Now they overcome — the paperwork.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Here’s the first thing that jumped out at me, and it’s the thing most of the coverage skated right past: according to the FT’s sources, Sam Altman took this idea to Donald Trump, Howard Lutnick, and Scott Bessent — and to Bernie Sanders.

Sit with that lineup. The President, his Commerce Secretary, his Treasury Secretary, and the democratic socialist from Vermont who wants to tax the entire industry into the ground. When a founder is reportedly floating the same concept to the administration and its loudest ideological enemy in the same stretch, he’s not advancing a principle. He’s buying insurance, and he doesn’t much care who writes the policy.

Per the Financial Times — citing people familiar with the talks, not anyone on the record — OpenAI has floated handing the U.S. government an equity stake of up to 5%, worth roughly $42.6 billion against its $852 billion valuation, donated rather than sold, through a proposed “Public Wealth Fund” modeled on Alaska’s oil-dividend setup. The arrangement reportedly extends to Anthropic, Google, and Meta ceding matching slices. Bloomberg, CNBC, and Forbes have all since picked it up. None of the other companies have agreed to anything. Reuters couldn’t independently verify the terms. It would likely require an act of Congress. And OpenAI’s own IPO may now be slipping to 2027.

One detail worth holding onto before we go further: there is no public statement from Sam Altman explaining any of this. Not a speech, not a post, not a quote in his own name. Everything we know about his reasoning comes secondhand, through sources describing private conversations with Trump, Lutnick, Bessent — and Bernie Sanders. Keep that in your pocket. It matters more than it looks.

Every outlet ran the number. I want to talk about the tell.

What I Learned in the Wild West

After the Marine Corps, I spent over a decade in Silicon Valley. In my early SaaS days out there, there was a number everyone repeated like scripture: ninety percent of startups fail. Nobody could source it cleanly, but it didn’t matter, because it captured something true about how founders actually think. You live inside a paradox: you have to sell everyone — investors, employees, yourself — on infinite upside, while some cold back-corner of your brain runs the survival math nobody says out loud.

That’s the lens I put on Altman’s offer, and once you see it, you can’t unsee it.

You do not give away five percent of a company you genuinely believe is going to ten trillion dollars. You just don’t. A founder who’s convinced the rocket clears the atmosphere clutches every basis point, because he knows what it’ll be worth on the other side. You give away five percent when the cold back-corner of your brain has run the survival math — and the answer came back scared.

So the question isn’t “why is Altman being generous?” Generosity has nothing to do with it. The question is: what does he see coming that makes $42 billion a cheap price to pay right now?

Two things. And this is where Part 1 and Part 2 come home to roost.

The Two Things He Sees

First, he watched what happened to Anthropic. The government pulled their frontier models off the market for nineteen days and handed them back only after Anthropic agreed to pre-release access, threat sharing, embedded federal partnerships, and a letter that lets Commerce reverse the deal whenever “circumstances change.” Altman saw a live demonstration that Washington owns an off-switch — and that being right on the merits, which Anthropic was, doesn’t buy you back your own product. It buys you a leash.

And if there was any doubt the administration sees it that way, Trump erased it on CNBC on July 2, talking to Joe Kernen about AI competition with China. His words: you want as little regulation as possible, but “when we think there’s a bad player out there, there’s a little danger, we stop the player quickly and effectively. We had a recent case where we did that.”

Read that twice. The President of the United States, on camera, confirming the off-switch and bragging about the trigger pull. “A recent case” is Anthropic. That’s not a leak, not a source familiar with the talks — that’s the man holding the switch describing how fast he’ll use it. Now put yourself in Altman’s chair, watching that interview with your IPO paperwork sitting at the SEC. You just got told, on national television, exactly what happens to a lab Washington decides is “a little danger.”

And note the tell inside the tell: Trump framed all of it as wanting “as little as possible” regulation. Light touch. Minimal restriction. That’s the rhetoric. The recall, the leash, the pre-release access, and now a floated equity stake are the record. Watch what they do, not what they say — the lightest-touch language in Washington is running the heaviest hand in tech.

