Compute companies expand with high-interest debt: Anthropic borrows $60 billion to buy TPUs, CoreWeave default probability exceeds 60%

[[**The compute power race is expanding on 10%–15% high-yield debt and circular financing: Anthropic borrows $60 billion to buy TPUs, CoreWeave's probability of default exceeds 60%, funders are tightening, and the leverage chain faces repricing.**

**Key points**

**1. Anthropic borrows $60 billion at a coupon above 10% to buy TPUs**

Blackstone invests $9 billion; the debt is backed by five years of chip lease payments; comparable BB-rated debt is about 7%. The company posted a net loss of $42 billion in 2025 and plans to invest $518 billion in the future.

**2. CoreWeave's probability of default exceeds 60%**

Bond yields break above 15%, and CDS pricing implies a five-year default probability above 60%; GPU-collateralized financing is about $29.3 billion, and total debt exceeds $40 billion.

**3. High-yield debt spreads to more compute players**

Volta's $5 billion loan yields more than 11%, with banks underwriting only $1.3 billion; xAI has about $12.5 billion in debt to purchase GPUs; Hongxin Electronics purchases chips for 2.983 billion yuan, equal to 88.36% of its first-half revenue.

**4. Circular financing raises fragility, and funders tighten**

Nvidia holds about 10% of CoreWeave and has committed to buying unsold compute through 2032; it is reportedly discussing about $250 billion in financing guarantees with OpenAI. Wells Fargo says cheap money has disappeared, and Goldman Sachs warns of an order cliff in 2027–2028. 2026-10-23 09:22 Observation | Compute companies enter a state of frenzy: Anthropic borrows 60 billion and pays 10% interest, CoreWeave's probability of default exceeds 60% Sikai Has Something to Say Quick Overview

This article is from the WeChat public account: Sikai Suisui Kan, author: Sikai, original title: "Observation | Compute companies enter a state of frenzy: Anthropic borrows 60 billion and pays 10% interest, CoreWeave's probability of default exceeds 60%"

In mid-October, three things happened at the same time, piecing together the most real side of the compute arms race.

Anthropic offered Blackstone a coupon above 10% to borrow $60 billion to buy Google TPUs. CoreWeave's bond yields soared above 15%, and the credit default swap market priced its five-year default probability above 60%. Domestically, Hongxin Electronics announced 2.983 billion yuan in compute chip purchases, equivalent to 88% of its total first-half revenue.

This is not "investing in the future." This is borrowing money from the future to gamble with today's life.

Anthropic: a company profitable on an "adjusted basis," so why pay 10% interest?

Anthropic's pricing on the debt side provides the starkest yardstick.

This $60 billion financing is called "Project Polaris" and is used to purchase about 1.5 gigawatts of Google TPUs. A Blackstone fund will invest $9 billion to buy debt backed by Anthropic's chip lease payments over the next five years, with a 10% coupon and an even higher total yield after discounts and fees.

A key comparison: according to the ICE BofA BB US High Yield Index, borrowers with similar credit profiles typically pay about 7% when raising funds in the public market. The extra 3 percentage points Anthropic pays is a vote of no confidence by creditors in "AI chip lease cash flow."

And Anthropic is not without revenue. It recently disclosed third-quarter revenue of $16.8 billion and said it had achieved adjusted profitability for two consecutive quarters. But there is an easily overlooked difference in basis: "adjusted profitability" is a non-GAAP metric, while the company's full-year 2025 net loss was $42 billion. Creditors price accounting earnings and cash flow, not an adjusted story.

Even if it is "profitable" on an adjusted basis, it still has to pay 10% interest. The reason is that its future commitments far exceed its current cash flow: it plans to invest $518 billion over the next several years in cloud computing and infrastructure. If the money on hand is not enough, it has to borrow. The cost of borrowing is set by creditors—and the price creditors give is: you are riskier than a BB-rated company.

CoreWeave: the bonds of Nvidia's "own child" have fallen into "distressed" territory

If Anthropic is "actively borrowing expensive money," then CoreWeave is being "priced as distressed" by the market.

This company, whose core business is renting Nvidia chip compute, issued $2.75 billion in bonds in April with a 9.75% coupon; on October 15 they traded intraday at 81 cents on the dollar, with yields breaking above 15%. Both unsecured bonds maturing in 2031 and 2032 fell below 80 cents—a level typically seen in the market as the threshold for a borrower in financial distress. Its five-year CDS spread widened at one point to 1,194.5 basis points, and the market-priced default probability exceeds 60%.

Musk reposted the related report on a social platform, leaving only one word: "yikes."

CoreWeave's business model is the ultimate sample of this frenzied logic. It uses Nvidia GPUs themselves as collateral in exchange for bank loans, then uses the borrowed money to buy more GPUs, and then rents out the compute. To date, it has signed seven similar GPU-collateralized loans, raising about $29.3 billion in total. Wall Street analysts estimate total debt exceeds $40 billion (about $35 billion under different accounting bases), most of which needs to mature gradually over the next five years.

The premise for this model to work is: Nvidia chip prices do not plummet, and customers keep placing orders. Once any assumption collapses, the entire chain turns from a virtuous cycle into a debt noose.

Nvidia's binding with CoreWeave makes this even more complicated. Nvidia holds about 10% of CoreWeave's shares, the two sides have signed a cloud computing service order initially worth $6.3 billion, and it is agreed that when CoreWeave's computing power is not fully utilized, Nvidia is obligated to purchase the remaining unsold capacity, an arrangement that extends to April 2032. In other words, whatever computing power CoreWeave cannot sell, Nvidia buys under contract. Jensen Huang publicly endorsed CoreWeave, emphasizing its "improved financial strength," but deliberately drew a line between "expressing confidence" and "financial backstop."

