AI Compute Debt Is Soaring: Can On-Chain Credit Fill the Bank Gap?

Original Author: @PinkBrains_io

Original Compilation: AididiaoJP, Foresight News

AI Compute Debt Soars, Onchain Credit Fills the Bank Gap

Musk poured cold water on the AI compute frenzy. In a response to Chamath, he said that over the past few weeks, the debt market has essentially closed its doors to one type of company: those whose revenue depends heavily on AI data centers and that lack other large sources of income.

SpaceX is an exception. It signed a $40 billion financial lease to purchase Nvidia Vera Rubin chips, led by Apollo and Goldman Sachs, expected to close before year-end. SpaceX calls it the largest GPU financing in history. It has an investment-grade rating and two revenue streams: Starlink and rocket launches.

Musk's assessment is even more stark: more than 200 neoclouds will face massive consolidation and shakeout. Neoclouds are cloud companies that only rent out GPUs, including CoreWeave, Nebius, Lambda, Crusoe, and hundreds of smaller companies. Amazon, Google, and Microsoft are "hyperscalers."

The credit market is already pricing this in. In mid-October, CoreWeave's credit default swaps hit a record, implying a default probability of over 60% within the next five years. Gartner predicts that by 2030, more than 80% of neoclouds will cease to operate independently.

But compute demand is still rising. Silicon Data's neocloud index shows that from September 28 to October 27, the hourly rental rate for B200 on specialized GPU clouds rose from $5.78 to $6.47. Even more telling is the H100: it launched in 2022, and Nvidia has since released two new generations of products, so it should have depreciated—yet its rental rate is 45% higher than last December. Demand is so strong that even older chips are in high demand.

When demand is strong, operators need to buy more GPUs. GPUs are expensive and expansion is fast, so their own capital isn't enough—they have to borrow. But bank money only flows to large neoclouds (GPU cloud providers), and smaller operators with only a few hundred GPUs mostly can't get loans.

This creates a financing gap: small operators have rental income and chips but can't get bank loans. So GPUs are starting to become collateral, and onchain credit backed by stablecoin deposits is stepping in to fill the segment that banks are unwilling to cover.

1. Compute Becomes a Debt Business, Banks Only Do Big Deals

AI compute has already become a debt business. According to American Compute, since 2023 there have been 110 public GPU and data center financings, totaling $197 billion in new capital, of which 92% was raised in 2025 and 2026.

Banks dominate the market. Since July 2025, of 49 transactions with clearly identified lenders, 32 were led by JPMorgan, Goldman Sachs, Morgan Stanley, or MUFG.

Pricing depends on the tenant, not the chip. In 2026, loans backed by investment-grade lease contracts carry rates of SOFR+2.25% to 3.0%; unrated or sub-investment-grade loans carry rates of SOFR+4.5% to 5.5%, or 9% to 10% fixed. SOFR is the overnight borrowing rate collateralized by U.S. Treasuries and is the pricing basis for most U.S. loans.

CoreWeave took the largest loan on the list and also pays the highest rate. Its loan rate is 5.50 percentage points above the SOFR benchmark, 3 percentage points more expensive than Nebius. This level has hit the upper end of the range banks offer sub-investment-grade borrowers—borrowing the most and paying the most.

Large buyers typically don't borrow directly at the corporate level but use SPVs. An SPV is a separate legal entity that buys chips, rents them out, and carries the debt, placing the chips, leases, and debt outside the corporate entity.

This structure works at the top, but as you go down the scale, it stops making economic sense. According to The Token Dispatch, bank due diligence costs don't vary with transaction size. On a $1 billion loan, it's a rounding error; on a $30 million loan, due diligence costs can kill the deal outright.

The key change is this: banks control GPU credit pricing, but only cover borrowers large enough to justify due diligence. Who serves small neoclouds? Onchain credit.

2. GPU Collateralized Lending: Like Auto Loans, Not Mortgages

USD.AI is one example. It lends out stablecoin deposits, taking GPUs and lease contracts as collateral. Its October report showed 20 financings originated, totaling $531 million, ranging from hundreds of thousands of dollars to $128.9 million per deal.

It brings the bank SPV structure down to mid-sized transactions. The borrower's SPV owns the GPUs, and USD.AI holds a first-priority lien on the servers and the SPV's equity. Loans typically cover 70% to 80% of hardware costs, with the borrower paying the rest; funds are held in escrow and only disbursed after server installation is verified. Rental income then repays the loan.

