The Pentagon spent the past year learning how to buy artificial intelligence. It is now learning how to finance it.
The Wall Street Journal reported on Thursday that the Office of Strategic Capital, the Defense Department's in-house lending arm, is negotiating a loan of roughly $5 billion to Fluidstack, a nine-year-old AI cloud-computing company, to shore up the domestic supply chain behind American data centers. Reuters relayed the report the same evening and noted it could not immediately verify the details. Fluidstack's loan application is being advised by Erebor Bank, the hard-tech national bank founded by Anduril's Palmer Luckey, which opened in February and has already gathered more than $4.6 billion in deposits.
The money is not earmarked for a single gleaming campus. It would go toward U.S. manufacturing capacity for the unglamorous components that gate the buildout: power equipment, switchgear, transformers, cooling systems. The binding constraint on AI infrastructure is less chips than the multi-year queue for the hardware that makes a rack of accelerators useful, and Washington now treats that queue as a national security problem. Trump signed an executive order last month declaring a national emergency and barring certain foreign equipment from the U.S. electricity grid.
The scale problem
What is startling is not the figure in AI terms, where $5 billion is a rounding error beside a $664 billion cloud backlog. It is the figure relative to the lender. The Office of Strategic Capital, written into law by the fiscal 2024 defense authorization act, advertises direct loans of up to $150 million. Its first direct loan, announced in August 2025, was exactly that: $150 million to MP Materials for heavy rare earth separation at Mountain Pass, California — no collateral, interest pegged to the ten-year Treasury yield plus one point, twelve years to repay. A $5 billion facility would be more than thirty times larger than anything the office has written. The Pentagon has asked Congress for about $20.2 billion for the credit program in fiscal 2027, up from less than $1.5 billion this year.
Officials are explicit about the theory of the case. “The Chinese Communist Party has implemented aggressive actions to endanger the global critical minerals supply chain and ultimately undermine American economic and national security,” said Patrick Witt, the office's acting director, when the MP Materials loan was announced. The President, he added, “directed the DoD to use its authorities and capital tools to confront the threat posed by the CCP.” Under Secretary of Defense for Research and Engineering Emil Michael framed it as industrial revival: “Through this loan, we are taking decisive action to restore our domestic critical minerals supply chain, revive our industrial base, and rebuild our military.”
Swap rare earths for switchgear and the logic transfers cleanly. Whether the statute does is less clear. The program's published list of 31 eligible technology categories includes edge computing, data fabric, battery storage, solar and microelectronics fabrication. It does not include data centers.
Fluidstack is an unusual vehicle for the ambition. Founded in 2017 as an Oxford spinout by Gary Wu, Cesar Maklary and Jamie Cox, it began by renting idle GPU blocks from corporate and university labs and grew into a general contractor for other people's AI supercomputers. It owns essentially no chips. It closed a $1.5 billion round led by Jane Street on September 5 at an $18 billion valuation, more than double the roughly $7.5 billion it carried earlier this year. Anthropic named it anchor partner for a $50 billion U.S. compute buildout across Texas and New York; Google has backstopped $1.8 billion of its lease obligations.
What it means when the state becomes a lender
Since May, when the Pentagon put Nvidia, Microsoft, AWS, Google, OpenAI, SpaceX and Oracle onto classified systems, Washington has been AI's most prestigious buyer. A buyer can let a contract lapse. A lender is married to the borrower's execution, its counterparties and the demand cycle beneath the entire sector. If capacity stays scarce, cheap public credit will look like shrewd statecraft. If too many megawatts chase too few durable workloads, taxpayers will have helped finance the overshoot.
Then there is the question of who is in the room. Erebor won its charter in February to serve AI, defense, crypto and manufacturing clients; Luckey, its founder and a board member, also built one of the Pentagon's fastest-growing prime contractors. Advising a loan application is not improper. But a bank founded by a major defense supplier advising on the department's largest-ever loan is precisely the structure Congress has questioned before — Senator Elizabeth Warren wrote to the Pentagon in 2023 raising conflict-of-interest concerns about this very office.
Analysts press a sharper version of the worry. Of the private fund managers the Pentagon is routing credit through, Julia Gledhill, a research analyst with the Stimson Center's National Security Reform Program, said: “They are not necessarily primarily motivated, in terms of financial incentives, to create a successful business and fulfill these production promises that DoD is making.” There is, she argued, “a very high risk” of public funds serving private gain, and “taxpayers deserve more transparency and accountability if they are helping foot the bill.”
The defensible answer is that private capital is not the constraint on data centers, which are drowning in it, but on the factories that supply them, where payback runs five to ten years and no venture fund wants the exposure. That is a genuine market failure — and also an argument that works for nearly every industrial input in the economy, which is why the terms will matter more than the headline.
Watch three things. Whether the size and structure survive negotiation — the Journal stressed nothing is settled. Whether the office discloses rate, collateral and repayment terms as it did for MP Materials, or whether a $5 billion loan to a private company draws less sunlight than a $150 million one. And whether Congress funds the $20.2 billion request, which is the difference between a one-off and a standing federal lending desk for AI infrastructure.
“They are not necessarily primarily motivated, in terms of financial incentives, to create a successful business and fulfill these production promises that DoD is making.”— Julia Gledhill, Research analyst, Stimson Center National Security Reform Program