Historical source report: November 9, 2026, at 19:16:30 UTC+8. Written by Pink Brains and translated into Chinese by TechFlow. Figures and developments below reflect that report, not independently verified current conditions.
GPU-backed lending connects stablecoin deposits to physical servers, cloud operators and off-chain enforcement. The central question is not simply what yield those loans offer, but who owns the equipment, who controls repayments and what happens when depositors want their money back before a loan matures.
A report by Pink Brains, translated by TechFlow, describes how USD.AI and its GPU Loan Exchange, GLX, seek to finance cloud operators that struggle to obtain bank credit. The model combines special-purpose vehicles, hardware collateral and stablecoin-funded loans. Moving the funding on-chain does not remove the underlying borrower, depreciation or liquidity risks.
Who owns the GPUs—and who bears the loss?
Under the structure described in the report, a borrower’s special-purpose vehicle, or SPV, owns the GPU hardware. USD.AI holds first-priority liens over the servers and all equity in the SPV. Loans typically finance 70%–80% of hardware acquisition costs, with borrowers supplying the remainder.
Funds remain with a third-party escrow provider until the servers are installed in a data center and their performance is verified. Rental revenue then goes first toward loan principal and interest.
If a borrower defaults, the resolution agent initially seeks to keep the SPV operating under a replacement operator or sell the entity at a price sufficient to repay lenders in full. Individual server auctions through IT asset disposition specialists are described as the last step after restructuring options fail.
Those arrangements establish a proposed recovery process, not a guaranteed recovery value. The distinction is especially important for non-recourse loans: creditors may have claims against an SPV’s assets without recourse to its parent company.
The stablecoin and the lending receipt are different exposures
The report describes USDai as backed 1:1 by PayPal’s PYUSD, whose underlying reserves include U.S. Treasuries. USDai itself does not pay interest.
Depositing USDai into the staking contract produces sUSDai, the yield-bearing receipt. Its pool finances GPU loans and collects lending interest. The report identifies sUSDai holders as the parties bearing GPU credit-default risk.
The backing description should therefore not be read as a promise that GPU loans are immediately redeemable at par. The source does not provide a complete legal analysis of depositor claims, a full reserve-and-liability reconciliation or an independent audit establishing customer protection.
Why borrowers would pay more than bank rates
The financing gap, according to the report, reflects both credit quality and transaction economics.
American Compute’s figures cited in the article show 110 publicly disclosed GPU and data-center financings since 2023, totaling $197 billion. Of that funding, 92% was raised in 2025 and 2026. Among 49 transactions since July 2025 with named lenders, 32 were led by JPMorgan, Goldman Sachs, Morgan Stanley or MUFG.
Pricing depended heavily on tenant credit. The report puts 2026 loans backed by investment-grade leases at SOFR plus 2.25–3.0 percentage points, compared with SOFR plus 4.5–5.5 points—or fixed rates of 9%–10%—for unrated or below-investment-grade borrowers. SOFR is a dollar overnight borrowing benchmark secured by U.S. Treasuries.
A bank-grade SPV could cost $200,000–$500,000 to establish and take months to approve, according to the source. Those fixed costs made financings below $50 million unattractive to banks, leaving smaller operators underserved.
USD.AI’s pitch was access and speed rather than cheaper debt. The report says it had originated 20 financings totaling $531 million, with individual loans ranging from several hundred thousand dollars to $128.9 million. Its advertised pricing ranged from 7% for borrowers with investment-grade leases to 15% without long-term rental agreements.
Across 15 loans tracked from January through August 2026, executed rates were 10%–15%, with a 12% median. USD.AI claimed transactions could close in as little as seven days; that was a project claim, not an independently established service standard.
A monthly exit window against roughly three-year assets
The most consequential funding mismatch sits between sUSDai redemption processing and loan maturities.
The report describes redemptions processed in batches every 30 days, while GPU loans generally run for about three years. A monthly queue can organize withdrawals, but it does not itself turn long-dated loans into cash.
The vault’s stated CHIP Hurdle Rate was 7%. On November 1, when sUSDai’s annualized yield was 6.53%, USD.AI reportedly contributed 127,927.66 USDai to bring it to 7%. That contribution is distinct from recurring borrower interest and does not establish that the target is guaranteed.
