AI infrastructure borrowing was expanding rapidly by November 2026, but access to credit remained uneven. Large operators could finance GPU fleets through banks. Smaller GPU cloud businesses increasingly looked to stablecoin-funded lenders.
A November 9, 2026 report attributed to @PinkBrains_io, translated into Chinese by AididiaoJP of Foresight News and published by Odaily, examined that gap through USD.AI. Its central distinction: growing demand for compute can support a lending business without making its collateral safe—or guaranteeing gains for its token.
Rising GPU rents, increasingly selective credit
The report cited Elon Musk’s response to Chamath that debt markets had effectively closed over the preceding weeks to businesses heavily dependent on AI data-center revenue and lacking substantial alternative income.
SpaceX was presented as an exception. It had signed a $40 billion finance lease to acquire Nvidia Vera Rubin chips, with Apollo and Goldman Sachs among the lead arrangers and completion expected before year-end. SpaceX described it as the largest GPU financing ever. Its investment-grade rating and revenue from Starlink and rocket launches distinguished it from GPU-only operators.
Musk expected extensive consolidation among more than 200 “neoclouds”—specialist GPU rental businesses including CoreWeave, Nebius, Lambda and Crusoe. The report also cited:
- CoreWeave credit-default swaps reaching a record in mid-October, implying a five-year default probability above 60%.
- Gartner’s forecast that more than 80% of neoclouds would cease operating independently by 2030.
Yet rental demand remained strong. Silicon Data’s neocloud index showed B200 hourly rental prices rising from $5.78 to $6.47 between September 28 and October 27. H100 rental prices were 45% above the previous December, despite the chip’s 2022 launch and two subsequent Nvidia product generations.
That combination—strong rental demand alongside worsening credit concerns—is the setting for onchain GPU lending. It is not evidence that demand eliminates financing risk.
Why banks leave smaller borrowers behind
American Compute’s figures, as cited in the report, identified 110 publicly disclosed GPU and data-center financings totaling $197 billion since 2023. Of that funding, 92% was raised in 2025 and 2026.
Banks dominated the disclosed lending market. Of 49 transactions identifying lenders since July 2025, 32 were led by JPMorgan, Goldman Sachs, Morgan Stanley or MUFG.
Pricing depended heavily on the tenant behind the rental contract:
| Financing category | Rates reported for 2026 |
|---|---|
| Backed by investment-grade leases | SOFR plus 2.25–3.0 percentage points |
| Unrated or sub-investment-grade | SOFR plus 4.5–5.5 percentage points, or fixed rates of 9–10% |
SOFR is an overnight borrowing benchmark secured by US Treasuries. CoreWeave’s loan—the largest in the report’s comparison—was priced at SOFR plus 5.50 percentage points, three percentage points above Nebius’s.
Large transactions commonly use a special-purpose vehicle, or SPV, that owns the chips, collects rent and carries the debt separately from the parent company. But the structure becomes less economical at smaller sizes. Citing The Token Dispatch, the report noted that bank due-diligence costs do not shrink proportionately with a loan: manageable costs on a $1 billion transaction can make a $30 million deal uneconomic.
The opening for onchain lenders is therefore not necessarily cheaper credit. It is access for borrowers too small to justify conventional underwriting costs.
USD.AI: hardware-backed credit at higher rates
USD.AI uses stablecoin deposits to finance loans secured by GPUs and rental contracts. Its October report listed 20 financings totaling $531 million, ranging from hundreds of thousands of dollars to $128.9 million.
The borrower’s SPV owns the GPUs. USD.AI holds first-priority liens on the servers and SPV equity. Loans typically cover 70–80% of hardware costs, with the borrower contributing the balance. Funds remain in escrow until installation is verified, after which rental income services the debt.
Its advertised pricing ranged from 7% for investment-grade contracted deployments to 15% for deployments without contracts. Among 15 loans tracked from January through August 2026, rates ranged from 10% to 15%, with a 12% median. USD.AI claimed it could complete financing within seven days.
If a borrower defaults, the agent first seeks a replacement operator or a sale of the operating SPV. Selling individual servers through an IT asset-disposal provider is the fallback.
The source compared this collateral to a car rather than a house: GPUs can become obsolete within two or three years. Typical USD.AI loans run for roughly three years. Eligible chips listed in the report were B200, B300, GB200, GB300, VR200, and AMD’s MI455X and MI430X. A residual-value warranty costs 150 basis points annually, deducted from yield, and covers part of the resale risk.
Where sUSDai’s yield comes from
The report described USDai as backed one-for-one by PYUSD, whose reserves include US Treasuries. USDai itself does not earn interest. Staking it produces sUSDai, the vault instrument that funds loans and accumulates interest; sUSDai holders bear the GPU credit exposure.
Redemptions are processed in 30-day batches, subject to available liquidity, against loans generally lasting about three years.
The stated yield target, called the CHIP Hurdle Rate, was 7%. On November 1, sUSDai’s annualized yield was 6.53%; USD.AI added approximately 128,000 USDai to bring it to 7%. Average yield from January through early June 2026 was also reported at 7.0%. That target should not be confused with a return generated entirely by borrower payments in every period.
