Debt up 22x in three and a half years: what the 'CoreWeave shock' reveals about how AI's compute factories get built

The phrase in the Japanese headlines that day was "the CoreWeave shock".

CoreWeave is a US company that rents GPU computing power to AI developers such as OpenAI and Microsoft. In effect, it builds AI "compute factories" and leases them out.

And it has built those factories with debt.

Debt outstanding went from $1.6 billion at the end of 2023 to $35.6 billion at the end of June 2026. Then, on October 14, yields on the company's bonds jumped to nearly 15%, and its stock fell 15% in a day.

Bond yields climb when lenders start to worry they may not get their money back. What happened to a company at the very center of the AI boom?

The fuse was lit a month earlier.

On September 17, CoreWeave announced a convertible bond (debt that can later be swapped for shares) and a program to sell up to 35 million new shares into the market over time. The convertible was upsized from $3.0 billion to $3.7 billion, with an option for up to $500 million more. It would keep raising money for its factories with both debt and new shares. The stock fell about 4% that day.

On October 13 (US time), a chart posted on X by a prominent investor went viral. It showed the yield on CoreWeave's $1.25 billion of 9.625% notes due 2032 briefly touching 15%. Elon Musk was reported to have replied "Yikes."

On October 14, the bad news stacked up.

• US inflation: September consumer prices rose 3.7% from a year earlier, above expectations. The Federal Reserve had just raised rates in September, and bets on further hikes grew
• Gartner's forecast: the research firm said the same day that by 2030, more than 80% of neocloud providers will cease independent operations
• Bond prices: according to Bloomberg, the 2032 notes traded below 80 cents on the dollar

CoreWeave's stock closed down 15.46% at $65.78, more than 60% below its record close of $183.58. Fellow neocloud Nebius fell 10%, Oracle, which is racing to add AI data centers, fell nearly 9%, and the semiconductor index dropped 4.3%.

In an internal memo, CEO Michael Intrator reportedly wrote that markets sometimes misunderstand and act irrationally, arguing that the backlog more than covers the debt.

Then came Tokyo on the 15th. The Nikkei closed down 3,318.34 yen (4.90%) at 64,425.95. According to Reuters, chip-testing equipment maker Advantest fell 15%, SoftBank Group 9.3% and Tokyo Electron 7.8%. Worries about US inflation and rate hikes collided with doubts about AI spending. Japan's leading business daily, the Nikkei, called the slide the "CoreWeave shock."

How did CoreWeave manage to borrow so much? The mechanism is surprisingly simple.

First, it signs multi-year contracts with large customers. Next, it sets up a subsidiary to hold the GPUs for those contracts, and borrows from banks and credit funds. The collateral is the subsidiary's GPUs and the customer contracts themselves. According to its quarterly report, the amount it can borrow is capped at a percentage of what the GPUs cost, and repayment comes from fees paid "generally from investment grade counterparties."

The contract terms are what make it work. 98% of CoreWeave's revenue comes from committed take-or-pay contracts: customers pay whether or not they use the capacity. From a lender's point of view, the source of repayment is locked in up front.

So who has made those promises?

Sources: CoreWeave 10-K (2025) and disclosures of each agreement. Amounts are contract totals or caps.

That last row is the other defining feature. All of CoreWeave's GPUs are NVIDIA's. NVIDIA bought $2 billion of CoreWeave shares in January 2026 and has agreed to buy up to $6.3 billion of unsold capacity through April 2032.

The company selling the GPUs has invested in its buyer and promised to take any unsold capacity. The money goes around in a loop. In September, the Bank of England's Financial Policy Committee flagged "circular arrangements" and opacity in how AI is being financed.

With the mechanism in view, look at the cash.

From January to June 2026, the business brought in $3.7 billion of cash. Debt repayments and interest took $6.2 billion, and new equipment another $14.1 billion. What it paid on its debt alone exceeded what it earned. New borrowing and new shares fill the gap.

In April–June alone, revenue more than doubled to $2.58 billion, but interest came to $640 million and the company posted an operating loss of $49 million. It has guided to interest expense of $860–940 million in July–September. Scheduled principal repayments total $4.4 billion for the rest of 2026 and $6.2 billion in 2027.

Markets have two main worries.

1. Is it borrowing at more than it can earn? Gil Luria of brokerage D.A. Davidson said on Bloomberg TV that borrowing at 9% to earn a 5% return destroys value.

2. Will the GPUs keep earning until the debt is repaid? CoreWeave depreciates its GPUs over six years. Gartner thinks spending on AI compute will grow too slowly for most providers to recover their investment before the current generation of chips is replaced.

There are clear supports, too.

• Backlog: about $104 billion of contracts not yet booked as revenue, according to its earnings release, about three times its debt. It says it signed more than $25 billion of new commitments early in the third quarter
• Demand for compute: one-year rental contracts for NVIDIA H100s rose from $1.70 per GPU-hour in October 2025 to $2.35 in March 2026. Spare capacity is scarce, so demand is still strong
• Falling borrowing costs: the company's average interest rate fell from 14.9% in 2023 to 8.3% at the end of June 2026

So the problem isn't a lack of demand. It's which runs out first: the customers' ability to keep their promises, or the company's ability to carry the interest. That's the race the market is pricing.