If you’re OpenAI, filing confidential IPO paperwork with that hanging over you, you have a math problem. You cannot take a company public when the government can switch it off on a Tuesday. No underwriter prices that. No institutional investor swallows it. The off-switch isn’t a policy annoyance; it’s an existential discount on your valuation. Five percent of the company is a rounding error against making that risk disappear.

Second — and this is the part the boosters won’t say into a microphone — the bubble knows it’s a bubble.

You don’t hedge a sure thing. You hedge when you’re scared. And the entire posture of this industry right now, underneath the superintelligence sermons, is defensive. SoftBank pouring in another ten billion. Reports that Nvidia has been renting back its own compute from cloud providers — a strange move that only makes sense if the end-user demand to absorb the buildout isn’t fully there. Valuations detached from revenue by an order of magnitude. Then the market leader reportedly offers Washington a $42 billion chunk of itself and calls it a gift to the public. That’s not the move of a company that believes its own trillion-dollar story. That’s the move of a founder running the survival math and deciding a government partner is cheaper than a government adversary — especially if the whole thing wobbles.

Ninety percent fail. Altman knows the number too. He’s just the rare founder rich enough to buy his way out of the tenth percentile — with equity that was never guaranteed to be worth anything anyway.

The “Public” Sleight of Hand

Now watch the framing, because it’s a magic trick and the misdirection is elegant.

The pitch attributed to Altman, per the FT’s sources, is that giving the public a financial stake is the best way to share AI’s upside. Democratic. Populist. Who could argue? The Alaska Permanent Fund cuts every resident a check; picture that, but powered by superintelligence. It’s a beautiful story — and notably, one told about Altman by people familiar with the talks, rather than by Altman at a podium where he’d have to defend it.

It’s also, on the structure reported so far, mostly theater. A donated, non-voting stake routed through a government fund is not you getting rich. It’s the company buying political goodwill with paper that costs it nothing where it counts — control. No board seat changes hands. No voting power moves. Altman keeps the wheel; Washington gets a dividend stub and a talking point.

You want to know what actually ceding power looks like? Sanders has separately pushed a far more aggressive version — reporting has him seeking a stake on the order of fifty percent of major AI firms, the kind of position that comes with real ownership and real votes. Set the two concepts side by side. Sanders wants ownership. The OpenAI proposal, as reported, is offering a tip. And here’s the thing about floating the smaller number into a debate where the larger one is already live: whether or not it’s deliberate, it makes 5% look moderate by simply standing next to 50%. That’s how anchoring works. You don’t have to intend it for it to benefit you.

The Steelman, Because He’s Owed One

Here’s where intellectual honesty costs me something, so I’ll pay it.

The generous read is that Altman means it. OpenAI published a “Public Wealth Fund” concept in an April white paper, and reporting traces his interest in broad-based AI benefit back to early 2025 — well before the recall, before the off-switch got demonstrated. So you can’t call this purely reactive panic. There’s a documented throughline, on the company’s own paper and in more than a year of reported talks, of the argument that AI’s gains should reach ordinary people. Grant that he may have simply found a mechanism.

Fine. Grant all of it. The idea can be sincere and the timing can still tell on him. Because a sincere conviction that surfaces as a concrete $42 billion trial balloon the same week your competitor gets leashed and your IPO looks shaky is a conviction that found its moment for reasons that have little to do with conviction. Intent and structure are different animals. The intent may be public benefit. The structure — donated, non-voting, floated quietly, timed to the leash — is insurance. Both can be true. Only one is load-bearing.

The Pattern Is the Point

Strip away the 5% figure, which may not survive negotiation, and the thing underneath is what matters — because it’s already happening and it doesn’t depend on this one deal.

The government took a 10% stake in Intel last year. It negotiated a cut of Nvidia’s and AMD’s chip revenue on sales to China. And it’s reportedly taken or explored positions in other strategic companies during this administration. Now the leading AI lab is offering equity, and floating that its rivals do the same. Whatever the exact tally, the direction is unmistakable: this isn’t a one-off act of corporate generosity. It’s the latest move in a sequence, and the sequence points one way — the U.S. government is becoming a shareholder in the companies it regulates.

