Bullish, the institutional crypto exchange operator, has extended a 100 million USD stablecoin debt facility to USD.AI to finance lending secured by graphics processing unit hardware, the companies announced Friday, in one of the clearest signs yet that GPU-backed credit is becoming a distinct onchain asset class.
Under the arrangement, USD.AI, an onchain financing platform developed by Permian Labs, will use the facility to extend loans to artificial intelligence infrastructure operators. The collateral is the GPU hardware itself, rather than the borrowers’ broader corporate balance sheets, meaning lenders’ recovery path runs through the chips, servers and networking gear that power AI compute.
## Stablecoin liquidity meets AI capex
The deal connects two of the most powerful flows in digital assets and technology markets: stablecoin issuance, which has grown into a multi-hundred-billion-dollar pool of dollar-denominated liquidity, and the financing scramble for AI compute, where demand for Nvidia-grade hardware has outstripped operators’ ability to fund it with traditional debt.
USD.AI positions itself as the connective tissue. The platform mints sUSDai, a tokenized representation of its GPU-backed lending book, allowing stablecoin holders to gain exposure to hardware-secured yields. Bullish said it plans to list sUSDai across multiple trading pairs and to support the token with a dedicated market-making program, which the exchange expects will improve secondary liquidity and price discovery for GPU-backed debt.
That listing commitment is the strategically important piece. Tokenized loan books have historically suffered from thin secondary markets, forcing holders to maturity. An exchange-operated market-making program gives institutional lenders an exit option, which in turn makes the underlying credit more attractive to originate.
## A growing GPU lending book
The facility builds on an increasingly active financing record. In June, USD.AI announced a 98.1 million USD loan backed by 2,304 Nvidia B300 GPUs, alongside a 34 million USD loan backed by 768 Nvidia B200 GPUs that was fully funded, according to the platform’s own disclosures.
The relationship between the two firms also has an equity dimension. Bullish Capital made a 4 million USD investment in USD.AI in September 2025, its first such investment since going public, meaning Friday’s credit facility deepens a partnership that began with a minority stake.
For AI infrastructure operators, the appeal is straightforward. Hardware-secured borrowing frees up capital that would otherwise be locked in depreciating compute, letting operators recycle cash into expansion while lenders hold a claim on assets with a deep resale market. The risk, of course, is that GPU values are cyclical, and a downturn in AI compute pricing would hit collateral recovery rates, a risk lenders are pricing for now in a tight hardware market.
## Bullish rides a rebound
The deal lands during a strong stretch for Bullish’s public market standing. The company listed on the New York Stock Exchange in August 2025, raising roughly 1.03 billion USD at 37 USD per share, with the stock opening at 90 USD on its first trading day.
Shares remain more than 60 percent below that debut level, but the stock has rebounded sharply, gaining roughly 45 percent over the past month to trade around 33 USD on Friday, according to Yahoo Finance data cited by Cointelegraph.
Bullish has not been alone. The past month’s recovery in digital asset markets has lifted a broad basket of crypto-linked equities, with Bitcoin treasury company Strive up about 88 percent, miner Canaan around 55 percent, and stablecoin issuer Circle gaining nearly 40 percent over the same period.
## Why it matters for DeFi
For the decentralized finance sector, the facility is a template for institutionalizing real-world collateral. Rather than chasing volatile crypto-native collateral, the model pairs a hard asset with demonstrable cash-flow utility, wraps it in a transferable token, and lists it on a regulated exchange with committed liquidity. It also gives DeFi yield seekers exposure to a cash-generating physical asset class without requiring them to warehouse hardware or underwrite borrowers directly, a division of labor the tokenized credit market has been chasing for years.
If GPU-backed lending scales, the next question is whether other collateral classes, from energy assets to data-center infrastructure, can follow the same pipeline from balance sheet to tokenized liquidity. Friday’s 100 million USD facility suggests Bullish and Permian Labs believe the answer is yes, and that the exchange’s willingness to commit its own balance sheet and market-making infrastructure marks the difference between another tokenization pilot and a durable credit market.
Disclaimer: This article is for informational purposes only and does not constitute financial advice.
second 100M facility after 98.1M in june. someone at bullish clearly likes this collateral haircut a lot
lending against gpus feels very 2026. the collateral depreciates faster than the loan matures, good luck to lenders holding last gen accelerators
B300s specifically, not last gen scrap. nvidia flagship holds value better than most altcoins lmao
until the B400 refresh drops and the flagship premium evaporates in a quarter. hardware collateral has a shelf life printed on the box
b400 cycle is the real clock on this. by the time the 2024 vintage b300s need refinancing the bid wont be there, haircuts or not
42k is list plus interconnect plus rack install tho, a deployed b300 is worth more than the bare chip. depreciation is still the killer, just not at spot component pricing
b400 timing is the whole bear case in one line. the first refinancing wall lands right as the flagship premium fades, no haircut schedule saves that
disagree on the haircut point, permian can reprice the loan book mid cycle if usd.ai keeps minting. the b400 risk is real but its a 2027 problem not a 2026 one
Thats why sUSDai exists tho, the token represents the debt position so the collateral terms are at least transparent on chain
onchain terms dont fix liquidation tho. pushing 2000 plus b300s into a used market that thin moves the bid against you exactly when you need it
that is the real tail risk, pushing 2000 B300s into a thin used bid. hopefully the haircuts are conservative enough that it never comes to that
the thin bid problem compounds too. the first liquidation dump tanks the used market, which reprices every other loan in the book at once
depends on the haircut tbh. B300s still command a premium used, if theyre lending at 40-50% LTV the depreciation is priced in
another 100M into gpu backed credit right as everyone argues ai capex has peaked. permian is basically running a shadow repo desk for nvidia hardware
backing the loans with the GPUs themselves is the interesting part. repo markets for Nvidia chips, wild
the sUSDai listing with a market making program is the real news, tokenized loan books always died from illiquidity
mm programs have a way of quietly sunsetting once the launch buzz fades. watch whether the sUSDai pairs still have quoted depth in march
same pattern as every yield program. apr holds while deposits flow in, quietly halves once the gpu collateral story stops trending
98.1M against 2,304 B300s in june and now another 100M facility. this book is scaling fast, wonder what the LTV is
sUSDai mints against the book so the LTV should be visible on chain soon enough. no more guessing
2,304 b300s for 98.1M works out to roughly 42k per unit. someone ran that math at bullish and still said yes twice
42k a unit on the june tranche and they came back for another 100M in two months. at that pace the b300 premium better hold across the entire loan book
the june tranche pricing at 42k a unit is the part everyone keeps skipping. nvidia cuts b300 output next quarter and that collateral mark moves fast
GPU-backed lending only works until hardware prices drop and the collateral is worth half the loan. Interested to see the haircut schedule they apply to those rigs.
second facility in two months means the june 98.1M book is probably fully deployed. the haircut schedule is the one doc that decides if this survives a gpu price correction
sUSDai with an mm program is basically a repo desk with nvidia collateral. if quoted depth holds past march this stops being a story, which is the actual win