USD.ai has become one of the louder yield stories in DeFi this quarter. Its CHIP governance token went live at the end of March, the airdrop has just finished distributing, and the protocol recently crossed $20M in cumulative yield paid out as its loan book grew, and sUSDai now holds close to $300M.

The most useful thing to understand before buying is that USD.ai is two tokens with two very different risk profiles. USDai is a plain PYUSD wrapper. sUSDai, the token that pays the yield, is closer to the equity tranche of a GPU lending fund. The yield is real and it comes from real borrowers, but it is credit yield, not deposit yield, and that distinction is what this issue is about.

USDai is backed one-to-one by PYUSD, the Paxos-issued, T-bill-collateralized stablecoin. Its entire job is to hold reserves. By the protocol's own documentation, USDai carries no exposure to GPUs or any loan the protocol originates. Hold USDai and your only real risk is Paxos and PYUSD.

sUSDai is the opposite. You stake USDai into a vault, and that vault holds the loan book. The GPU credit risk, the yield upside, and the potential principal loss all live in sUSDai. It is not a savings account. It is a first-loss claim on a portfolio of loans made against AI servers.

Roughly half of protocol assets sits in the PYUSD reserve earning 4.5% APY, about 15% was drawn into active loans at ~11% APY, and the remaining was staged commitments waiting to fund future loans. The blended 12.5% expected APY includes future accrual on that committed-but-unfunded capital. The returns actually realized on funded assets today is lower, closer to 7%, because the large PYUSD reserve earns 4.5% while the loan book progressively deploys over time.

This is normal for a credit book that is still scaling. CHIP incentives and the airdrop pulled deposits in faster than Permian Labs, the team behind USD.ai, can originate and fund loans, so fresh capital sits in the reserve until it is lent. The realized rate should converge toward the projected rate as commitments draw, provided the loans fund and stay current. It is worth knowing which of the two numbers you are actually buying.

The loans go to neoclouds and AI infrastructure operators, secured by installed, revenue-generating GPUs, mostly Nvidia B300 and B200 servers, at an 80% loan-to-value cap. Each one sits in a bankruptcy-remote SPV with a perfected lien on the hardware, escrow at Wilmington Trust, property insurance, and value reinsurance from Barkr. The legal stack is more thorough than most of DeFi.

Pricing the collateral is not the problem it is sometimes made out to be. There is deep public pricing for these GPUs, from rental marketplaces like Vast.ai and the Ornn compute price index to secondary-sale benchmarks. The recent signal is strong: B200 rental rates have roughly doubled over the past six weeks on the ORN index, which supports both borrower cash flow and resale value. What the protocol does not do is run an on-chain price oracle or margin-call when hardware values move. Default is triggered by missed payments, not by collateral value.

Several features contain the loss when a borrower does fall behind. The loans amortize fast, returning roughly 4% of principal and interest every month, so outstanding exposure on any loan falls steadily from day one (roughly half paid back after just 1 year). A missed payment moves a loan into default after a 30-day grace period, at which point the servers are sold through IT-asset-disposition partners rather than left in limbo. And Barkr value reinsurance is designed to cover the shortfall when a liquidation clears below the contracted floor. The honest residual gaps are that the Barkr policy limits are not publicly disclosed, and that Nvidia B-series resale markets are still thin because the cards are new, so forced-sale timing carries more uncertainty than the rental data alone implies.

The clearest risk in the book today is operator concentration, and it is sharper than the protocol's many-locations framing suggests. Grouped by the operators that actually host the financed hardware, the active loan book leans heavily on one name. QumulusAI accounts for roughly 79% of drawn loans, about $44.6M across three loans in Washington, Missouri, and Pennsylvania, at a 73% weighted loan-to-value. Crucible is the only other meaningful exposure at about 15%, $8.6M in Texas at a more conservative 54% LTV. Lyceum in France is around 4%, and everything else is rounding error.

The reassuring side is that these are named, identifiable operators rather than anonymous wallets, the loans amortize quickly, and the weighted LTVs leave a real equity cushion. The concern is simply that the diversification is thin. One operator's payment behavior drives most of the book, so the question to underwrite is less how diversified the fund is and more how comfortable you are with top operators like QumulusAI.