Run that to its conclusion, because nobody in Washington will say it plainly. If the government owns a piece of OpenAI, it sits on both sides of every table. It’s the referee and the team owner. It writes the safety rules and collects the dividend when the rules crush a smaller competitor who can’t afford the compliance freight. It approves the model release and books the upside. There is no arms-length oversight left when the overseer is on the cap table.

My colleague Jason Hubbard flagged exactly this the morning Trump first floated the idea back in June, and he put it plainer than I can:

“The regulatory body responsible for overseeing an industry just announced it might become a financial stakeholder in that same industry — the very definition of a conflict of interest.” — Jason Hubbard, Trump’s Decided to Buy a Timeshare on the Titanic

He called it a five-alarm governance catastrophe on day one. A month later, it’s not a hypothetical — it’s a proposal with a dollar figure attached. Read his full breakdown of the debt math underneath it; it’s the fiscal case that runs alongside my political one.

That’s not sharing the wealth. That’s fusing the state and the platform into a single entity with one bank account — and calling it a citizen dividend so it goes down smooth.

What the Citizen Actually Gets

Here’s where I land, and it ties off everything since Part 1.

The recall proved the government can turn the models off. The restoration proved it can leave them on with a leash. The framework — arriving as early as next week, per the FT — will set the rules the labs already agreed to follow. And now the equity offer proposes to bind the regulator and the regulated together at the balance sheet, permanently.

Four moves. One direction. And at no point in any of it does the ordinary citizen — the veteran waiting on a claim, the founder trying to build without a federal minder, the taxpayer told he’s getting a “share” — get a seat where the decisions actually happen. He gets a dividend stub and a press release. The people who wrote a blank check to this country get told they’re now part-owners of a company whose wheel they will never touch.

You were promised a share of the upside. What you got was a government more entangled with frontier AI than any regime in American history, a market leader that just told on itself, and a story pretty enough that you’re supposed to say thank you.

Ninety percent fail. The house always runs the math. And the house just made the government a silent partner.

THE BUBBLE AND THE BACKLASH — FOUR-PART INVESTIGATION

PART 1: The Bubble and the Backlash
THE ALLIANCE CRACKED
Washington and Silicon Valley start fighting over control.

PART 2: The Vault Got a Revolving Door
THE DOOR CAME WITH TERMS
Anthropic gets access back, but Washington retains leverage.

YOU ARE HERE → PART 3: The Tell
THE $42 BILLION TELL
OpenAI reportedly floats giving Washington equity as political risk becomes impossible to ignore.

PART 4 / FINALE: The Bubble and the Backlash: What It Cost
THE RECEIPT CAME DUE
Washington built the control system. Veterans, citizens, and builders absorbed the cost.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

UPDATE — July 2, breaking. Hours after this published, the Financial Times reported OpenAI has offered to donate the U.S. government a stake of up to 5% — worth around $42.6 billion — and floated Anthropic, Google, and Meta ceding similar slices through a sovereign-wealth-fund vehicle. Read who’s doing the offering. This isn’t Washington seizing the labs; it’s the labs trying to make the referee a shareholder and calling it a gift to the public. The off-switch worked so well that the industry is now volunteering equity to keep it holstered. That’s not the backlash easing — it’s the backlash getting cheaper. Full breakdown in Part 3.

Previously on The Control Grid: In Part 1, we watched the pro-AI coalition splinter in real time — Trump’s own AI czar breaking ranks, Congress fracturing off party lines, and the administration floating equity stakes in the companies it regulates. We closed with a breaking line: the Fable 5 reversal was under negotiation, with only the Pentagon and NSA left to sign off.

That was this morning.

By this afternoon, it was done. Commerce Secretary Howard Lutnick announced the export controls on Fable 5 and Mythos 5 were lifted. Anthropic said access starts flowing back today. Nineteen days from worldwide confiscation to full restoration.

We called the shot in Part 1. I’d take a victory lap, but there’s no time — because everyone else is covering this as a happy ending, and the actual story is sitting in the fine print they’re skimming.

The models came back. The question Sacred Loop exists to ask is: what did the keys cost, and who holds the copy?