The article contrasts borrower rates around a 12% median with depositor yields around 7%. Relevant costs and income components included:
- Residual-value insurance costing 150 basis points annually, deducted from staking yield.
- A protocol share described as 10% of net interest.
- Lower Treasury-rate returns on cash not yet deployed into GPU loans.
DefiLlama figures quoted in the report showed October fees of $4.87 million and protocol revenue of $487,000, with $24.75 million in gross revenue over the preceding year.
Deployment was uneven. At an earlier point, vault deposits approached $700 million while only about $1 million had been lent against computing assets; the article attributed much of the inflow to token-airdrop expectations. By mid-October, it reported $177 million in active loans and another $134 million in escrow awaiting chip delivery, broadly matching DefiLlama’s $309.5 million loan figure.
That breakdown matters: deposits, escrowed commitments and operating loans do not generate the same returns or offer the same liquidity.
GLX turns loan retention into loan distribution
GLX launched on October 1, according to the report, to help move loans from USD.AI’s balance of funded assets to external institutional buyers.
In the described model, sUSDai supplies initial bridge financing. After several months of repayment history, loans can be sold to professional asset managers, releasing cash for new originations. Institutions can also co-fund loans at inception.
On October 29, Silver Point Capital—which the article described as managing $48 billion—and Janus Henderson’s Victory Park Capital joined USD.AI through GLX in a $106 million, four-year senior secured loan to NexGen Cloud. The financing supported deployment of 1,728 Nvidia B300 GPUs in Finland.
The source reported more than $200 million of assets passing through GLX. It also described a governance proposal setting these sale conditions:
- A minimum sale price of par, or 100, with 100.5 or higher recommended.
- A 99.5 floor permitted only when redemption requests extending beyond a 30-day cycle exceeded available liquidity.
- A split of premium-sale profits: 60% to USD.AI Inc and 40% to a reserve vault supporting sUSDai’s 7% benchmark.
- No profit split for sales at par or a discount, and requirements for buyers to be independent of the originating entity.
Secondary sales can release cash and transfer future loan exposure. Their usefulness nevertheless depends on willing buyers and executable prices. Proposed pricing rules do not, by themselves, establish that a buyer will be available during stressed withdrawals.
A reported exit—and the limits of recourse
The article presents Duos Edge AI as an example of a completed loan exit.
On June 5, 2026, USD.AI announced a $98.1 million, three-year loan to the Duos Technologies subsidiary, secured by 2,304 Nvidia B300 GPUs. Hydra Host managed the equipment through its Brokkr platform.
The SPV financing was fully non-recourse: USD.AI could pursue the SPV and its hardware, but not the Duos parent company.
In September, Axe Compute acquired the SPV and, according to the report, paid approximately $87.8 million to settle all outstanding principal and interest. The acquired cluster had a five-year lease and projected cash flow of $364.6 million. That cash-flow figure was a projection, not realized revenue.
The report says USD.AI recovered its outstanding principal and interest in less than four months and recycled the capital into further lending. One successful exit does not establish recovery outcomes for the wider portfolio.
Token buybacks are a separate claim on the economics
The report says a CHIP buyback program routes sUSDai net-interest-margin revenue to market purchases after CHIP Foundation operating expenses. No fixed purchase amount or schedule had been set.
Separately, on October 26, USD.AI Inc reportedly spent $1 million USDC buying 22,343,372 CHIP at an average price of $0.044756, equivalent to 1.1% of circulating supply.
These flows should not be conflated with depositor protection. Origination fees of 2%–3% accrue to USD.AI Inc, while the protocol’s recurring revenue depends on its share of lending income. Faster loan turnover can create more origination fees without implying an equivalent increase in token-holder value.
The remaining risk is physical and contractual
The source argues that GPU lending benefits from strong rental demand, but also identifies rapid hardware depreciation and financially vulnerable borrowers as central risks. USD.AI’s reported mitigants included approximately three-year loan terms, restrictions on eligible GPU architectures and residual-value insurance.
Those measures do not eliminate uncertainty over equipment resale prices, tenant payments, insurance recovery or the ability to sell loans when depositors seek exits. On-chain funding can make parts of the transaction visible; ownership, enforcement and recoveries still depend on off-chain entities and contracts.
The key distinction is between funding access and risk removal. GLX may help distribute loans, and SPVs may define collateral rights, but neither makes a long-dated GPU loan equivalent to liquid cash.