The Pentagon Just Underwrote Fluidstack — And CoreWeave's $35.6B Debt Stack Suddenly Looks Optional

The Wall Street Journal reported on Sept 10, 2026 that the Pentagon's Office of Strategic Capital (OSC) is in talks to lend roughly $5 billion to Fluidstack, a UK-headquartered AI cloud startup that already controls more than 100,000 GPUs and operates data centers in Texas and New York. The figure alone forces a re-read of who funds American AI infrastructure. OSC's direct-loan cap today sits at $150 million per project; the program's largest conditional commitment to date — $820 million to drone-maker Performance Drone Works on July 31, 2026 — is roughly one-sixth the size of what is now on the table for a single private AI-cloud tenant.

The mechanism is not a commercial loan. OSC's pilot authority, established in 2024 and re-upped under P.L. 119-21, allows the Pentagon to provide direct debt, loan guarantees, and technical assistance to companies building dual-use technologies in 34 covered categories — a list that now explicitly subsumes AI infrastructure. FY2026 funding for the program came in below $1.5 billion. The Pentagon's FY2027 request, sent to Congress in April 2026, asks for $20.2 billion in new OSC credit authority, layered on top of a separate $46 billion mandatory \"Sovereign AI Arsenal\" line and a $30 billion AI-supercomputing buildout. Fluidstack, in other words, is not a one-off bet — it is the first visible deployment of a much larger war chest.

Fluidstack's operating footprint — not its balance sheet — is what made it bankable to the Pentagon. The company manages more than 100,000 GPUs across its network and serves named AI labs including Anthropic, Meta, Mistral, Poolside, Black Forest Labs and Character.AI. Its largest commercial anchor is a $50 billion commitment from Anthropic (announced Nov 12, 2025) to build data centers in Texas and New York, with first sites already coming online through 2026. The company is projecting2026 revenue of about $660 million — roughly tripling the $200 million it booked in 2025 — and added a $1.5 billion private-equity round led by Jane Street at an $18 billion valuation on Sept 3, 2026, bringing total disclosed funding to more than $2.6 billion.

Three details sharpen the Pentagon's logic. First, Fluidstack already has hyperscaler-grade counterparties on long-dated contracts, including a 25-year lease structure on its largest U.S. site — the kind of cash-flow profile federal credit can underwrite without equity dilution. Second, in March 2026 the company publicly abandoned an $11.5 billion project in northern France to concentrate capital in the United States, a signal of strategic alignment with Washington's industrial-base priorities. Third, Google has effectively pre-vetted the credit on the Abernathy campus by agreeing to cover Fluidstack's lease obligations if the startup defaults — a structure the OSC can now replicate at scale with federal balance-sheet capacity behind it.

The listed AI cloud sector has spent18 months trading on two narratives: a $100B-plus order book and a debt-financed capex super-cycle. CoreWeave's Q2 2026 results put both under fresh pressure. Revenue doubled to $2.575 billion, up 112% year-on-year, and contracted backlog ballooned to $104.2 billion — but three customers accounted for 72% of the quarter's revenue, and total indebtedness reached $35.6 billion against an operating loss. Net interest expense alone was up 140% to $640 million. Capex guidance for 2026 sits between $35 billion and $39 billion, funded by $13.5 billion of debt and roughly $1 billion of stock issuance.

A Pentagon loan to a competitor does not change CoreWeave's backlog or its customer list overnight. It does change the discount rate at which the market should value that backlog. Federal credit below Treasury rates — what OSC is structured to deliver — is cheaper than 10–15% GPU-backed term loans, and it removes the equity-dilution overhang that has hung over CoreWeave's $35.6B debt stack since IPO. For the first time, a U.S. AI-cloud operator will operate with a sovereign-style cost of capital. Listed peers cannot match that without an explicit federal match, which puts pressure on CoreWeave's roughly 12.5x trailing EV/sales multiple and on Nebius's 44x — the latter now has to defend why a smaller, less-diversified backlog warrants a premium.

The Pentagon underwrites Fluidstack, but the marginal dollars still flow through the same physical supply chain — and three listed names already have skin in the Fluidstack build, with no competing sovereign-backed bidder in sight. TeraWulf is the most exposed. Its October 2025 joint venture with Fluidstack at Abernathy, Texas commits the two companies to deliver 168 MW of critical IT load (240 MW gross) on a 25-year lease, representing approximately $9.5 billion of contracted revenue to the JV, with Google backstopping $1.3 billion of Fluidstack's lease obligations. In July 2026, TeraWulf also signed a separate 20-year lease directly with Anthropic at its Justified Data campus and sold a majority interest in the Abernathy JV to Fluidstack — effectively converting TeraWulf into the real-estate and power layer underneath a sovereign-backed tenant.

NVIDIA is the simplest read: more GPUs deployed by more competitors, regardless of who carries the balance sheet, means more accelerator demand. Vistra is the indirect read: Fluidstack's Texas campuses — including Abernathy and any expansion sites — need gigawatts of dispatchable generation in ERCOT, and Vistra controls 38,700 MW across nuclear, gas, coal, solar and battery. Power remains the binding constraint on the AI buildout, and the Pentagon's loan does not relax ERCOT's interconnect queue — it simply ensures the project that gets built first has sovereign credit. Applied Digital is the awkward peer: 1.2 GW of leased capacity and a $31 billion contracted backlog are real, but the company has no federal credit enhancement, putting it on the wrong side of the funding gap the OSC loan just opened.

Three dates will tell investors whether the OSC loan is symbolic or a new template. First, whether the Pentagon signs a final term sheet before the end of FY2026 — current program funding caps the headline size unless Congress approves the $20.2 billion FY2027 ask. Second, whether Fluidstack passes muster on the standard OSC criteria (national-security alignment, hastening of outcomes, creditworthiness), with the $50 billion Anthropic anchor and Google's existing lease backstop making that test largely procedural. Third, whether the next round of $30 billion in FY2027 AI supercomputing procurement flows preferentially to operators with an OSC credit relationship — the same way SBIR winners compound follow-on federal work.

Over a one-to-three-year horizon, the bigger question is whether the OSC template extends to other neocloud operators — Nscale, Crusoe, even Nebius's U.S. buildout — or stays a single bespoke arrangement. Congressional scrutiny is already visible: a July 2026 proposal would impose restrictions on data centers built on DoD land, and a separate Anthropic–Pentagon dispute over usage policy has dragged into the courts. That political friction is the main downside risk to the thesis. The upside is that the Pentagon has just demonstrated it will underwrite the balance sheet of any AI operator that meets its dual-use bar — which structurally lowers the cost of capital for the entire sector while raising the bar for the listed players whose entire investment case was that they were the only ones who could finance this buildout.