USDai: who loses if the GPU depreciates by 17.9% per year

The example that USD.AI publishes on its own risk underwriting page implies that the protocol accepts an annual depreciation of 17.9% on the GPUs it finances. This figure does not appear written anywhere: it is derived by solving for two statements from the issuer—"hardware is valued at a maximum LTV of 80% at origination" and "a loan originated at 80% LTV deleverages to approximately 65% by the end of the first year"—using a 36-month linear amortization (LTV, loan-to-value, is the percentage of the collateral value that is lent; the calculation is ours, not the issuer's). Who pays when the GPU loses that 17.9% is split between two tokens that are similar in name and nothing else: those holding USDai do not bear that loss—the documentation states the token "has no exposure to GPUs or any loans originated by the protocol"—and those holding sUSDai bear it entirely, although its price won't reflect it until a borrower stops paying. This article deconstructs that structure using the issuer's technical documentation, the implicit arithmetic of its example, and fifteen months of DefiLlama deposits and loans, including the day $52 million entered all at once.

What exactly is a depositor buying in USD.AI?

USD.AI is a credit protocol on Arbitrum that lends to artificial intelligence infrastructure operators with GPUs as collateral. On August 28, 2026, at 07:00 New York time, Bullish (NYSE: BLSH) announced a $100 million stablecoin debt facility to the protocol and the listing of its yield token in several pairs with market making. The usd.ai landing page describes the product as a dollar "collateralized by the physical backbone of the AI revolution"; the technical documentation, which is the governing authority for knowing what is being bought, describes two instruments with opposite functions:

The architecture is a tranche-based securitization using crypto vocabulary: the same division of roles that separates USDe from sUSDe in Ethena, applied here to hardware instead of derivatives and private credit. The senior tranche —USDai— earns zero and risks nothing: it is a PYUSD wrapper, with direct minting announced to be restricted to verified institutions starting in April 2026, although the documentation consulted on September 6 still describes it as open to anyone. The junior tranche —sUSDai— is an ERC-4626 vault with ERC-7540 redemptions that absorbs all the business risk, and the documentation itself warns bluntly: "sUSDai is not a stablecoin. It is not instantly redeemable at par."

The "GPU-backed dollar" mentioned in marketing is, on the balance sheet, a dollar backed by U.S. public debt through a PayPal stablecoin. The GPUs back the other token.

Where does the 17.9% annual depreciation implied by the USD.AI example come from?

The USD.AI risk underwriting page publishes three parameters for a typical loan: maximum LTV of 80% on verified acquisition cost, a standard three-year term with linear amortization, and a deleveraging trajectory—"a loan originated at 80% LTV deleverages to approximately 65% LTV at the end of year one, because the principal amortizes faster than the collateral value falls." With these three data points, the depreciation assumption is determined. This calculation is CleanSky's own analysis, not a figure published by the issuer:

That 17.9% is the assumption supporting the commercial phrase "the principal amortizes faster than the collateral drop." If the actual first-year depreciation were to exceed 17.9%, the loan would stop deleveraging: the LTV would rise instead of falling toward 65%, and each month of the loan's life would leave the lender with less cushion than the previous month. The threshold separating one scenario from the other is that figure the issuer does not publish.

Can an H100 withstand that 17.9% annual depreciation?

Regarding actual hardware depreciation, confidence in available sources must be tempered: there is no audited residual value index for data center GPUs comparable to used car price guides. What exists are reports from secondary market sellers and analysts. The GPU Depreciation & Residual Value Report 2026 by AMCompute and HashrateIndex tracking place the value retention of an H100 at around 75-85% at 24 months and 45-60% at 36 months, with refurbished units trading between $18,000 and $22,000 during 2026. Annualized on a compound basis, that 36-month range equates to a depreciation of between 15.7% and 23.4% per year (own calculation): the 17.9% implicit in the official example falls within that, closer to the optimistic end than the pessimistic one.

What is not published is the piece that decides the outcome when something goes wrong. Barkr, the firm that underwrites the protocol's residual value insurance—150 annual basis points automatically deducted from interest yield, with coverage up to 80% of the guaranteed amount—pays the difference between the guaranteed price and what is fetched in the sale. The documentation defines that guaranteed price as a function of a "predefined RVI schedule" (RVI, residual value insurance), and that schedule does not appear in any public document. It also adds a warning worth reading in full: "coverage exclusions, claim eligibility conditions, and insurer counterparty risk are not modeled."

Why doesn't the sUSDai price reflect the GPU's loss in value?

The method for calculating the vault's Net Asset Value (NAV) contains the piece that reorders everything above. Loan positions are valued "conservatively at the remaining loan balance, or optimistically at the remaining balance plus interest accrued since the last installment": the entry price uses the optimistic valuation and the redemption price uses the conservative one.

