According to Woofun AI, the price of CoreWeave, Inc. (ASX: CRWV) credit default swaps (CDS) has swung violently, with the implied probability of default breaking past 60%. This market signal does not point directly to a bursting of the AI industry bubble. Instead, it provides a crucial revaluation anchor for the on-chain core yield asset USD.AI and its token sUSDai.
CoreWeave's debt financing structure and USD.AI's lending model are highly isomorphic at the underlying logic level. The only difference lies in counterparty credit quality. The former relies on giants such as Meta Platforms, Inc. (ASX: META) and OpenAI, while the latter mainly serves small and medium-sized neocloud vendors. This structural similarity means that the traditional financial market's risk pricing of CoreWeave is, through a mapping effect, redefining the true risk premium of on-chain GPU credit. A deep analysis of this transmission mechanism is the key to understanding the current valuation split in the Web3 credit market.
CDS, as an insurance tool for measuring debt default risk, is quoted in basis points, with 100 basis points corresponding to a 1% annual premium rate. CoreWeave's five-year CDS hit a peak of 1194.5 basis points intraday on October 15. This means investors holding its bonds must pay an annual premium of 11.945% to hedge default risk. This figure is nearly 20 times the investment-grade corporate CDS index (about 60 basis points). The market inferred a 60% implied default probability from this, but this is not a direct bankruptcy probability. It is a risk-neutral probability based on specific recovery rate assumptions. It is worth noting that this indicator is dynamic. By October 19, the CDS price had fallen back to 1082 basis points. This fluctuation reveals market concerns about CoreWeave's short-term liquidity pressure, rather than a final verdict on its long-term viability. Understanding the essence of this data is the prerequisite for stripping away market noise and gaining insight into true risk stratification.
As of the first half of 2026, CoreWeave's total liabilities were approximately $52 billion. This includes about $35 billion in total debt, about $16.3 billion in operating leases, and only about $5 billion in shareholders' equity. The leverage ratio is as high as around 10 times. Its $35 billion debt structure shows clear stratification. The first layer is the largest portion, the DDTL (Delayed Draw Term Loan), with about $15 billion drawn and about $9 billion undrawn. The core risk control logic of DDTL lies in bankruptcy isolation. Funds are not lent directly to CoreWeave's parent company. Instead, they are injected into independent project companies (SPVs). Each SPV is tied to specific GPU assets, customer contracts, and data center leases. Banks focus their assessment on the credit of the contract counterparty (such as Meta Platforms, Inc.) rather than CoreWeave itself. The second layer is convertible bonds, with a scale of about $11 billion and coupons ranging from 1.75% to 2.875%. The third layer is unsecured notes, which triggered market panic. With a scale of about $1.5 billion, they are structurally subordinated debt, ranking behind tens of billions of dollars in secured claims in the liquidation order. This structure indicates that the CDS surge mainly reflects the risk of the most subordinated claims, not the entire debt system.
Barclays analyst Rajadhyaksha vividly pointed out that this market turmoil is like questioning the solvency of mortgage borrowers across an entire city because of high interest rates on small credit cards. It is a mismatch of reference frames. Based on this logic, Goldman Sachs upgraded CoreWeave's rating from 'Neutral' to 'Buy' on October 19, setting a target price of $106 (the stock price was $61.39 at the time). Goldman Sachs' revision was based on the fact that CoreWeave's true weighted financing cost should be calculated based on annualized interest expense of $2.56 billion for the second quarter divided by total debt of about $30 billion. The result is about 800 to 900 basis points, far below the 15% implied by market panic. The 15% yield applies only to that $1.5 billion in subordinated notes. It cannot represent CoreWeave's overall financing cost. This revision of institutional views provides an important benchmark for reassessing on-chain credit products with similar structures.
Comparing CoreWeave's DDTL model with USD.AI's loan structure reveals that the two are highly consistent in risk control logic. Both follow the principle of 'not looking at borrower credit, but looking at contract counterparty credit.' USD.AI divides loans into three tiers. Tier 1 requires the off-taker to be investment grade, with yields of 7–9%. Tier 2 is non-investment grade but has multi-year contracts, with yields of 10–12%. Tier 3 is spot scattered leasing, with yields of 12–15%. This tiering essentially codifies the risk control logic of DDTL. However, the key difference lies in counterparty quality. CoreWeave's tenants include hyperscale cloud providers such as Meta Platforms, Inc., OpenAI, and Microsoft. In contrast, USD.AI's publicly disclosed borrowers are mostly small and medium-sized neocloud vendors such as QumulusAI, Corvex, Hydra Host, Duos Edge AI, and Sharon AI. JPMorgan clearly stated in its October 20 report that 'not all project companies have hyperscaler backing.' This risk is more prominent in USD.AI's asset portfolio. Therefore, USD.AI is not safer because it has collateral. It takes on higher risk because of lower contract quality.
A deeper systemic risk lies in the fact that although USD.AI's 13 borrowers appear diversified, they are all exposed to the same GPU rental price curve. This is not 13 independent risks. It is 1 systemic risk multiplied by 13. Once the rental curve declines, it could trigger concentrated defaults. This would then lead to a flood of GPUs entering the secondary market, causing liquidation prices to collapse. This forms a vicious cycle of 'default → liquidation price drop → insufficient collateral → more defaults.' The current GPU rental market shows clear divergence. On the contract side, CoreWeave raised prices across its entire product line by about 25% in July, reflecting tight supply and demand in the long-term contract market. On the spot side, H100 hourly rental fell from $3.5 in early 2025 to $1.29 in early 2026, a decline of 63%. B200 hourly rental fell from $6.1 in early June to $4.37 in July, a 31% drop in three weeks. Since collateral liquidation can only be executed at spot prices, this gap constitutes the most fragile link in the entire credit structure. Assuming a liquidation price of 30% of spot, this cyclical effect would be further amplified.
