The Pentagon Is Talking About Lending $5 Billion to a Company That Owns No Chips. The Money Is Not for Compute, It Is for Transformers.

The Wall Street Journal reported on Thursday, September 10, 2026 that the Pentagon is in talks to lend roughly $5 billion to Fluidstack, the AI cloud operator, through the Office of Strategic Capital. The headline everywhere this morning is the number and the borrower. Both are interesting. Neither is the story.

The story is the reported use of proceeds. Fluidstack would not spend the money building an AI data center. It would spend it building domestic manufacturing capacity for the components a data center needs, which reporting describes as power and cooling gear.

Read that again with the borrower in mind. Fluidstack owns no chips. It does not own the buildings it operates. It has never manufactured anything. And the Department of Defense is discussing handing it the largest loan its industrial base credit office has ever written, to go stand up a supply chain.

The Loan Book

The Office of Strategic Capital exists to put federal credit behind industries the Pentagon considers strategically load-bearing and that private capital underprices. Rare earths, magnets, batteries, drone components. It has been operating at a particular scale. The proposed Fluidstack facility is not that scale.

The last row is the one that reframes the first. A single borrower is in discussions for a facility roughly the size of everything OSC committed across every sector in the entire fiscal year, and close to seven times its largest individual loan on record. This is not an incremental step up the ladder. It is a different ladder.

The budget line says the same thing prospectively. The administration requested $20.2 billion for OSC in fiscal 2027 against $1.5 billion in fiscal 2026, a roughly thirteenfold increase in a credit office most people outside defense finance had never heard of. A $5 billion loan makes no sense against the 2026 office. It makes perfect sense as the first transaction of the 2027 one.

Two honest caveats before this goes further. First, this is WSJ reporting sourced to people familiar with the discussions, not a signed term sheet, and OSC deals routinely move as conditional commitments that take quarters to close or quietly do not. Second, the published details are thin on structure, tenor and conditions, all of which matter enormously for what the loan actually does. Treat the $5 billion as a direction, not a fact.

Why the Borrower Is Not a Manufacturer

The obvious objection is the right one. If the goal is domestic manufacturing capacity for transformers, switchgear, turbines and cooling systems, there are companies that already make those things. Lend to them.

The answer is that a factory is not financed against a machine. It is financed against an order book. Long lead time electrical equipment plants are capital intensive, slow to build, and catastrophic to build speculatively, which is exactly why the domestic capacity gap exists in the first place. Manufacturers have spent two decades watching demand for this gear arrive in unpredictable bursts and have sized their lines accordingly.

What Fluidstack brings is not a production asset. It is a signed, multi-year, very large demand commitment. The company is the operator standing up the sites where Anthropic's reported one million Google TPUs will run, under an arrangement reported at $50 billion, which made it the first publicly known operator of TPU capacity outside Google itself. An entity with that book knows, with more precision than any manufacturer does, how many megawatts of switchgear it needs delivered in which quarter.

So the Pentagon is not underwriting a factory. It is underwriting a purchase order, and letting the purchase order pull the factory into existence behind it. That is a coherent piece of industrial policy design and a genuinely unusual one, because it means the federal credit risk sits on a demand forecast rather than on a physical plant with salvage value.

Three Repricings in Nine Weeks

The other thing worth putting in one place is how fast the counterparty has moved. We covered the equity side of this four days ago, when Jane Street led Fluidstack at $18 billion while simultaneously sitting as the anchor customer behind a competitor's round.

Nine weeks, a 2.4x equity remark, and then a federal credit facility roughly three times the size of the round that set the mark. The loan application is being advised by Erebor Bank, the newly chartered federal bank whose client base skews defense and technology startups, which is its own small signal about how this category of transaction is being assembled now.

There is a recursion here worth naming plainly. The federal credit decision rests on Fluidstack's demand book. That book rests on a reported Anthropic commitment. And the company's equity price, the number any underwriter will look at first, was set four days earlier by a proprietary trading firm that is itself one of the sector's largest compute buyers. None of that is improper and all of it is public. It just means the number of independent opinions underneath a $5 billion federal exposure is smaller than four separate headlines make it look.

