Debt is becoming the defining constraint on the AI computing boom, with banks directing large facilities toward established GPU cloud operators while smaller providers seek higher-cost, stablecoin-funded credit secured by servers and customer leases. American Compute has counted 110 disclosed GPU and data-center financings worth $197 billion since 2023, with 92% of that capital raised during 2025 and 2026.
The divide leaves mid-sized neocloud companies businesses focused chiefly on renting GPU capacity in an awkward position. Demand for Nvidia's newest chips remains strong, but traditional lenders increasingly favor borrowers large enough to support extensive underwriting, legal work and collateral monitoring. On-chain lenders are filling part of the gap, though at interest rates that reflect the rapid depreciation risk of AI hardware and the uneven credit quality of GPU renters.
Credit markets have already signaled concern about the sector's weaker operators. In mid-October, credit default swaps on CoreWeave reached a record level implying a more than 60% probability of default over five years, according to market pricing cited by American Compute. Gartner has projected that more than 80% of neoclouds will no longer operate independently by 2030, underscoring the pressure on companies that must keep buying new hardware while competing with much larger cloud platforms.
GPU rental demand remains elevated
Rising financing risk has not coincided with a collapse in GPU pricing. Silicon Data's neocloud index showed hourly rental rates for Nvidia B200 chips increasing from $5.78 on Sept. 28 to $6.47 on Oct. 27. Rates for older H100 processors were 45% above their level last December, despite the chip's 2022 launch and Nvidia's subsequent release of newer generations.
Those figures help explain why providers continue to borrow heavily for equipment. A leased GPU cluster can produce recurring revenue quickly if it is installed and contracted, yet the economics turn difficult when deployment is delayed, utilization falls, or a new chip generation reduces the value of older machines. The useful financing life of a high-end GPU is often measured in years rather than the longer timeframes familiar in conventional infrastructure lending.
Bank credit has concentrated around the largest transactions. Since July 2025, JPMorgan, Goldman Sachs, Morgan Stanley and Mitsubishi UFJ led 32 of 49 disclosed deals with named lenders, according to American Compute's deal tracking. Such concentration reflects the costs of underwriting facilities backed by thousands of servers, complicated lease arrangements and specialized-purpose entities.
The price of borrowing depends heavily on who rents the machines. American Compute found that 2026 loans supported by investment-grade lease contracts were priced at roughly the Secured Overnight Financing Rate, or SOFR, plus 2.25 to 3 percentage points. Unrated and sub-investment-grade borrowers paid SOFR plus 4.5 to 5.5 points, or fixed rates around 9% to 10%.
CoreWeave, which secured the largest loan in American Compute's dataset, paid SOFR plus 5.5 percentage points roughly 3 percentage points more than Nebius. That pricing places even large operators with substantial contracted revenue closer to high-yield credit conditions than to conventional asset finance.
Smaller borrowers turn to stablecoin credit
The bank model becomes much less efficient for smaller facilities. Large buyers can place GPUs, lease contracts and related debt inside special-purpose vehicles, or SPVs, separating the financed equipment from the main operating company. The same structure is harder to justify for a $30 million loan than for a $1 billion facility, because legal, engineering and collateral-review costs consume a far larger share of the transaction.
That opening has fueled GPU-collateralized lending through on-chain protocols. USD.AI reported in October that it had originated 20 financings totaling $531 million, with loans ranging from several hundred thousand dollars to $128.9 million. Its structure uses stablecoin deposits to finance borrowers that place GPU hardware in SPVs.
The lender generally receives a first-priority claim on both the servers and the SPV's equity. Loans typically cover 70% to 80% of hardware costs, while the borrower funds the balance. Capital is released after installation is verified, a safeguard designed to ensure that the financed equipment exists and has been deployed.
Speed is the appeal for borrowers, but it comes with higher rates. Across 15 loans tracked between January and August 2026, borrowing costs ranged from 10% to 15%, with a median of 12%. That compares with roughly 9% to 10% fixed rates for sub-investment-grade bank borrowers in American Compute's dataset.
USD.AI has kept its lending criteria narrow. The protocol's typical loan tenor is about three years, and it has accepted only recent-generation equipment, including Nvidia's B200, B300, GB200, GB300 and VR200 models, as well as AMD's MI455X and MI430X. It also charges a 150-basis-point annual residual-value warranty intended to account for the possibility that collateral loses value before a loan is repaid.
Depositors take the equipment credit risk
USD.AI separates its stablecoin from the token that earns lending returns. USDai is backed one-for-one by PYUSD, whose reserves include U.S. Treasuries. Depositors who stake USDai receive sUSDai, a yield-bearing token whose holders are exposed to GPU loans. Redemptions are processed in 30-day batches, subject to available liquidity.
The protocol has targeted a 7% return under what it calls its CHIP Hurdle Rate. On Nov. 1, sUSDai's annualized yield was 6.53% before an injection of roughly 128,000 USDai brought the rate to 7%, according to USD.AI. The average yield from January through early June was also reported at 7%.
That return is funded by the spread between borrower costs and depositor yield. With borrowers paying a median rate near 12% and sUSDai holders targeted for 7%, the remaining income supports the residual-value warranty, a protocol share of interest payments, and returns earned on idle capital held in Treasuries before loans are deployed.
DefiLlama listed USD.AI fees of $4.87 million in October and protocol revenue of $487,000, consistent with the protocol retaining 10% of fees. It recorded $24.75 million in gross protocol revenue over the previous year, including a roughly $3 million spike on June 4, a day before USD.AI announced a loan to Duos Technologies.
The model has also faced the practical problem of matching stablecoin deposits with equipment purchases. At one point, deposits approached $700 million while loans stood near $1 million. By mid-October, active loans were reported at $177 million, alongside $134 million in escrow awaiting chip delivery. DefiLlama later showed $309.5 million in active loans, approximately matching the combined figure.
Loan sales offer a route to recycle capital
USD.AI launched GLX on Oct. 1 to sell seasoned loans and recycle proceeds into new financing. On Oct. 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. GLX cumulative turnover subsequently exceeded $200 million.
GLX's terms provide for loan sales at a minimum price of 100, or par. The platform sets non-binding commercial guidance above 100.5, while a 99.5 floor is allowed only when redemption requests remain larger than available cash after the 30-day queue. Sale proceeds are divided 60% to USD.AI Inc., which operates GLX, and 40% to a Yield Reserve supporting sUSDai returns. Discounted sales do not receive that split.
An early exit involving Duos Edge AI offers a test case for the collateral structure. On June 5, USD.AI issued a $98.1 million, three-year loan secured by 2,304 B300 chips. In September, GPU operator Axe Compute acquired the SPV and repaid roughly $87.8 million, taking over a cluster tied to a five-year lease with expected revenue of $364.6 million.
On-chain lending remains marginal beside the conventional GPU-debt market. American Compute recorded $48.4 billion in GPU-supported debt issuance during 2026, while on-chain lenders accounted for less than 1% of the total. USD.AI represented 93% of the on-chain segment, according to reported market-share figures, ahead of GAIB, Forum Markets and Falcon Finance.
That concentration makes the sector's expansion dependent on whether protocols can consistently enforce collateral standards, manage chip obsolescence and sell loans without disrupting depositor liquidity. GPU demand and rental pricing have so far supported aggressive hardware purchasing, but the funding structure ultimately rests on the cash flow generated after the servers are installed not on the tokens issued around the financing.
Explore how on-chain private credit is reshaping financing for GPU-backed AI infrastructure in 2026 with Toobit Academy’s deep dive.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