"Confidence" can be expressed verbally, but a "backstop" is a contract clause. When a chip company's revenue depends on customers' purchases, and customers' ability to repay depends on collateral and contractual commitments provided by the chip company, the fragility of the entire chain is pushed to its limit.

This is not an isolated case: from Volta to xAI, high-interest debt is becoming the "new normal" in the computing power industry

Volta, another AI cloud service provider, completed a $5 billion leveraged loan issuance on October 15 through JPMorgan and Goldman Sachs, priced 6.5 percentage points above the benchmark rate, at 95 cents per $1 of face value, for a yield of more than 11%. More notably, the two banks only underwrote $1.3 billion of the term loan, choosing not to underwrite the remaining $3.7 billion of chip procurement loans, leaving the risk outside the door.

Banks are "scaling back." When underwriters themselves are unwilling to keep all the risk on their balance sheets, risk pricing in this market is already undergoing a qualitative change.

xAI's financing structure is equally aggressive. Of its $20 billion financing in January 2026, about $12.5 billion is debt financing, channeled through special purpose vehicles (SPVs) specifically for purchasing Nvidia GPUs. The debt Apollo provided for the related chip vehicles is expected to carry an interest rate of 9.5%.

The pricing logic of the dollar bond market is reproduced in A-shares in another way.

Hongxin Electronics announced 2.983 billion yuan in computing power chip procurement, equivalent to 88.36% of its total revenue in the first half of the year. The company's original main business was FPC flexible circuit boards, with gross margins once as low as 2.39%, and consecutive losses of 268 million, 308 million, and 436 million yuan from 2021 to 2023. After entering the computing power track in 2023, its computing power business revenue grew from 1.988 billion to 2.788 billion yuan, and its share of total revenue rose from 33.84% to 38.12%. A manufacturing company that was originally thin-margin and loss-making for years is using money close to its full-year revenue to stockpile chips, betting that the spread from computing power leasing can cover funding costs and depreciation. The logic of this bet is the same as CoreWeave's, but its risk resilience is weaker.

The root of the frenzy: circular financing turns "real demand" into an unverifiable question

Behind these cases is a common financing mechanism: circular financing.

Nvidia invests in CoreWeave, CoreWeave uses that money and GPU-collateralized loans to buy Nvidia chips, and then leases the computing power to customers such as Meta. Nvidia's relationship with OpenAI is even more aggressive—according to media reports, the two sides are discussing providing about $250 billion in financing guarantees for OpenAI's data centers, while also discussing up to $350 billion in additional financing for OpenAI's chip procurement.

When the supplier is simultaneously the investor, the guarantor, and the source of revenue, the boundary between "end demand" and "demand subsidized by the supplier" disappears. You think someone is buying computing power, but in reality it may be the shovel seller lending money to the shovel buyer so he can buy his shovels.

The market has already begun pricing: from "cheap money" to "every dollar scrutinized"

Wells Fargo's global head of fixed income strategy gave the clearest summary of this turning point: "The kind of 'cheap money' era we experienced in the first nine months of this year is beginning to disappear. Now, every dollar or marginal dollar will be scrutinized more strictly, and investors will carefully assess what kind of return conditions the issuer can offer."

Goldman Sachs warns that tighter credit and rising capital costs could delay computing power infrastructure construction, and the semiconductor supply chain may face an "order cliff" in 2027 to 2028—demand cannot be converted into actual purchases.

A D.A. Davidson analyst laid out CoreWeave's math very bluntly: "If CoreWeave borrows at a cost of 9% and the best return it can get is only 5%, then no matter how large it scales, it makes no sense: they are destroying value at every step."

Lian Ping, chairman of the China Chief Economist Forum, offered a more macro warning: "The rise in debt scale and leverage will increase the financial fragility of AI companies and raise the overall risk exposure of the investment-grade bond market" and "the interest burden on highly indebted AI companies will continue to increase, squeezing free cash flow and amplifying credit risk."

AI's technological progress is real, and computing power demand is real. But "demand is real" and "meeting demand with 10% to 15% debt" are two completely different things.

Anthropic pays 10% interest because it believes future revenue can cover current costs. CoreWeave uses GPU-collateralized loans because it believes chip prices will not plummet and customers will not walk away. Hongxin Electronics uses 88% of half-year revenue to stockpile chips because it believes the spread from computing power leasing can beat funding costs.

Each assumption looks reasonable on its own. But strung together, they form a chain of layer upon layer of leverage, dependence, and assumptions. If the assumption at any link is disproven, the pricing of the entire chain will have to be redone.

When creditors start demanding 10%, when the CDS market prices a 60% default probability, when the underwriting banks themselves are "scaling back," the market is already saying: at this price, I don't accept it.

The peak of frenzy is never when everyone has gone mad. It is when the calmest money begins to withdraw.

AI's technological progress is not false, and the real demand for computing power is not false either. But real demand does not mean one can overdraft the future with 10%~15% high-interest debt. The entire computing power leverage chain is built on a series of fragile assumptions: chips will not plummet, customers will not leave, and funding will remain loose. As long as any link breaks, valuations will have to be rewritten. The peak of frenzy is never when everyone has gone mad. It is when the calmest money begins to withdraw. Channel: Finance and Economics. This content is published with the author's authorization, and the views represent only the author himself, not Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]