From Rare Earths to GPUs: The Pentagon's $5 Billion Loan Pitch to Fluidstack Would Be Its First Into AI Compute

The office inside the Pentagon that was built to keep America supplied with rare-earth magnets and battery chemicals has picked its next potential target — and it is not a mine. According to a Wall Street Journal report published September 10 and echoed by Reuters, the Department of Defense’s Office of Strategic Capital (OSC) is in talks to lend roughly $5 billion to Fluidstack, the AI “neocloud” that leases GPU and TPU data-center capacity to the largest AI labs in the world. If the loan closes, it would be the first time this specific defense financing office has reached past physical minerals into AI compute supply chains — a crossover that analysts say could set the template for how Washington funds the infrastructure layer of the AI economy.

The lender: an office built for things you can touch

The Office of Strategic Capital exists to secure the physical inputs of the defense industrial base, and every deal it has signed so far has financed something tangible. It has committed $1.95 billion in conditional loans across battery maker Sila Technologies, rare-earth magnet producer Niron Magnetics and Australia’s Sunrise Energy Metals. It put $85.5 million into a bauxite mine in Guyana through the Industrial Base Analysis and Sustainment program, took a 10 percent stake in Trilogy Metals’ Alaska minerals project for $35.6 million, and became the largest shareholder in rare-earths producer MP Materials.

Against that ledger, a loan to Fluidstack would be categorically new. Fluidstack owns no mine and makes no chip. What it owns is gigawatt-scale data-center capacity — the buildings, power contracts and cooling systems inside which Nvidia GPUs and Google TPUs actually run — and the Pentagon appears to have concluded that this capacity is now as strategically essential as the minerals it has spent two years underwriting.

The loan, as reported, is meant to shore up the U.S. data-center supply chain and its component manufacturing capacity rather than fund one specific facility outright. If it closes, it would support the same civilian compute build-out that Google and Anthropic already pay for commercially, sitting underneath AI training and inference capacity rather than a weapons program.

The borrower: a neocloud whose customers are Google and Anthropic

Fluidstack’s rise has been one of the fastest in the AI infrastructure boom. The company raised a $200 million Series A in February 2025, closed $450 million more in January 2026, and by September 3, 2026 had reached an $18 billion valuation after a $1.5 billion round led by quant fund Jane Street. A leaked investor memo reported by Forbes projected Fluidstack’s revenue growing to roughly $660 million in 2026 — more than tripling from about $200 million the year before.

The demand that built that valuation came from two of the largest AI labs, not from the Pentagon. In November 2025, Anthropic signed a $50 billion agreement with Fluidstack to build data centers custom-designed for its needs in Texas and New York. By April 2026, Fluidstack held $6.7 billion in contracted revenue backstopped by Google, which has also discussed investing roughly $100 million as part of its own push to lease out TPU accelerators through third-party data-center partners. Its largest customers are Google and Anthropic — not the Department of Defense.

The advisor: a bank that did not exist a year ago

Perhaps the most striking detail in the Journal’s reporting is who is advising Fluidstack’s loan application: Erebor Bank, the hard-tech-focused bank that Anduril founder Palmer Luckey opened in February 2026 with backing from Palantir co-founder Joe Lonsdale. By the time this loan was reported, Erebor had already gathered more than $4.6 billion in deposits. Erebor was built to bank founders and companies working on defense, energy and other capital-intensive physical technology that traditional lenders often avoid — and its role advising a Pentagon-adjacent AI transaction puts a bank not yet two years old at the center of the first crossover between rare-earths-style defense financing and commercial AI compute.

Why this matters

Three implications stand out.

First, compute is now officially strategic. The OSC’s mandate covers the defense industrial base’s most critical physical inputs. Extending that logic to GPU data centers formalizes what national-security officials have been saying for two years: AI compute capacity is a strategic asset, on par with rare earths and batteries, and Washington is willing to lend directly against it.

Second, it signals how AI infrastructure will be financed going forward. AI data centers are extraordinarily capital-intensive, and traditional project finance has struggled to keep pace with the scale of demand. If the Pentagon becomes a willing lender at the $5 billion scale — and if this is “not a one-off,” as one analysis put it — it sets a template where American AI infrastructure vendors gain access to defense-grade credit lines that foreign competitors bidding for the same hyperscaler contracts do not have. European officials watching from Brussels will read it as a signal to accelerate their own Chips Act successor programs and IPCEI funding rounds.

Third, the Anduril-adjacent financing ecosystem is maturing fast. A bank founded by Palmer Luckey, backed by Joe Lonsdale, advising a Pentagon loan to a neocloud serving Google and Anthropic — that is a complete parallel financial stack for the AI-defense-industrial complex, assembled in under two years.

The caveats

Neither the Pentagon, Fluidstack nor Erebor has issued a public statement confirming the deal, and Reuters said it could not independently verify the Journal’s reporting. The size, terms and even whether Fluidstack ultimately accepts defense financing could still change. That is the normal state of an in-talks story — but worth bearing in mind before treating the loan as settled policy.

What is not in doubt is the direction. A defense financing arm that spent two years treating rare-earth magnets and battery cells as strategic assets worth underwriting directly is now willing to apply the same logic to AI compute capacity. If you build, finance or supply AI infrastructure, treat data-center supply-chain financing as a live line item in U.S. AI infrastructure policy — and expect more announcements like this one over the next year, not fewer.