Read the Letter, Not the Headline

Lutnick’s letter to Anthropic co-founder Tom Brown, obtained by The Hill, says a license “is no longer required” for the models. Everybody stopped reading there. Keep going.

Per the letter, Anthropic agreed to proactively detect and address security risks, work with the U.S. government on protocols and standards “for these and future models,” and inform the government of any malicious activity. And then the sentence that should be the headline everywhere: Commerce “reserves the right to reevaluate the decisions made in this letter” if Anthropic fails to adhere to its commitments — or if “circumstances change.”

Circumstances change. Two words doing the work of an entire leash.

This is not a legal framework. A framework tells you the rules before you play. This is a permission slip with a revocation clause, personally signed, and it covers not just the two models that got confiscated but future models that don’t exist yet. The recall proved the off-switch was real. The restoration just proved the on-switch has a price — and the price is a standing appointment with Washington before your product legally exists.

Here’s the Secretary announcing it himself — note what gets celebrated and what goes unmentioned.

Alignment across the U.S. government. For a commercial software product. Say that sentence out loud in 2019 and watch where it lands.

What Anthropic Actually Paid

Here’s the part that makes this a Sacred Loop story and not a business-page story. Anthropic spent the entire standoff insisting the government had it wrong. Their own statement called the jailbreak narrow, the vulnerabilities minor and previously known, and warned that if this standard were applied across the industry “it would essentially halt all new model deployments for all frontier model providers.”

Then they paid anyway. Per Anthropic’s own resolution announcement, the company will now:

• Give the government pre-release access to any model that “materially advances the capability frontier in areas relevant to national security”

• Share threat intelligence on how its tools get abused

• Join the vulnerability clearinghouse created by Trump’s June executive order

• Substantially scale up federal partnerships, including dedicated personnel and compute

• Build a “shared, voluntary security and evaluation standard” with the government and industry peers

Read that list twice. That’s not a company that lost an argument. That’s a company that won the argument — the White House never produced public evidence contradicting Anthropic’s technical assessment — and concluded that being right was beside the point. When you can be switched off with ninety minutes’ notice, the merits of your case are decoration.

My colleague Jason Hubbard saw this mechanism coming before the confiscation even happened, back when the “voluntary” review order was just ink:

“Pre-release review regimes that quietly turn into soft licensing.” — Jason Hubbard, When the Ump Buys the Team

Seven words in June. A signed reality by July. Nobody passed a licensing law. Nobody had to. You just demonstrate the off-switch once, and every lab in America starts scheduling its Washington checkups voluntarily.

The Trillion-Dollar Reason Nobody Fought Harder

Why would the company that publicly called the government’s case overblown hand over this much? Fortune said the quiet part with unusual clarity: Anthropic has filed SEC paperwork for an IPO at a roughly trillion-dollar private valuation, and open hostility with the administration could stand in the way — “as would the overhanging risk that the government could step in to turn Anthropic’s models off whenever it sees fit.”

There it is. The bubble and the backlash, converging in one sentence. A valuation built on limitless promises now structurally requires staying friendly with the referee. Jason Hubbard framed this months ago in blunter terms: a business plan that quietly assumes the government will have to backstop you doesn’t make you a bold innovator — it makes you an off-balance-sheet liability waiting for a crisis. Anthropic just proved the inverse is also true. When the government can pull the plug, staying in its good graces stops being politics and becomes fiduciary duty.

Now stack Part 1’s Timeshare front on top: the same administration holding the off-switch has floated taking equity stakes in these companies. Rule-writer, early-access customer, prospective shareholder — and now, demonstrated repo man. Every incentive in that structure points the same direction, and none of them point toward you and me getting the unfiltered version of anything.

To be fair to Anthropic — and fairness is cheap when the facts do the work — they were the only ones who pushed back at all. Dario Amodei reportedly resisted administration demands, including, per Politico, requests to voluntarily pull Fable off the market entirely. They disputed the threat assessment in writing while their IPO paperwork sat at the SEC. That took spine. It just wasn’t enough spine to change the outcome, which tells you everything about where the leverage actually sits.

