Source publication: November 9, 2026, at 19:00 UTC+8. This historical-source rewrite reflects the figures and developments reported at that time.
Banks dominate financing for AI infrastructure, but smaller GPU operators can struggle to obtain loans even when they own hardware and collect rental income. An analysis credited to @PinkBrains_io, translated by AididiaoJP of Foresight News and published by Odaily, examines how stablecoin-funded lenders are trying to serve that gap.
USD.AI is the central example. Its model connects stablecoin deposits to loans secured by GPUs and rental contracts. The critical questions are who owns the hardware, which entity owes the debt, and whether depositors can obtain cash before the underlying loans mature.
The ownership and repayment chain
According to the source, USD.AI uses a structure familiar from conventional infrastructure finance:
- A borrower’s special-purpose vehicle, or SPV, owns the GPUs.
- USD.AI holds a first-priority lien over the servers and the SPV’s equity.
- Loans generally cover 70%–80% of hardware costs, with the borrower funding the remainder.
- Funds remain in escrow until server installation has been verified.
- Rental income then services the debt.
If a borrower defaults, an agent first attempts to replace the operator while keeping the SPV running, or sell the operation as a whole. Individual servers are sold through IT asset-disposal firms only if those approaches fail.
That structure does not necessarily give the lender recourse to a borrower’s parent company. In the Duos Edge AI transaction described in the article, recovery was limited to the SPV and its chips, not Duos Technologies itself.
On the funding side, the source describes USDai as backed 1:1 by PYUSD, whose reserves include U.S. Treasuries. USDai itself does not earn yield. Staking it produces sUSDai, the vault token that funds loans and accumulates interest. The article identifies sUSDai holders as the parties bearing GPU credit risk.
These are distinct layers: stablecoin backing, hardware collateral and borrower repayment obligations are not interchangeable protections.
Why smaller GPU operators pay more
The article cites American Compute data showing 110 publicly disclosed GPU and data-center financing transactions totaling $197 billion since 2023, with 92% raised in 2025 and 2026. Of 49 transactions since July 2025 that identified lenders, 32 were led by JPMorgan, Goldman Sachs, Morgan Stanley or MUFG.
Borrowing costs depended heavily on the tenant supporting the lease. The source reported the following 2026 pricing:
| Financing category | Reported pricing |
|---|---|
| Loans backed by investment-grade leases | SOFR plus 2.25–3.0 percentage points |
| Unrated or sub-investment-grade loans | SOFR plus 4.5–5.5 percentage points, or 9%–10% fixed |
| USD.AI loans tracked from January through August 2026 | 10%–15%, with a 12% median |
SOFR is an overnight borrowing benchmark secured by U.S. Treasuries. CoreWeave’s reported spread of 5.50 percentage points over SOFR was three percentage points above Nebius’s.
Large borrowers can place chips, leases and debt inside SPVs. Smaller transactions face a different constraint: due-diligence expenses do not fall proportionately with loan size. Citing The Token Dispatch, the article contrasts a $1 billion loan, where those costs are relatively minor, with a $30 million transaction, where they can undermine the economics.
USD.AI’s proposition is therefore access and speed rather than cheaper funding. Its advertised pricing ranged from 7% for investment-grade leases to 15% for deployments without contracts, and it said loans could close within seven days. An October report listed 20 financings totaling $531 million, ranging from hundreds of thousands of dollars to $128.9 million each.
Strong rental demand does not remove credit risk
The source presents a market in which demand for compute remained strong while financing conditions became more selective.
Silicon Data’s neocloud index showed hourly B200 rental prices rising from $5.78 to $6.47 between September 28 and October 27. H100 rental prices were reported to be 45% above the previous December’s level, despite the chip having launched in 2022 and Nvidia having released two subsequent product generations.
At the same time, the article reported that CoreWeave credit-default swaps reached record levels in mid-October, implying a five-year default probability above 60%. It cited Gartner’s forecast that more than 80% of neoclouds would cease operating independently by 2030.
The analysis also relayed Elon Musk’s warning that debt markets had largely closed to companies heavily dependent on AI data-center revenue. SpaceX was presented as an exception: it had signed a $40 billion finance lease for Nvidia Vera Rubin chips, led by parties including Apollo and Goldman Sachs and expected to close before year-end. SpaceX called it the largest GPU financing on record; the article noted its investment-grade rating and revenue from Starlink and rocket launches.
For GPU lenders, the tension is straightforward: rental demand supports repayment, but rapidly changing hardware can weaken both rental income and collateral resale value.
USD.AI’s loans run for approximately three years, not decades. Its eligible collateral list was restricted to newer chips: B200, B300, GB200, GB300, VR200, AMD MI455X and MI430X. A residual-value warranty costing 150 basis points annually, deducted from yield, covered part of the resale risk—not all of it.
Where sUSDai’s yield comes from
The stated CHIP Hurdle Rate was a 7% target, rather than a fixed borrower-paid return passed directly to depositors.
On November 1, sUSDai’s annualized yield was 6.53%. USD.AI contributed approximately 128,000 USDai to bring it to 7%. The source reported an average yield of 7.0% from January through early June 2026.
