The largest loan the Pentagon’s Office of Strategic Capital has ever made was $725 million, to a rare earth producer called Energy Fuels. Its second largest was a $400 million commitment to a scandium miner in Australia. The office exists to keep supply chains that Washington considers strategic, mostly critical minerals and drones, from depending on China.
On September 10, the Wall Street Journal reported that the same office is in talks to lend roughly $5 billion to an AI cloud startup named Fluidstack. That is not a rounding difference. It is nearly seven times the office’s previous record, aimed at a company most people outside the data-center business have never heard of, in a sector the office was never built to fund. Reuters said it could not immediately verify the figure, and the talks are early, so treat the number as reported rather than signed. But the direction is the story, and the direction is clear.
The money is not for chips, and not for a data center
The first assumption to drop is that this is the government buying GPUs or bankrolling a new AI facility. It is not. According to the reporting, Fluidstack would use the loan to expand domestic production capacity for certain data-center components and to shore up the U.S. supply chain, not to construct a new site outright.
That distinction matters more than the headline. The binding constraint on AI stopped being chips a while ago. It became power, which is why this site has spent 2026 tracking an $87 billion data-center build in France chosen for its nuclear supply and the first utility mega-merger built specifically to feed AI load. Now the constraint is moving one layer deeper still, to the physical hardware that delivers that power: transformers, switchgear, high-voltage cabling, the grid-interconnect equipment that a gigawatt-scale campus cannot run without and that currently has lead times measured in years.
The policy context makes the target explicit. Last month President Trump signed an executive order declaring a national emergency and barring some foreign-made equipment from the U.S. power grid that data centers depend on. A loan to build domestic capacity for grid and data-center components is the financing arm of that order. Washington is treating the guts of power delivery the way it already treats rare earths and battery materials: as a supply chain too important to leave offshore.
Who Fluidstack actually is
Fluidstack is not a household name, but it sits closer to the center of the AI buildout than most companies that are. Founded in London in 2017, it relocated its global headquarters to New York in December 2025. Forbes profiled it on September 3 under a headline that captures the position precisely: a tiny startup helping Google take on Nvidia, now worth $18 billion.
The specifics behind that valuation are worth stating plainly. Fluidstack raised an $830 million Series A at a $7.5 billion valuation in July 2026, then moved into talks for a further round, reported at $18 billion and led by Jane Street, only months later. An investment memo cited by Forbes has the company managing up to 1.3 gigawatts across more than ten sites this year, with revenue reaching $660 million in 2026 and an aspiration to run more than 17 gigawatts by 2030, which would exceed the current U.S. data-center capacity of Amazon and Microsoft combined. It manages more than 100,000 GPUs today for customers including Mistral AI, Character.AI, Poolside, and Black Forest Labs. Its pitch is speed: it claims it can deliver gigawatt-scale capacity in roughly six months against the 18 to 24 months a traditional hyperscale build takes.
Two relationships explain why the government cares. Fluidstack builds infrastructure for Google, part of Google’s effort to lean on its own TPUs rather than Nvidia. And Anthropic has listed a $50 billion U.S. infrastructure commitment with the company, financed in part through Apollo and Blackstone, with a first gigawatt already deployed. This is the same Anthropic that has assembled a five-vendor compute stack including 220,000 GPUs rented from SpaceX. Fluidstack is one of the physical layers underneath the frontier labs, which is exactly why a national-security lender would want it healthy and building on American soil.
The lender of last resort keeps moving up the stack
There is a pattern here that predates this loan, and it is the real reason to pay attention.
When the AI buildout outran what customers and normal debt markets could fund, the first backstop was the chip maker itself. In July this site described Nvidia’s reported $250 billion guarantee behind OpenAI’s Ohio campus as the chip maker becoming the lender of last resort. Then the pattern spread sideways into the vendors, with AMD taking equity in Anthropic so Anthropic would buy AMD chips and Google, Microsoft, and Nvidia writing checks into the customers and clouds that recycle straight back into their own revenue. The industry started, in one skeptic’s phrase, holding each other’s paper.
The Fluidstack talks are the next rung. The backstop is no longer a supplier, a hyperscaler, or a private-credit fund. It is the federal balance sheet. The Office of Strategic Capital has authority to deploy loans and guarantees at a scale the Pentagon has been expanding aggressively, requesting more than $20 billion for the program in its fiscal 2027 budget and opening it in August to private credit funds that agree to pool federal money with their own. A $5 billion single loan fits that trajectory even as it dwarfs everything the office has done to date.
That is a meaningful shift in who carries the risk of the buildout. Private capital prices AI infrastructure on the assumption that demand holds and the tokens keep flowing. When the government co-signs, it is making a national-security bet that the capacity has to exist regardless of whether any single lab’s economics work out. Those are not the same bet, and the taxpayer is now on one side of it.
The Palmer Luckey layer
One more detail says a great deal about how this deal is wired. Fluidstack’s loan application is being advised by Erebor Bank, the defense-and-crypto-focused institution founded by Anduril’s Palmer Luckey and 8VC’s Joe Lonsdale, with Peter Thiel among its reported backers. Erebor received a national bank charter in February 2026, the first granted under the second Trump administration, launched with $625 million in capital, and grew deposits to $4.6 billion by July while chasing a valuation near $9.5 billion. It was built to serve exactly this clientele: AI, crypto, defense, and advanced manufacturing companies operating close to government priorities.
So the structure is a defense-tech bank founded by a defense-tech entrepreneur advising an AI cloud on a loan from the defense department’s strategic-investment arm, to build components a Trump executive order just made a national emergency. None of that is illegal or even unusual for how strategic industries get financed. It is worth naming because it shows how tightly the AI infrastructure layer, the national-security apparatus, and a specific political-financial network have fused in 2026. This is the same fusion driving Europe’s scramble to fund its own gigafactories on sovereignty grounds, and it is the practical expression of the thesis that compute has become a proxy for national economic power.
What a buyer should take from it
I have spent more than twenty years in IT operations, much of it on the procurement and infrastructure side, and the part of this story that generalizes is not the politics. It is the components.
The lesson for anyone planning capacity, at any scale, is that the pacing item has moved. It is not the accelerator anymore, and increasingly it is not even raw power availability. It is the long-lead physical hardware that stands between a signed power contract and a running rack. When the Pentagon is willing to make its largest loan ever to expand domestic production of that hardware, it is telling you the shortage is real and structural, not a quarter-long blip. The same dynamic already showed up downstream in the memory-price surge that put the AI buildout on ordinary buyers’ invoices. Grid components are the next line item to watch.
Two practical implications follow. First, treat transformers, switchgear, and interconnect as procurement risks on par with silicon: order early, qualify more than one supplier, and price the schedule risk explicitly rather than assuming availability. Second, be honest about concentration. A market where the same handful of labs, clouds, chip makers, and now the government are all financing one another’s capacity is efficient right up until it is not. The circular-financing web that makes the buildout possible also means a shock in one place travels fast. Keep your own dependencies legible, and do not assume a backstop you did not negotiate will be there when you need it.
The Fluidstack loan may or may not close at $5 billion. What it already confirms is that the AI infrastructure race has reached the point where the federal government considers the picks and shovels a matter of national defense. The buildout is no longer just an industry. It is becoming policy, and the bill is starting to run through Washington.