The loan would come not from a bank but from an office inside the Defense Department that most of Wall Street has never heard of. The Pentagon is negotiating a loan of about $5 billion to Fluidstack, an AI cloud provider that runs high-performance GPU clusters for training and inference, according to people familiar with the matter. The money would flow through the Office of Strategic Capital, a unit created to steer financing toward critical technology supply chains.
If the deal closes, it would be the largest loan the office has ever made, according to the people. The Wall Street Journal reported the talks on Thursday.
Fluidstack, founded in 2017 and based in New York, has grown into one of the operators that big AI developers lean on when their own capacity runs short. The company has said its customers include some of the largest model builders, and it has expanded quickly by assembling fleets of the most advanced chips on short timelines. It raised a large round of financing in recent years that valued it in the billions, and it has positioned itself as a fast alternative to the hyperscalers’ own clouds.
The purpose of the money is narrower than building new data centers. The people said Fluidstack is weighing whether to use the financing to strengthen the domestic supply chain and manufacturing capability for specific components that data centers depend on, rather than adding raw computing capacity. That framing matters: it would make the loan an industrial-policy instrument as much as a financing deal.
The application was advised by Erebor Bank, a lender founded by Palmer Luckey, the defense entrepreneur who built Anduril. The choice of adviser signals how closely the deal sits to the defense-industrial world rather than conventional tech finance, where government-backed loans to AI providers are still rare.
Terms, including the interest rate, have not been set, and the loan has not received final approval. Officials could still walk away or restructure the arrangement.
The Office of Strategic Capital was stood up in 2023 with a mandate to use loan guarantees and financing to pull private capital into technologies the military needs, from advanced batteries to microelectronics. Its early activity was modest by Pentagon standards; a $5 billion loan to a commercial AI cloud provider would mark a step change in its ambition.
Analysts said the talks reflect a shift in how Washington thinks about AI infrastructure. The Pentagon and its allies have framed advanced computing as a strategic asset on the order of munitions or shipyards, and loaning money to a GPU operator treats the compute supply chain as part of the defense industrial base. The Chips Act and related programs pushed the same logic in semiconductors; this loan would extend it to the operators of the machines themselves.
For Fluidstack, a government backstop would lower the cost of capital at a moment when data-center construction is expensive and private financing has tightened. A five-billion-dollar facility would also give it the balance sheet to lock in long-term chip and power commitments, which have become the real currency of the AI buildout.
For the Pentagon, backing a commercial operator is a way to secure access to computing power without owning and running it directly, an arrangement that is cheaper and faster than standing up its own facilities. The trade-off is oversight and control, questions the two sides would have to settle in final terms.
The talks also reflect how quickly Fluidstack has scaled. The company has secured large allocations of the most advanced chips and has marketed itself as a faster alternative to the big cloud providers, with clusters that can be stood up in a fraction of the time a hyperscaler needs to build a campus. Its ability to keep signing such commitments depends in part on financing that private lenders are increasingly cautious about extending.
National security officials have privately worried that the US computing supply chain is too concentrated in a handful of firms and too dependent on parts made overseas. A loan aimed at domestic manufacturing of the components data centers need, rather than at more data centers themselves, would address that concern directly, which is part of why the Pentagon has engaged at all.
Whether the loan proceeds will depend on terms the two sides have not yet agreed on, including pricing, collateral, and what the Pentagon gets in return for backing a commercial operator. People familiar with the talks said those details could take months to settle.
The deal is not done, and government loans of this size move slowly. But the fact that a defense office is in the room at all, people familiar with the talks said, shows how far AI compute has moved from a Silicon Valley niche toward something the military now budgets for like any other weapons system.