This Is the Template Now

If this were one company’s turbulence, fine. It isn’t. Look at the pattern that locked in over the past week:

OpenAI: GPT-5.6 remains limited to government-approved partners, vetted customer by customer — the first time in American history the government preemptively restricted a model’s release. OpenAI’s own statement: “We don’t believe this kind of government access process should become the long-term default.”

Meta: the last major U.S. lab without a voluntary model-sharing agreement, pressed by the administration last week to start submitting its models.

The framework itself: still doesn’t exist. Trump’s June 2 executive order promised a voluntary 30-day review process. Three weeks later there are no finalized benchmarks, no formal submission process, and — by Anthropic’s own admission — no agreed standard for even classifying how severe a jailbreak is. The measuring stick hasn’t been built, but the measuring has very much begun.

Brad Carson runs Public First — a bipartisan super PAC that is explicitly pro AI safety, so nobody gets to wave him off as industry spin. His description of the current process: “ad hoc, personalized, opaque, and possibly lawless.” He added that governments recalling dangerous products is appropriate — but it has to happen “with transparency and basic fairness.”

Ad hoc. Personalized. Opaque. That’s not a critic describing a policy. That’s a supporter of regulation describing the absence of one.

And that absence may be about to fill fast. As of this afternoon, the Financial Times reports the White House is accelerating its plans for AI model standards — with guidance that could land as early as next week, following the government’s interventions in the Anthropic and OpenAI rollouts.

Notice the sequence, because it’s the whole game. First you demonstrate the off-switch on two companies. Then you announce the rulebook — after the industry has already learned, in public, what happens to labs that don’t cooperate. That’s not writing standards. That’s writing them once everyone’s agreed in advance to sign whatever you put in front of them. The negotiating leverage got established in June; the paperwork arrives in July.

And who’s building the missing measuring stick? Anthropic announced it’s developing the jailbreak severity framework with Amazon, Google, and Microsoft. Sit with that. The companies being regulated are writing the ruler their regulator will use.

Jason Hubbard ran the whole review process through his own published engineering standard today — the checklist he uses to decide whether a workflow is even allowed to ship — and it failed every test. His verdict is the one sentence this entire episode comes down to:

“You don’t ship a gate without defined criteria and documented collapse behavior. Not in a CI/CD pipeline. Not in an agentic system. Not in federal AI governance... a gate without criteria isn’t a safety mechanism. It’s a permission slip for whoever holds the authority to make the call.” — Jason Hubbard, The Gate With No Test Suite

That’s the engineering translation of everything above. I made the political case that this is a template; Jason proved the template has no test suite behind it. Read his piece next — it’s the other half of this one.

And here’s the company that just won its argument and lost its leverage, marking the moment.

If the post strikes a grateful or relieved note — and relief is the natural tone when your product goes back on sale — sit with what that gratitude is for. Not a court win. Not a repealed rule. Permission. A company thanking the government for the right to ship the thing it already built. That’s the template, and everyone in the industry just watched it work.

The Sorting Hat, Updated

Last week I told you what the recall button really was:

“The recall button didn’t malfunction. It worked exactly as designed. It was never a kill switch. It was a sorting hat.” — Eric Mitchell, You’re Already Inside the Vault. Now What?

Today the sorting hat published its criteria. You get sorted back into access by agreeing to pre-release review, threat-intel sharing, embedded federal partnerships, and a reevaluation clause with no expiration date. Compliance is the house you get sorted into. The alternative house has ninety minutes’ notice and no appeal.

Here’s what I want you to carry out of this series so far. The bubble means these companies cannot afford to fight — trillion-dollar valuations don’t survive wars with the referee. The backlash means the government doesn’t need legislation — the off-switch demonstration did what no bill through Congress ever could. Put the two together and you get the quietest possible version of state control over the most powerful information technology ever built: no law passed, no vote taken, no court ruling. Just a letter, a leash, and a thank-you post.

The models are back. The precedent isn’t going anywhere.

Part 3 is coming — because the framework they promised is now, per the FT, arriving as early as next week. The fight over who writes it, what it measures, and who it locks out will make these three weeks look polite. The leash came first. The rulebook comes next. We’ll be reading it line by line.