Neither looks at the market value of the hardware. A loan with an H100 that has lost 30% of its resale price is worth the same in the NAV as one with a newly installed GB300, as long as both borrowers pay their installments. The documentation is explicit about when that changes: the adjustment for collateral sale and insurance collection is "the only discrete NAV event in the lifecycle of a default."

The sUSDai holder does not suffer depreciation continuously, as they would with a hardware manufacturer's stock: they suffer it all at once, and only if a borrower stops paying. A GPU that plunges 40% on the secondary market will not move the sUSDai NAV by a single basis point as long as the operator continues to pay the monthly installment.

What does USD.AI look at to set a loan's interest rate?

The loan documentation states it in a way that organizes the entire analysis: rates are fixed, without reference to SOFR, and are set at origination "based on two variables: the quality of the offtake contract—a binding agreement by which a counterparty commits to buying compute capacity under defined terms—and the loan-to-value ratio, not the borrower's creditworthiness."

Tiers 1 and 2 require at least 24 months of remaining contract at origination; within each band, the rate is decided by the LTV, with 70% at the floor and 80% at the ceiling. Alongside the rate, there are two more locks, both on cash flow rather than hardware: a debt service reserve of around 10% of the gross amount—about three months of peak service—provided by the borrower's own capital before anything is released, and a minimum debt service coverage ratio of 1.15x reviewed every quarter against actual receipts from compute purchase contracts.

The rental market explains this hierarchy. The Silicon Data index puts the H100 peak at $12.00 per GPU-hour in 2023; by late 2025, the marketplace segment was trading between $1.92 and $2.00 per hour, and the hyperscaler segment between $6.20 and $6.64. The distance between those two segments is the same as that separating a Tier 2 multi-year contract from a Tier 3 spot rental.

How much of the 8.28% comes from a GPU and how much from a PayPal incentive?

The protocol's homepage showed on September 6, 2026, a current APR of 8.28% (APR: the annualized yield advertised by the protocol itself), an expected APR of 11.81%, $546.9 million in total deposits, and a loan pipeline of $321.6 million. The 4-week US Treasury bill yielded 3.72% in coupon equivalent on September 4, the latest data published at the time of the query. The spread charged for assuming the risk of the junior tranche is, therefore, 4.56 percentage points.

That spread does not come solely from the business of lending against GPUs. The documentation describes an agreement with PayPal that distributes an annual incentive of 4.5% on the PYUSD held by the protocol, with a cap of 1,000 million dollars in loan backing and a stated horizon of 2026, which "accrues at the protocol level and flows into the sUSDai yield." PayPal's logic is that of any stablecoin issuer: every dollar parked in PYUSD inflates its currency circulation and the reserves backing it, and paying for that balance is cheaper than acquiring it through distribution. With the frozen utilization of September 6 —263.1 million lent out of 546.1 million total, or 48.2%— the remaining 51.8% earns that incentive or the yield from T-bills, depending on where each dollar is parked.

The following breakdown is our own calculation, as the issuer does not publish the APR breakdown. It assumes the upper bound of the incentive (that all non-lent capital earns the full 4.5%):

At the opposite extreme —where everything not lent out is in T-bills at 3.72% without collecting incentives— reserves contribute 1.93 points and the book would have to yield 13.2%. Regardless of the distribution, the implicit yield of the book remains between 12.3% and 13.2%: above the ceiling of the level 2 band —non-investment grade counterparties with multi-year contracts— and already within the level 3 band, that of compute sold without a multi-year contract. Furthermore, more than a quarter of the 8.28% APR of sUSDai —2.33 of those 8.28 points, the upper bound of our calculation— comes from the 4.5% PayPal incentive, with a $1 billion loan backing cap and a declared expiration in 2026.

What do fifteen months of USD.AI deposits and loans tell us?

The DefiLlama daily series allows us to reconstruct something no press release mentions: for almost all of 2025, money entered the protocol and that money was not lent. Utilization is our own calculation (lent divided by the sum of unlent and lent deposits).

On January 1, 2026, USD.AI held $684.4 million and had $1.1 million lent: a utilization of 0.16%. Those holding sUSDai at that time were, in practice, earning Treasury bill yields and the PayPal incentive, with GPU exposure near zero. Between January 1 and August 1, $524.8 million in unlent deposits left while the loan book grew by $235.0 million—both figures measured by the same yardstick, the DefiLlama series: the capital that arrived seeking risk-free yield left when the risk actually began to exist.

There is one specific movement worth noting by date. On August 27, unlent deposits were $173.7 million; on August 28, $225.7 million. $52.0 million in a single day, the same day as the Bullish facility announcement. DefiLlama does not attribute deposits to counterparties, so the date coincidence is the only verifiable fact: there is no public confirmation that this money is from the facility.