Borrowers pay more than at banks. USD.AI's rate card ranges from 7% for investment-grade lease contracts to 15% for deployments without contracts. For 15 loans tracked from January to August 2026, rates ranged from 10% to 15%, with a median of 12%. Banks charge sub-investment-grade borrowers about 9% to 10% fixed. The selling point isn't price—it's access and speed: USD.AI says loans can be completed within 7 days.

What happens in case of default? The agent first tries to have the SPV switch operators and continue running, or sell the whole thing to repay the loan; only if that fails does it sell servers one by one through IT asset disposal firms.

GPU collateralized lending is more like auto loans than mortgages. A house can last 100 years; a GPU can become obsolete in two or three. So USD.AI loan terms are about 3 years, not 30; collateral rules only accept the latest chips: B200, B300, GB200, GB300, VR200, as well as AMD's MI455X and MI430X. A residual value warranty of 150 basis points per year is deducted from yield to cover part of the resale risk.

3. Where sUSDai's Yield Comes From

USDai is 1:1 backed by PYUSD, whose reserves include U.S. Treasuries; USDai itself does not bear interest. Staking yields sUSDai, a vault that funds loans and accrues interest. Only sUSDai holders bear GPU credit risk. Redemptions are processed in batches every 30 days, subject to available liquidity, with loan terms of about 3 years.

The target yield is called the CHIP Hurdle Rate, set at 7%. On November 1, sUSDai's annualized yield was 6.53%; USD.AI injected about 128,000 USDai to push it to 7%. From January to early June 2026, the average yield was 7.0%.

Borrowers pay a median rate of about 12%, depositors earn about 7%, and the spread flows to three places: the 150-basis-point residual value warranty, the protocol's 10% cut of interest, and idle cash sitting at Treasury rates while waiting to be lent out.

DefiLlama shows October fees of $4.87 million and protocol revenue of $487,000. Revenue is 10% of fees, consistent with the net interest margin the protocol retains from sUSDai yield. Over the past year, gross protocol revenue was $24.75 million; origination fees inflate the protocol's share, including a revenue peak of about $3 million on June 4, the day before the Duos loan was announced.

At one point deposits far exceeded loans. At a peak of nearly $700 million, loans were only about $1 million, with inflows tied to airdrop expectations. By mid-October, active loans were $177 million, with another $134 million in escrow awaiting chip delivery; DefiLlama's $309.5 million in active loans is close to the sum of the two. Idle cash still earns interest, but only at Treasury rates. GLX exists to deploy idle capital faster.

4. GLX: Sell the Loan, Then Lend Again

For USD.AI, holding a 3-year loan means bearing 3 years of default risk. And once borrowers scale up, they may switch to cheaper bank loans—Sharon AI moved to a bank at a 9.95% rate.

There's pressure on the funding side too. Depositors can redeem every 30 days, but loans run up to 3 years. Similar maturity mismatches brought down Celsius. A 30-day queue can slow runs but can't eliminate the mismatch. Much of the deposits were also chasing airdrops: at a peak of nearly $700 million, less than $1 million was lent out. By October 15, about $311 million was deployed, including $177 million in active loans and $134 million in escrow.

GLX launched on October 1, breaking the one-to-one correspondence between deposits and loans. sUSDai funds a loan, the loan is "seasoned" through several months of on-time payments, then sold, and the cash is redeployed into new loans. External lenders can also join at origination. On October 29, Silver Point Capital and Victory Park Capital joined USD.AI in providing a $106 million 4-year loan to NexGen Cloud for 1,728 B300s in Finland. GLX has now circulated over $200 million cumulatively.

On pricing, the minimum sale is at par of 100—that is, no discount, sold at original price. Above 100.5 is a "commercial guidance price," encouraging sales at that level but not a hard minimum. 99.5 is the floor, allowed only in one case: after the 30-day redemption queue, redemption requests still exceed available cash—that is, only when liquidity is tight is a slight discount permitted.

On economics, 60% of sale proceeds go to USD.AI Inc., which operates GLX, and 40% to the Yield Reserve. The Yield Reserve is an earnings reserve used to top up sUSDai's yield toward the 7% target. Discounted sales don't participate in the split.