Read the engineering companion to this piece: Jason Hubbard’s The Gate With No Test Suite breaks down why the government’s review process fails as a system, not just as policy. Part 1 made the political case; Jason made the technical one; together they’re the full picture.

THE BUBBLE AND THE BACKLASH — FOUR-PART INVESTIGATION

PART 1: The Bubble and the Backlash
THE ALLIANCE CRACKED
Washington and Silicon Valley start fighting over control.

YOU ARE HERE → PART 2: The Vault Got a Revolving Door
THE DOOR CAME WITH TERMS
Anthropic gets access back, but Washington retains leverage.

PART 3: The Tell
THE $42 BILLION TELL
OpenAI reportedly floats giving Washington equity as political risk becomes impossible to ignore.

PART 4 / FINALE: The Bubble and the Backlash: What It Cost
THE RECEIPT CAME DUE
Washington built the control system. Veterans, citizens, and builders absorbed the cost.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms

Nothing says “trust the process” like the process turning on itself in real time.

When the referee starts asking for a cut of the betting pool and the playbook, you’re not watching a fair game anymore. You’re watching a merger.
— Jason Hubbard, When the Ump Buys the Team

Previously on Sacred Loop: we told you the vault was real. We told you the recall button works. We showed you the Lutnick letter — the receipt proving the “national security” risk evaporates for roughly 100 hand‑picked “trusted partners” while everyone else waits.
You’re Already Inside the Vault. Now What?

This week, the people who built the vault started fighting each other over who gets to hold the keys.

That’s not me reading tea leaves. That’s the actual headline from Axios on Monday: “The pro‑AI movement is splintering.” When the people who created this mess start publicly turning on each other, you know the architecture is under real pressure. Let’s get into it.

Front Zero: The Timeshare on a Sinking Ship

Start here, because this number makes everything else make sense.

The same week the administration was deciding who gets access to the most powerful AI models, it floated something even bigger: buying equity stakes in the AI companies it’s supposed to regulate.

“The regulatory body responsible for overseeing an industry just announced it might become a financial stakeholder in that same industry, the very definition of a conflict of interest. That alone is enough to scream bad idea! That’s the whole thing. We can stop there and it’s already a five‑alarm governance catastrophe.”
— Jason Hubbard, Trump’s Decided to Buy a Timeshare on the Titanic

Jason did the math, and the math is ugly. Nine major AI players raised $122 billion from bond markets in 2025 alone. OpenAI is projecting roughly $14 billion in losses on $13 billion in revenue this year. Sixty‑one percent of all global venture capital is now flowing into a single sector. And the United States is sitting on $36 trillion in gross national debt, with $9–10 trillion maturing this year at rates nearly double what we’ve been paying.

“I have to be missing something here. Right?!? … The country is not in a position where ‘let’s also splash the pot and stock up on shares of structurally unprofitable AI companies’ should come up outside of a not‑so‑funny joke.”
— Jason Hubbard, Trump’s Decided to Buy a Timeshare on the Titanic

Keep that $36 trillion number in your head. It’s the backdrop for everything below.

Front One: Even Their Own Guy Is Calling BS

Here’s where it gets genuinely interesting, because this isn’t Sacred Loop saying the policy is broken. This is David Sacks — Trump’s own former AI and crypto czar — saying it on the record.

Axios reported Monday that the pro‑AI coalition that helped get Trump elected on a “win the AI race” platform is now openly fracturing.gov.ca

“A year ago, President Trump declared that America was in a global AI race and that the way to win it was to be pro‑innovation. President Trump was exactly right. We deviate from that strategy at our peril.”
— David Sacks, via X, reported by Axiosgov.ca

That’s not a Democrat. That’s not a critic. That’s the architect of the administration’s own AI strategy, publicly telling the administration it’s eating its own playbook.

Here’s the fuller picture, because Sacks’s actual position is sharper than a one‑line warning. He didn’t just defend Anthropic on principle — he criticized the company’s “minimizing language” around the underlying jailbreak, describing the vulnerability as enabling the “operability of a cyber weapon.” Sacks has spent months accusing Anthropic of regulatory capture and fear‑mongering. So when even Sacks — no friend of Anthropic — is warning that the broader restriction strategy risks American competitiveness, that’s not industry spin. That’s someone who thinks Anthropic deserved scrutiny saying the government’s response went further than the threat justified.