The difference between borrowers’ roughly 12% median interest rate and depositors’ roughly 7% yield reflected several factors:
- The annual 150-basis-point residual-value warranty charge.
- The protocol’s 10% share of interest.
- Undeployed cash earning Treasury rates rather than GPU-loan rates.
DefiLlama figures cited in the article showed $4.87 million in October fees and $487,000 in protocol revenue. Reported gross revenue over the preceding year was $24.75 million. Origination fees could produce spikes, including approximately $3 million on June 4, the day before the Duos loan announcement.
Deployment mattered as much as the quoted lending rate. Deposits had approached $700 million while loans were around $1 million or less, with inflows linked to airdrop expectations. By October 15, approximately $311 million was deployed: $177 million in active loans and $134 million in escrow awaiting chip delivery. DefiLlama’s roughly $309.5 million active-loan figure was close to the combined amount, rather than funded, operating loans alone.
The liquidity mismatch behind GLX
sUSDai redemptions were processed in 30-day batches, subject to available liquidity, while the underlying loans generally lasted about three years. A redemption queue can slow withdrawals; it does not make long-term loans immediately liquid.
Launched on October 1, GLX was designed to recycle capital. sUSDai could fund a loan, allow it to establish several months of repayment history, then sell it so that the cash could finance another borrower. Outside lenders could also participate at origination.
On October 29, Silver Point Capital and Victory Park Capital joined USD.AI in a $106 million, four-year loan to NexGen Cloud, financing 1,728 B300 GPUs in Finland. The article reported more than $200 million in cumulative GLX loan turnover.
Reported sale rules included:
- A normal minimum price of 100, or par.
- A nonbinding commercial guideline of 100.5 or higher.
- A 99.5 floor permitted only when redemption requests still exceeded available cash after the 30-day queue.
- A prohibition on buyers having an economic relationship with the proposal’s sponsor.
Gains on sales were split 60% to GLX operator USD.AI Inc. and 40% to the Yield Reserve, which supported the 7% sUSDai target. Discounted sales did not generate such a split.
The source illustrates the economics with a $10 million loan sold at 100.5: the $50,000 gain would split into $30,000 and $20,000. Re-originating another $10 million loan at USD.AI’s quoted 2%–3% origination fee could generate $200,000–$300,000. The larger commercial incentive was repeat origination, not the sale premium alone.
What the Duos exit demonstrated
The article identified Duos Edge AI as the only loan that had exited at the time of publication.
On June 5, 2026, USD.AI provided a $98.1 million, three-year loan to the subsidiary of Nasdaq-listed Duos Technologies, secured by 2,304 B300 chips. In September, GPU operator Axe Compute bought the entire SPV and repaid approximately $87.8 million of debt.
Axe acquired a compute cluster with a five-year lease and expected revenue of $364.6 million. USD.AI recovered most of its funding in less than four months.
That transaction showed one route to early capital recovery: selling the operating entity with its equipment and lease. It did not establish that every loan could be sold on comparable terms during market stress.
CHIP buybacks depend on lending economics
The article said 100% of sUSDai’s net interest margin, after CHIP Foundation costs, was allocated to programmatic CHIP purchases, without a fixed amount or timetable. Origination fees stayed with USD.AI Inc., which could use surplus funds for CHIP purchases. USDai’s fee switch was not enabled.
Using the October 26 purchase price, the source put CHIP’s circulating market value at approximately $90 million, based on 2 billion tokens at $0.0448 each. It estimated annual programmatic purchases equivalent to roughly 2.5% of circulating supply using the trailing-year average, or 6.5% at the then-current pace. These were rate-based estimates, not commitments.
Separately, on October 26, two days before the buyback router began operating, USD.AI Inc. manually spent 1 million USDC on approximately 22 million CHIP, at an average $0.044756—about 1.1% of circulating supply.
Reward emissions had fallen from 3% in season one to 2.25% in season two. The stated longer-term goal was for purchases to exceed reward issuance, followed by monthly, revenue-funded rewards.
A financing niche, not a replacement for banks
The source reported $48.4 billion in GPU-backed debt raised in 2026, with on-chain lenders holding less than 1%. USD.AI accounted for 93% of the on-chain segment, followed by GAIB, Forum Markets and Falcon Finance.
Potential growth came from smaller operators underserved by banks and overseas demand. The article reported Nvidia’s sovereign-AI revenue had more than tripled to over $30 billion that year, with Canada, France, the Netherlands, Singapore and the UK among the main contributors. It also suggested possible U.S. data-center restrictions after the midterm elections could favor smaller sites; that remained a prospective development.
The limiting risks were equally concrete: fast hardware depreciation, concentrated exposure to smaller neoclouds, restricted recourse through SPVs and a mismatch between redemption windows and loan maturities. GLX could recycle capital, but loan sales still required buyers and acceptable prices.
The article’s conclusion was narrower than a claim that DeFi would replace bank lending: on-chain GPU credit was filling a financing gap. Growth in that business did not, by itself, imply a higher token price.
Original article on Odaily — credited to @PinkBrains_io; translated by AididiaoJP, Foresight News.