On buyers, buyers cannot have an economic relationship with the proposing party, preventing self-dealing.

For example: a $10 million loan sold at 100.5 generates $50,000, split $30,000 and $20,000. At USD.AI's quoted 2% to 3% origination fee, relending that $10 million earns $200,000 to $300,000. The money is made in relending, not in the sale price.

The only loan that has exited so far came from Duos Edge AI.

On June 5, 2026, USD.AI issued a $98.1 million 3-year loan to Duos Edge AI, a subsidiary of Nasdaq-listed Duos Technologies, collateralized by 2,304 B300 chips. If the loan defaults, USD.AI can only pursue the SPV and the chips, not Duos itself—the risk is isolated in the SPV.

In September, another GPU operator, Axe Compute, bought the entire SPV and repaid its roughly $87.8 million loan. Axe got a compute cluster with a signed 5-year lease and expected revenue of $364.6 million.

For USD.AI, the loan recovered most of its capital in under 4 months and can be redeployed into new lending.

5. CHIP Buybacks and the Scale Ceiling

For CHIP holders, all of the above ultimately comes down to the protocol's 10% cut of sUSDai interest. The more the vault lends, the larger the cut.

GLX cuts both ways. Selling a loan takes away the interest, and origination fees go to USD.AI Inc.; but the cash comes back and is relended at about 12% instead of sitting at Treasury rates.

Buybacks have already begun. sUSDai's net interest margin, after deducting CHIP Foundation costs, is 100% used for programmatic purchases of CHIP. USD.AI has not set a fixed amount or schedule. Origination fees stay with USD.AI Inc., and any excess can buy CHIP; loan sale proceeds are split 60/40. USDai's fee switch has not yet been turned on.

At the October 26 buyback price, CHIP's circulating market cap was about $90 million, corresponding to 2 billion tokens at $0.0448 each.

How much the buyback program can purchase in a year depends on which basis you use: at the past 12 months' average, about 2.5% of circulating supply per year; at the current pace, about 6.5%.

On October 26, two days before the buyback router launched, USD.AI Inc. also manually bought about 22 million CHIP with 1 million USDC at an average price of $0.044756, about 1.1% of circulating supply.

That single purchase was equivalent to 5 to 6 months of the program at the historical basis, or 2 months at the current pace—buying urgently and heavily.

The long-term goal is for buybacks to exceed reward emissions. Emissions have already dropped from 3% in Season 1 to 2.25% in Season 2; once purchases exceed new issuance, rewards will shift to a monthly, revenue-funded arrangement. Most DeFi buybacks rely on trading fees; CHIP buybacks rely on GPU loan interest and can only grow with loan scale.

Scale is still small. In 2026, GPU-backed debt raised $48.4 billion, and onchain lenders hold less than 1%. At roughly 12% loan rates and a 10% protocol cut, each 1% of market captured would fund about $5.8 million a year in CHIP buybacks, about 6.5% of current circulating market cap; 5% would be about one-third. USD.AI accounts for 93% of the onchain segment, with GAIB, Forum Markets, and Falcon Finance behind it.

6. The Growth Boundaries of Onchain GPU Financing

GPU loans work because the collateral generates rental income, and rents are still rising. The risk is also in the asset: GPUs depreciate quickly, and Nvidia's next generation could simultaneously push down rental and resale prices, while the small neoclouds served by onchain lenders are precisely the group most likely to be consolidated.

But there are also plenty of reasons for growth. Banks have essentially closed their doors to small operators, and state regulatory rules are inconsistent, giving an edge to lenders that can complete funding within days. Data center restrictions that may emerge after the U.S. midterm elections could push construction toward smaller sites, better suited to a model of fast, cluster-by-cluster lending.

Demand outside the U.S. is also rising. Nvidia's sovereign AI business revenue grew more than twofold this year to over $30 billion, with Canada, France, the Netherlands, Singapore, and the U.K. as major contributors. Some of that capacity will come online through local neoclouds such as NexGen Cloud and Sharon AI, precisely the mid-sized borrowers onchain lenders are now targeting.

Currently, onchain GPU finance still accounts for less than 1% of total GPU-backed debt in 2026. It fills the gap banks are unwilling to cover rather than replacing them. And the protocol's own growth does not equal token price appreciation.