The Executive Order Underneath It

This loan did not appear from nowhere. On August 26, 2026, the President signed Executive Order 14420, declaring a national emergency over foreign-produced bulk-power system equipment and authorizing the Department of Energy to prohibit, condition or unwind transactions involving it.

The covered list is the important part: large transformers, high voltage circuit breakers, utility-scale inverters, battery energy storage systems, generation turbines and industrial control systems. Local distribution equipment is excluded. DOE has 120 days from the order to publish implementing rules, which puts the deadline around December 24, 2026.

Line the two events up. On August 26, Washington constrained the supply of exactly the equipment that gates a data center interconnection. Fifteen days later it was reported to be discussing the largest loan in its industrial credit office's history to expand domestic production of that same equipment. The order created a demand shock in a market with almost no domestic slack, and the loan is the supply-side answer. Fifteen days is not enough time to originate a $5 billion credit facility from scratch, which means the two were moving in parallel rather than one reacting to the other.

For anyone tracking the power constraint, this is confirmation of a thesis that has been building all year. The binding input stopped being accelerators a while ago. It is megawatts, interconnection queues and the eighteen to thirty-six month lead times on the heavy electrical gear that connects a building to a grid, a shift we have tracked through the FERC interconnection fights and the nuclear restarts. The Pentagon has now priced that constraint, and the price is five billion dollars.

Our Take

The number that matters is not $5 billion. It is zero, the number of transformers Fluidstack has ever built.

We do not think that is a scandal. We think it is a disclosure about how Washington has decided to run AI industrial policy, and it is worth stating explicitly because nobody involved will. The government looked at a capacity gap in heavy electrical manufacturing, concluded that the constraint was not engineering talent or capital but the absence of a creditworthy forward order, and went and financed the party holding the order instead of the parties holding the factories. That is a bet that demand certainty is the scarce input, not production capability.

It is also a materially different risk posture than OSC's prior book. A rare earth processing loan is secured against a plant, a mineral stream and an offtake agreement. A loan underwritten against a neocloud's forward book is secured against a contract with a frontier AI lab, which means the federal government has taken an indirect position on AI demand persisting through the tenor of the loan. Every previous OSC borrower could fail and leave behind an asset. This structure can fail and leave behind a spreadsheet.

We would also note what this does to the phrase national security AI infrastructure, which has been doing a lot of undefined work in policy documents this year. It now has a concrete referent, and the referent is not a model, a chip or a dataset. It is switchgear. The most consequential AI policy action of the week is about cooling loops and high voltage breakers, and it is going to be covered as a startup funding story.

What This Means If You Build on APIs

Directly, in the next two quarters: nothing. No inference price moves because of this. The serving fleets that price your tokens are already energized, and the cost curve under them is silicon and memory bandwidth, which is why DeepSeek could put a cache read at $0.003 yesterday in the middle of the tightest power market on record. The pricing floor is not set by transformer lead times.

Indirectly, on a 2028 horizon: this is the second consecutive quarter where the state has intervened on the supply side of compute rather than the regulatory side of models. If domestic grid equipment capacity actually expands, the ceiling on how much serving capacity comes online in 2028 and 2029 goes up, and more serving capacity has meant lower token prices every single time. If it does not expand, or expands behind schedule, the power constraint binds harder and capacity gets allocated by contract size rather than by price, which is the environment where small builders get deprioritized quietly. Worth watching which way that resolves before you commit to a 2028 cost model.

Three signposts for the next two quarters. First, whether the loan closes at all and at what size, since the gap between a reported discussion and a signed OSC commitment is where most of these stories end. Second, whether DOE's implementing rules under EO 14420, due around December 24, name specific equipment categories and country of origin thresholds, because the specificity of that rulemaking determines whether the demand shock is real or rhetorical. Third, whether a second AI infrastructure operator applies to OSC, which is the direct test of whether Fluidstack is a one-off or whether the Pentagon credit office has quietly become a standing financing channel for the AI buildout.

We track the compute supply side on our compute providers page. The next number that moves this story is a signed term sheet, or a DOE rule with real thresholds in it.