USD.AI provides $34M loan to finance 768 Nvidia GPUs for a Swedish customer

GPU financing is starting to look a lot less like venture capital and a lot more like equipment finance. That is the real story here. USD.AI says it has provided a $34 million loan to a Swedish customer to finance 96 servers carrying 768 Nvidia GPUs, using the kind of legal-and-onchain structure it has been building out for AI infrastructure deals. If that sounds niche, the stakes are not — whoever figures out how to fund GPU fleets cheaply gets a real edge in the AI buildout. ### What actually changed? The immediate news is the loan itself. USD.AI says the facility totals $34 million and finances 768 Nvidia GPUs across 96 servers for a customer in Sweden. The company has been pitching itself as a lender for “neoclouds” and AI infrastructure operators, and this deal fits that template almost exactly. (docs.usd.ai) ### Why is a GPU loan interesting? Because GPUs are expensive, depreciating, and hard for normal lenders to underwrite. A startup can raise equity to buy them, sure, but that is slow and dilutive. A bank could in theory do equipment finance, but USD.AI’s whole pitch is that banks usually do not want this risk profile, especially when the collateral is fast-moving compute hardware rather than aircraft, trucks, or real estate. (us([docs.usd.ai)inance-gpu-credit)) ### So what is USD.AI really selling? Basically, not just money. It is selling a financing wrapper that makes GPUs legible to lenders. The protocol originates non-recourse loans secured by the physical hardware and related cash flows, then exposes that debt onchain so depositors can fund it and earn yield. The company’s docs describe this as institutional-grade GPU financing rather than speculative crypto lending. (docs.usd.ai) ### How does the structure work? This is the part that matters. USD.AI uses a bankruptcy-remote SPV to hold the hardware, escrow to control the flow of funds, and lien filings plus bailment arrangements so the lender has an enforceable claim on the servers sitting in the data center. Onchain tokens sit on top of that legal stack — they do not replace it. That is the only way a “GPU-backed” loan is more than marketing. (usd.ai)the economics look like? USD.AI has not posted a full page for this Swedish transaction on its site yet, but its recent live deals and product pages show the rough range. A loan announced on April 6 carried a 10.0% annual rate, 70% loan-to-value, and a three-year term. Its yield product advertises expected APR in the low teens, and borrower docs describe a debt service reserve of roughly 10% plus conservative LTVs around(usd.ai)% is not some outlier — it sits inside the band USD.AI has been telegraphing. (usd.ai) ### Is this a one-off? Not really. USD.AI’s April 2026 news flow shows a $26.8 million loan backed by 576 Nvidia B300 GPUs and a separate commitment of up to $500 million for Sharon AI. That does not prove this Swedish deal is huge by hyperscaler standards, but it does show a pattern — the company is trying to build an actual loan book, not just announce a pilot. (usd.ai)Europe as well as North America, Australia, and Southeast Asia. So the Swedish angle looks less like a special case and more like evidence that GPU credit markets are spreading beyond U.S. operators. If compute becomes a globally financeable asset class, geography stops being the bottleneck and cost of capital becomes the bottleneck. (usd.ai 1) (usd.ai 2)art is not that someone borrowed to buy Nvidia GPUs. Everybody wants Nvidia GPUs. The interesting part is that lenders are building repeatable machinery to treat those boxes like collateral — with escrow, SPVs, liens, and onchain distribution layered together. If that model holds up, AI infrastructure gets funded faster, and the winners may be the operators who can borrow cheapest, not just the ones who can raise the most equity.