A methodological note, as several figures for the book are circulating that measure different things. The issuer's own YTD report, published on June 8, 2026, declares a TVL of 398 million dollars "with 202 million deployed," 13 active loans, and an average sUSDai yield of 7.0% so far this year. DefiLlama, as of June 1, totaled 396.8 million between both items —the total matches almost to the dollar— but allocated only 103.8 million as "borrowed": half of what the issuer calls deployed. The totals match while the distribution does not, so the growth of the book can only be read within a single source: in DefiLlama's, the borrowed balance went from 103.8 million on June 1 to 263.1 on September 6, a 153% increase in three months. For the same reason, the 283.0 million in TVL shown by DefiLlama and the 546.9 million in "total deposits" from usd.ai do not contradict each other: the DefiLlama homepage excludes what is borrowed, and by adding it —283.0 + 263.1 = 546.1 million— it aligns with USD.AI with a deviation of 0.15%.

What happens the day a USD.AI borrower stops paying?

The loans are non-recourse: if the operator doesn't pay, the creditor goes after the hardware and the vehicle's flows, not the parent company's balance sheet. In exchange, the legal structure is dense, of the same lineage as the warehouse loans Maple sets up with Kraken. On-chain, the NFT representing the server package is frozen, can no longer be transferred or pledged, and the distribution logic shifts from the ordinary payment waterfall to the acceleration waterfall, with all collections directed to lenders pro-rata until the loan is amortized.

Off-chain, the agent accelerates the debt and exercises remedies under the U.S. Uniform Commercial Code—foreclosure, repossession, private sale—with lien waivers signed with the data center allowing them to enter, disconnect, and remove the servers, and the option to execute against the parent's stake in the vehicle. IT asset disposition (ITAD) partners manage removal and resale, Alliant covers damage insurance, and Aravolta monitors the hardware in real-time.

The residual value insurance kicks in at the end, and with a cap that should be understood: it pays the difference between the guaranteed price and the proceeds from the sale, and what returns to the protocol is the minimum of that sum and the frozen principal plus interest. As a lender and not a shareholder, the protocol recovers at most what it is owed and does not capture upside beyond its credit.

There are two identified gaps that a depositor should have located. The first: the auction mechanism that would allow paying to jump ahead in the exit queue —Queue Extractable Value, QEV, extensively described in the documentation— is "planned but not yet implemented," with the timeline and specifications still open. Today, exiting sUSDai means a FIFO queue —first-in, first-out— in 30-day cycles, which during periods of high utilization can extend across several cycles, as the protocol does not liquidate loans early to meet withdrawals. The second: the figure of the curator with first-loss capital described on the "FiLo Curator" page no longer appears in the documentation consulted on September 6, 2026 —zero occurrences of the term in the full dump of docs.usd.ai, and the URL returns a 404 error while adjacent pages respond normally—. The backstop that does appear today is sCHIP: the governance token in staking, which can be used to cover a deficit.

What can be checked before touching sUSDai?

From the USD.AI structure, five checks emerge that anyone can perform without privileged access and that discriminate better than the landing page APR.

1. Utilization. Dividing the "borrowed" figure from DefiLlama by the sum of both items reveals what proportion of the advertised yield depends on loans versus reserves: 48.2% on September 6, 2026, is a very different number from 0.16% on January 1.
2. The distribution of the live book by compute purchase contract tier. Between the 7% of an investment-grade counterparty and the 15% of a spot rental, there are two distinct businesses, and the implicit yield of 11.8-12.4% that balances today's APR is only sustainable with tier 2 weighting.
3. The RVI schedule of the residual value insurance. Until it is published, the coverage of up to 80% of the guaranteed amount underwritten by Barkr remains a percentage of an unknown base.
4. The expiration of the 4.5% PayPal incentive. It contributes up to 2.33 of the 8.28 points of the current APR according to our calculation, has a cap of 1,000 million dollars in loan backing, and a declared expiration in 2026.
5. The history of the sUSDai redemption queue. How many 30-day cycles have been required to clear it in each epoch since utilization exceeded 50%.

None of those five have to do with the resale price of an H100, and all have to do with whether the operator renting those cards can continue to pay their installment. The protocol has built a considerable structure around the collateral and has placed its two toughest conditions—the 1.15 debt service coverage and the three-month reserve—on cash flow. The risk the structure discounts and the risk the slogan announces are not the same.

Sources and links: USD.AI Documentation · USD.AI, Underwriting and Risk Management · USD.AI, 2026 YTD "Lighthouse" Report (Jun-8-2026) · DefiLlama, Protocol Historical Series · PR Newswire, $100M Bullish Facility (Aug-28-2026) · U.S. Treasury, Daily Bill Rates · Silicon Data, H100 Rental Index · AMCompute, GPU Residual Value Report 2026