The gap between approximately 12% median borrower interest and approximately 7% depositor yield reflected the residual-value warranty, a protocol charge of 10% of interest, and idle funds earning Treasury rates rather than GPU-loan rates.
DefiLlama figures cited in the report showed October fees of $4.87 million and protocol revenue of $487,000. Reported gross revenue over the preceding year was $24.75 million. Origination fees could create spikes, including roughly $3 million on June 4, the day before the Duos loan announcement.
Deposits and lending had not always grown together. Deposits approached $700 million at their peak while outstanding loans were only around $1 million; the report associated those inflows with airdrop expectations. By mid-October, approximately $177 million was in active loans and $134 million in escrow awaiting chip delivery. DefiLlama’s $309.5 million active-loan figure was close to their combined value, rather than funded, operating loans alone.
GLX recycles capital—but does not erase the mismatch
USD.AI launched GLX on October 1 to sell loans after several months of timely payments and recycle the proceeds into new lending. External 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 chips in Finland. Cumulative GLX turnover had exceeded $200 million, according to the report.
The sale rules distinguished normal pricing from liquidity-stress pricing:
- Ordinary sales must be at par, or 100, or higher.
- A price above 100.5 is commercial guidance, not a mandatory minimum.
- A 99.5 floor is permitted only when redemption requests still exceed available cash after the 30-day queue.
- Buyers cannot have an economic affiliation with the proposing party.
Gains on sales are split 60% to USD.AI Inc. and 40% to the Yield Reserve, which supports sUSDai’s 7% target. Discounted sales do not participate in that gain-sharing arrangement.
The economics favor repeated origination over small sale premiums. Selling a $10 million loan at 100.5 generates $50,000 in gains. Re-lending that capital at USD.AI’s quoted 2–3% origination fee could generate $200,000–$300,000.
The report identified one exited loan: USD.AI’s June 5, 2026, $98.1 million three-year loan to Duos Edge AI, a subsidiary of Nasdaq-listed Duos Technologies, secured by 2,304 B300 chips. Recourse was limited to the SPV and collateral, not Duos itself.
In September, Axe Compute bought the SPV and repaid approximately $87.8 million of debt. It acquired a cluster with a five-year lease and expected revenue of $364.6 million. USD.AI recovered most of its capital in less than four months.
That exit illustrates capital recycling, not proof that every loan will find a buyer. The report also noted that successful borrowers may refinance with banks: Sharon AI moved to bank financing at 9.95%.
More fundamentally, a 30-day redemption queue cannot eliminate the mismatch with multiyear loans. The source invoked Celsius as a warning about maturity mismatch—not evidence that USD.AI must follow the same path.
CHIP buybacks depend on lending economics
CHIP’s link to the business is the protocol’s 10% share of sUSDai interest. After CHIP Foundation costs, 100% of that net interest margin was designated for programmatic CHIP purchases, without a fixed purchase amount or schedule.
Origination fees remain with USD.AI Inc., which may use surplus funds to buy CHIP. Loan-sale gains follow the separate 60/40 split. The USDai fee switch was not active.
Using the October 26 purchase price, the report put circulating market capitalization near $90 million, based on 2 billion tokens at $0.0448 each. Estimated annual programmatic purchases equaled roughly 2.5% of circulating supply using trailing-year averages, or 6.5% using the then-current pace. These were pace-based calculations, not commitments.
On October 26, two days before the buyback routing began, USD.AI Inc. also manually spent 1 million USDC on approximately 22 million CHIP at an average $0.044756—about 1.1% of circulating supply.
The stated longer-term goal was for purchases to exceed reward issuance. Emissions declined from 3% in season one to 2.25% in season two; the report said rewards would shift to a monthly, revenue-funded arrangement once purchases exceeded new issuance.
A financing niche, not a guaranteed token cycle
The report put 2026 GPU-backed debt fundraising at $48.4 billion, with onchain lenders holding less than 1%. USD.AI accounted for 93% of that onchain segment, followed by GAIB, Forum Markets and Falcon Finance.
Its illustrative calculation—12% loan interest and a 10% protocol share—suggested that capturing 1% of that market could fund roughly $5.8 million in annual CHIP purchases, about 6.5% of the cited circulating market capitalization. A 5% share would imply roughly one-third. Those figures depended on the stated assumptions; they were not forecasts of market share or token returns.
Potential demand extended beyond the US. The report said Nvidia’s sovereign-AI revenue had grown more than threefold that year to above $30 billion, with Canada, France, the Netherlands, Singapore and the UK major contributors. Some deployments could reach midsized local operators such as NexGen Cloud and Sharon AI. Possible US data-center restrictions after the midterm elections were another prospective driver of smaller sites, not an established outcome.
The limiting risks remained concrete: rapid chip depreciation, potentially weaker rental and resale prices as new hardware arrives, vulnerable smaller borrowers, and a redemption schedule much shorter than the underlying loans.
Onchain GPU finance was filling a gap banks left behind, not replacing banks. More lending could expand protocol revenue, but neither strong compute demand nor buybacks alone established a bullish token outcome.
Original publication: November 9, 2026, 19:00 UTC+8. Read the original article on Odaily.