According to data compiled by Woofun AI, as of October 17, USD.AI's total deposits were $574.5 million. There were 15 active loans totaling $310.7 million. The weighted average borrowing rate was 10.46%, and the net APY paid to depositors was 6.25%. The 4.2 percentage point spread between the asset-side 10.46% and the liability-side 6.25% mainly stems from differences in capital deployment rates. Only $310.7 million (a deployment rate of 54.1%) was invested in interest-bearing loans. Of the remaining $263.8 million, some remained in Wilmington Trust custody accounts awaiting hardware installation verification (official guidance of up to 7%). Some was held as PYUSD reserves invested in 5-year US Treasuries (yield of 4.99%). Calculations show that 10.46%×54.1% + 7%×45.9% yields a gross return of 8.87%. After deducting about 2.6% in protocol fees and operating costs, this exactly matches the 6.25% net return. This means that of the 6.25% return depositors receive, more than one-third comes from bridge interest on 'undeployed funds,' not from true GPU credit risk premium. As the deployment rate rises toward the level corresponding to the official target of 12.4% yield, risk exposure will nearly double, while yield growth will be limited. This structural feature deserves vigilance.
On the credit loss path, USD.AI has set up three buffer mechanisms. First, it holds first-priority secured claims, with recourse limited to specific assets within the project company. Second, borrowers must contribute 20–25% of their own funds and establish a 10% debt service reserve and residual value insurance. Collateral must fall by more than a quarter before depositors are affected. Third, loans use straight-line amortization, with the loan balance dropping to 50% by month 12. Sensitivity analysis shows that even assuming a 50% default rate, a single loss would be 20% (loan balance of 50% minus liquidation price of 30%). Total losses would be about $31.07 million, or 5.41% of total deposits. This indicates that the threshold for credit losses is relatively high. However, the risk of a liquidity run is more insidious. sUSDai is not a stablecoin. Its net asset value fluctuates. Redemptions use a 30-day cycle with a first-come, first-served queue mechanism. This structure could evolve into a bank run during stress periods. Expectations of NAV decline trigger secondary market discounts. Discounts exacerbate redemption panic. The queue extends to 60 days or even longer, forming a self-reinforcing negative feedback loop. In this process, no loan needs to actually default. Liquidity exhaustion alone can cause sUSDai to trade at a significant discount.
Derivatives strategies further amplify this risk. On the Pendle platform, sUSDai's PT (principal token) trading yields are often above 9%. But note that the PT promises to return 1 sUSDai at maturity, not $1. Buying PT at a discount of $0.913, if sUSDai's NAV is still 1.00 at maturity, realizes a 9.5% return. If NAV falls to 0.913, it just breaks even. Below that, it is a loss. Therefore, the 9.5% fixed return can only absorb 8.7% of NAV downside risk. Moreover, PT lacks liquidity before maturity, making timely exit impossible. On the Morpho platform, looping strategies using sUSDai as collateral (such as the Frax USDai ecosystem vault and Bitwise RWA vault) have reached annualized yields as high as 27.13%. The arithmetic logic is: sUSDai yield of 6.25% multiplied by 5x leverage, minus 5% borrowing cost multiplied by 4x, yielding a net return of 11.25%. But under the 86%–91.5% LLTV range, risk is amplified simultaneously. Re7 Labs incurred $27.4 million in bad debt during the November 2025 Stream Finance event due to xUSD and deUSD de-pegging. The lesson was a failure to identify the looping collateral structure and lagging price feedback. sUSDai faces similar risks: a market discount can trigger liquidations without any actual credit loss. The looping strategy on Morpho is essentially a bet that sUSDai will not face a liquidity crisis, not a bet that borrowers will not default.
In summary, the economic substance of holding sUSDai is selling credit insurance to a basket of small and medium-sized neocloud vendors. In normal times, it collects a 6.25% premium (of which only 5.66 percentage points truly come from credit, with the rest being bridge interest). In extreme cases, it bears losses. Compared with CoreWeave CDS sellers who collect an 11.945% premium, sUSDai holders are clearly undercompensated for risk, and counterparty quality is weaker. Despite buffers such as 20% own funds and residual value insurance, the 10 percentage point spread reveals a severe split between public market and on-chain credit pricing. This split persists because on-chain GPU credit has not yet experienced a full credit cycle test. Since USD.AI issued its first $620,000 loan in July 2025, 15 months have passed without any defaults, with only one early repayment. The market lacks real data on secondary market liquidation prices for next-generation GPUs such as the B300, causing pricing to be based on untested assumptions. The probability of credit loss in the short term (3–6 months) is low, but discount risk may be self-fulfilling through expectations. The medium term (6–18 months) is the real stress test window. The first half of 2027 is expected to face a triple overlay of rising deployment rates, loans entering late amortization (collateral value dropping to 31% and 21%), and a declining rental curve. Investors are advised to monitor four key indicators: deployment rate, weighted borrowing rate, sUSDai secondary market discount, and changes in active loan balances. If participating, Morpho looping leverage should be limited to no more than 3x (liquidation only occurs if NAV falls 22.5%). Pendle should only be allocated to short-term PT. sUSDai should be treated as an equity position rather than a cash position. A barbell strategy should be used to manage risk. CoreWeave's CDS volatility provides the first public price signal for GPU-collateralized credit, prompting the industry to re-examine collateral ratios (recommended to drop to 60–70%) and amortization periods (recommended to shorten to 24–30 months) to address potential valuation reconstruction.
Comments