A $100m Stablecoin Line Against Graphics Cards

Bullish, the institutional crypto exchange, has committed a $100m stablecoin facility to USD.AI, a protocol that lends against artificial intelligence computing hardware, according to reports from The Defiant and Cointelegraph. The money funds loans secured on graphics processing units, the chips that run AI training and inference workloads. USD.AI's own API reports $265m of loan reserves and $491m of total value locked, The Defiant said.

The structure is a real-world asset trade with an unusually physical collateral base. Most onchain credit backed by off-chain assets is secured on treasury bills, invoices or private credit, all of which are paper claims. Here the security is equipment sitting in a data centre, and the borrower is typically an operator who needs working capital to buy more of it than cash flow allows.

Why lenders like it and why it is hard

The appeal is that GPU rental income is contracted, dollar-denominated and currently in short supply, which makes the cash flow easier to underwrite than most crypto-native collateral. The difficulty is everything to do with the hardware itself. A GPU is a depreciating asset on a schedule set by the next product generation, it is physically located in a jurisdiction with its own insolvency law, and recovering it after a default means repossessing racks rather than liquidating a position.

That is the underwriting question a $100m facility has to answer, and the size of it relative to the protocol matters. Against $265m of reported loan reserves, a $100m commitment is a substantial addition to lending capacity rather than a marginal one.

Stablecoins as the funding leg

The facility is denominated in stablecoins, which is the part that makes it an onchain story rather than a private credit one. Dollar tokens give the lender same-day settlement and a borrower base that already holds balances onchain, without either side arranging bank rails for each drawdown. It also means the credit risk and the settlement asset live on different layers: the loan performs or fails on the physical hardware, while the money moves as a token.

The wider pattern

Compute-backed lending is one of the few real-world asset categories where onchain capital is funding something that did not previously have easy access to it. Tokenized treasuries move an existing instrument onto a new rail. A loan against a rack of GPUs creates credit that a traditional lender may have declined to write, either because the collateral is unfamiliar or the borrower is too small.

That novelty cuts both ways. A category with no long default history and rapidly depreciating security is exactly where a credit cycle finds its first casualties, and the sector has not yet been tested through a downturn in GPU rental rates. Nothing in the reported figures indicates stress, and nothing in them indicates the loans have been through one.

What to watch is whether the reported reserves and total value locked move together as the facility is drawn. Loan reserves rising without a matching increase in locked value would mean capital is being deployed rather than accumulated, which is the point of a credit facility and also the moment when the underwriting starts to matter.