Even more telling: Trump himself walked it back.

“Well, not now, but a week ago, maybe.”
— President Trump, on whether Anthropic is a national security threat, via Axiosgov.ca

Read that twice. The President of the United States just admitted the national security threat that justified shutting down the strongest public AI model worldwide had an expiration date measured in days, dependent entirely on how fast the company complied with him personally. He praised Anthropic for responding “very quickly” and “responsibly.”gov.ca

That’s not national security policy. That’s a grudge with a timer on it.

BREAKING, as of today: Axios and Politico are now reporting the administration is close to lifting restrictions on Fable 5 entirely — possibly as soon as this week — with Anthropic co‑founder Tom Brown personally negotiating with Lutnick’s office. Only the Pentagon and NSA still have to sign off. Four months of standoff, a worldwide shutdown, a partial Mythos restoration, and now a full reversal under negotiation — all inside about three weeks. If you want a single data point for how fast “national security” can become “never mind,” this is it.

The trigger was simple: the same week Anthropic’s Fable 5 and Mythos 5 got suspended worldwide, the White House asked OpenAI to delay the broad rollout of its new GPT‑5.6 model too — releasing it in stages instead, “customer by customer,” until the government approves a general release. Two competing labs. Same treatment. Same control mechanism.

“We’ve made clear to the US government that this is not our preferred long term model, and will work with them and others in industry to achieve a more sustainable approach for future releases.”
— Sam Altman, internal memolablab

Translation: even the company playing ball thinks the current process is broken.

“The AI party now has a hall monitor who is also diluting the punch.”
— Paul Kedrosky, venture capitalist, via Axiosgov.ca

Investors are calling the deployment delays “hugely bearish” — meaning real money is now pricing in the risk that Washington, not the market, decides when and whether your product ships.gov.ca

And it’s not just industry voices. Congress is split too, and not cleanly down party lines.

“The United States, not China, needs to set standards for trusted AI.”
— Rep. Andrew Garbarino (R‑N.Y.), Chairman, House Homeland Security Committeegov.ca

Garbarino backed the national‑security framing but drew a sharp line, warning the response shouldn’t “unintentionally disadvantage American companies, allied partners, or critical infrastructure defenders.” That’s a Republican committee chairman telling his own administration: don’t win the security argument by losing the competitiveness war.gov.ca

On the other side, Democratic Ranking Member Bennie Thompson warned against “an ad hoc approach where decisions are made by political appointees in the White House rather than experts.” Independent Senator Angus King called the whole episode “pretty extreme,” openly questioning whether it was driven by the administration’s “otherwise announced antipathy” toward Anthropic rather than hard evidence.nhpr+1

When a Republican committee chairman and a Democratic ranking member are both telling the same administration it’s handling this wrong — just from different angles — that’s not partisan noise. That’s a bipartisan signal that the process itself has no credibility left.

And here’s the kicker that should make every patriot’s blood pressure spike: while the U.S. throttles its own labs, open‑source Chinese models are surging. Axios confirms Chinese models now occupy several top spots on OpenRouter’s usage leaderboard, driven by companies trying to minimize AI costs while American access gets rationed behind closed doors. The policy sold as “beating China” is functionally handing China the on‑ramp.gov.ca

This is precisely the asymmetry Sacred Loop flagged weeks ago.

“AI is dangerous when you use it. It’s patriotic when the government does. It’s criminal when China does.”
— Eric Mitchell, The AI Vault Is Real

Except now we’re watching American companies hand China the market share while we argue about who gets to look inside the vault.

