GPU‑backed loan on Arbitrum

# GPU-backed loan on Arbitrum shows how crypto rails are moving into AI infrastructure A $26.8 million loan tied to 576 NVIDIA B300 graphics processing units was drawn on April 6, 2026 through USD.AI on Arbitrum, with the hardware in Washington State serving as collateral. The borrower was Crucible Capital, and the loan terms disclosed by USD.AI were a 10.0% annual rate, a three-year term, and a 70% loan-to-value ratio. (usd.ai) That is unusual for two reasons at once. First, the collateral was not crypto, Treasury bills, or invoices, but physical AI hardware. Second, the loan was settled through Arbitrum, an Ethereum layer-two network usually associated with decentralized finance trading and stablecoins rather than equipment finance. (usd.ai) The basic idea is simple: instead of forcing an AI infrastructure operator to sell equity or wait for a bank credit committee, the operator posts tokenized rights to its hardware and borrows against it. USD.AI says its system lets borrowers use graphics processing units represented as electronic documents of title non-fungible tokens as collateral, then receive Tether in return and make payments every 30 days. (blog.arbitrum.io) For this to work, the token cannot just be a digital sticker attached to a server rack. USD.AI’s CALIBER framework wraps each loan in a bankruptcy-remote special purpose vehicle, places the hardware under a Uniform Commercial Code Section 7 bailment structure at a third-party data center, and tokenizes the enforceable claim as an ERC-721 token on Ethereum. (usd.ai) That legal structure is the bridge between onchain finance and real machines. USD.AI says the token metadata links to appraisals, Uniform Commercial Code filings, and custody documentation, while liquidation is designed to happen through orderly remarketing or re-leasing of hardware rather than a forced crypto-style fire sale. (usd.ai) The Arbitrum piece matters because it shows where this financing is being staged. USD.AI launched publicly on Arbitrum in August 2025 after saying it had already attracted more than $50 million in deposits during private beta, and both USD.AI and Arbitrum have framed the network as a hub for real-world-asset and infrastructure-linked lending. (usd.ai) This is part of a bigger shift in what crypto networks are being used for. In its August 2025 case study, Arbitrum said real-world-asset value on the chain had recently crossed $355 million, and it positioned deep decentralized finance liquidity and stablecoin activity as reasons infrastructure-backed credit products would launch there. (blog.arbitrum.io) The other half of the story is outside crypto. PitchBook reported in August 2025 that venture-backed companies were increasingly using graphics processing units as collateral for debt, with examples including CoreWeave, Lambda, Crusoe, and Fluidstack, as lenders chased the capital demands of AI data centers. (pitchbook.com) So the novelty here is not that GPUs can back loans at all. The novelty is that a mid-sized version of that trade is now being originated through crypto infrastructure, with tokenized claims, public smart-contract settlement, and a consumer-facing blockchain network acting as the rail for a piece of equipment finance. (usd.ai) There is also a clear target market. Arbitrum’s 2025 profile of USD.AI said the protocol was built for “sub-$20M” compute operators that are often too small for conventional leasing and bank financing, while USD.AI’s public launch described the product as serving underserved AI builders that can post tokenized hardware as collateral. (blog.arbitrum.io) That helps explain why this loan is being watched. Traditional private credit has already moved into large GPU-backed financings, but crypto rails may be better suited to smaller, faster, and more standardized deals where documentation, collateral tracking, and liquidity provisioning can be handled in a more programmable way. That is an inference from the borrower profile USD.AI describes and from the broader private-credit trend PitchBook documented. (blog.arbitrum.io) The risks are real, though. PitchBook noted that lenders still debate the economic shelf life of graphics processing units and how quickly a valuable chip can be overtaken by a newer model, which makes collateral valuation and liquidation harder than in older asset-backed markets. (pitchbook.com) That risk is exactly why the Crucible deal is notable. A loan backed by 72 NVIDIA B300 servers, rather than by crypto tokens or Treasury products, suggests that layer-two networks like Arbitrum are becoming venues for a new category of tokenized finance: loans against compute infrastructure, where the asset producing AI revenue is also the asset securing the debt. (usd.ai) If this model scales, Arbitrum and similar networks may end up doing more than hosting decentralized exchanges and stablecoin transfers. They may become settlement layers for a market where server racks, data-center contracts, and AI hardware are financed the way inventories, receivables, and real estate have been financed in traditional markets for decades. That final step is a forward-looking inference, but it follows from the structure of the Crucible transaction, USD.AI’s collateral framework, and the broader move toward GPU-backed lending. (usd.ai)