Not everyone in the industry is purely against government involvement, to be fair. Anthropic itself has urged stronger safeguards as models grow more capable.gov.ca

“The government being involved here is actually super important. They just need to find the right balance between safety and broad access.”
— Dan Shipper, CEO of Every, via Axiosgov.ca

That’s a real position, held by people who aren’t trying to dodge accountability. The fight isn’t innovation versus safety. It’s clear rules versus a system where one phone call from the wrong person changes everything overnight.gov.ca

“It’s hard to build when there’s a moving target.”
— Mark Pincus, Zynga founder, investor in both OpenAI and Anthropic, via Axiosgov.ca

That’s the whole indictment in one sentence. Nobody — not the labs, not the investors, not even Trump’s own allies — knows what the rules actually are from week to week.

Front Two: The Bill Is Already Here

While Washington fights over the vault, the bill for building it is landing on real companies, in real dollars, right now.

The Information reported this week that Amazon could end up paying significantly more for Anthropic’s Claude models under a renegotiated partnership. The shift: from a compute‑hours pricing model to a token‑based model — meaning costs now scale directly with usage rather than fixed infrastructure spend. Claude already powers core pieces of Amazon’s shopping tools, coding assistants, and workplace products. Amazon is reportedly already evaluating hedges, including leaning harder on its own in‑house Nova models.

Think about the position this puts Amazon in. They built core product infrastructure on Claude. Now the pricing model underneath that infrastructure is shifting in a way that could meaningfully raise their bill — at the exact moment the model they’re paying for has been subject to federal export‑control suspensions, restorations, and “trusted partner” tiering. You can’t build a stable business on top of infrastructure the government can switch off on a Friday afternoon.

Meanwhile, on the other side of the ledger: OpenAI engineers reportedly discovered optimizations that more than halve inference costs — a genuine technical breakthrough, since inference (the cost of actually running a model for users) is one of the largest ongoing expenses in this entire industry.

Here’s why that matters in this context, and why I’m not letting OpenAI off easy for it: a 50%+ cost reduction is exactly the kind of internal engineering win that should be making American AI more competitive on a global stage. Instead, it’s happening in the same week the administration is staggering GPT‑5.6’s release “customer by customer.” The labs are innovating. The policy layer is the bottleneck. Genuine technical progress, sitting behind a government waiting room.

The Turn: Who Actually Pays for This

Pull the camera back.

The federal government is weighing equity stakes in companies it’s also threatening to blacklist overnight — while sitting on $36 trillion of debt it can barely service. Trump’s own AI strategist is publicly warning the administration it’s sabotaging its own stated goal. Investors are pricing in policy risk that didn’t exist eight weeks ago. Amazon is facing a costlier Claude contract layered on top of federal access uncertainty. OpenAI found real efficiency gains and still can’t ship its own product on its own timeline.

Everybody in this story is paying a tax that has nothing to do with building better AI. They’re paying for the unpredictability of a government that can’t decide if it’s regulating, investing, or just freelancing national security policy from week to week.

That’s not strategy. That’s chaos wearing a flag pin.

THE BUBBLE AND THE BACKLASH — FOUR-PART INVESTIGATION

YOU ARE HERE → PART 1: The Bubble and the Backlash
THE ALLIANCE CRACKED
Washington and Silicon Valley start fighting over control.

PART 2: The Vault Got a Revolving Door
THE DOOR CAME WITH TERMS
Anthropic gets access back, but Washington retains leverage.

PART 3: The Tell
THE $42 BILLION TELL
OpenAI reportedly floats giving Washington equity as political risk becomes impossible to ignore.

PART 4 / FINALE: The Bubble and the Backlash: What It Cost
THE RECEIPT CAME DUE
Washington built the control system. Veterans, citizens, and builders absorbed the cost.

Eric Mitchell is a Marine Corps veteran, former Silicon Valley insider, and former national television political analyst whose work has appeared on NewsNation, CNN, BBC, and MSNBC. He is the Chief Marketing Officer of Sacred Loop AI and Chief Editor of The Control Grid, where he investigates artificial intelligence, Big Tech, government power, and the financial incentives shaping the future of technology. His reporting follows one principle: Follow the evidence. Explain the system.

Follow the Investigation

The conversation doesn’t end with one article.

X: https://x.com/EricLMitchell

LinkedIn: https://www.linkedin.com/in/ericlmitchell

For breaking investigations, behind-the-scenes research, and analysis that doesn’t make it into every article, follow along across all three platforms