Among companies building AI compute factories with debt, CoreWeave is the leanest and the most leveraged. But the same pattern reaches much bigger companies.

Start with spending. Capital expenditure guidance for 2026: Alphabet $195–205 billion, Amazon about $220 billion, Microsoft about $190 billion and Meta $130–145 billion. Simply adding up the four tops $700 billion (fiscal years and definitions differ slightly by company).

Gartner estimates AI providers will spend nearly $1 trillion on AI infrastructure in 2026 and another $6 trillion in 2027–2030. Recouping that would take $8.5 trillion of cumulative revenue by 2030, it calculates, while revenue attributable to AI data centers will reach no more than $3.6 trillion.

Cash on hand can't cover sums like that.

• Meta is building a giant Louisiana data center through a joint venture with investment firm Blue Owl, which issued about $27 billion of bonds. Issued with an A+ rating, those bonds had slipped to around 91 cents on the dollar by late September
• Oracle's borrowings rose from $92.6 billion in May 2025 to $129.5 billion in May 2026. In July, S&P cut Oracle to BBB-, the lowest investment-grade rating

And overall? According to the Bank of England committee's September record, Morgan Stanley estimates about $450 billion of AI-related debt had been issued by early September, more than double all of 2025. The same record cites JPMorgan's estimate that about $4.1 trillion of AI capital spending in 2026–2030 will be debt-financed. The Bank for International Settlements has noted that AI firms are shifting from cash flow to debt.

On October 13, the IMF counted 14 major deals since 2023: about $300 billion of equity and $650 billion of debt raised in exchange for $1.4 trillion in future revenue commitments. AI's compute factories are being built with money raised today against promises of tomorrow. CoreWeave's structure is a miniature of that.

Money is flowing in from Japan, too

Japan isn't outside this.

On October 1, SoftBank Group paid in the final $10 billion of a follow-on investment, bringing its total investment in OpenAI to $64.6 billion, for about 13%. It funded that last payment with foreign-currency bonds. In September it also sold about $11.1 billion of bonds, the largest high-yield corporate bond sale on record, and in Japan it sold ¥1 trillion of bonds to individual investors (seven years, 4.75%).

Banks are involved as well. CoreWeave's $8.5 billion loan in March 2026 was arranged by Morgan Stanley and MUFG. And the sharp falls in Advantest and Tokyo Electron on the 15th reflect how much of their sales depend on exactly this construction boom.

Build capacity on debt, run out of money before demand catches up: there's a well-known precedent. The telecom bubble of the late 1990s.

Betting on the internet, US telecom companies poured more than $500 billion into fiber and networks in the five years from 1996. At a 2002 Senate hearing, the chairman of the Federal Communications Commission put telecom debt at about $1 trillion worldwide, much of which would never be repaid.

It ended badly. US long-distance capacity grew about 500-fold from end-1997 to end-2001, while actual demand grew about fourfold. In 2002, 56.4% of the world's corporate bond defaults were telecom, and WorldCom and Global Crossing went bankrupt.

Side by side with AI today:

Sources: Brookings (2002), OECD (2003), Kansas City Fed (October 2026) and company filings; compiled by DATA WORLD.

What's similar: capacity is being built ahead on debt, and money circulates between sellers and buyers.

What's different, first, is who pays. In an October analysis, the Kansas City Fed found debt-to-asset ratios of 44% for hyperscalers versus 85% for neoclouds. The giants can fund investment from earnings; the debt-financed part is concentrated in neoclouds and special financing vehicles. There are signs of demand, too. Microsoft's commercial backlog rose 84% to $678 billion by the end of June 2026 (up 25% excluding OpenAI).

The biggest difference is asset life. The fiber left over from the telecom bubble was eventually put to use as the internet grew, because it lasts for decades. GPUs don't. In April, the IMF warned that GPUs could face obsolescence within about two years. Surplus fiber found its use later; surplus GPUs won't wait. The time for demand to catch up is much shorter than it was for telecom.

EDITOR'S TAKE ── Editor-in-Chief Ogasawara's reading

Interest rates now help set the price of AI

Reading through CoreWeave's filings, what came into view is that AI compute factories rest not only on technology but on promises about money.

GPUs are expensive, and data centers take years to build. Yet AI companies want compute now. Debt bridges that gap: borrow today against future usage fees, and install the GPUs today.

Because it's borrowed, rising interest rates feed into the price of compute. The roughly 15% yield the market demanded on October 14 is close to double the 8.3% average rate CoreWeave pays. The price of AI is now being set not only at power plants but by interest rates in financial markets.

CoreWeave's strongest support is that 98% of its revenue is take-or-pay. But a promise is only as strong as the one who makes it. Microsoft and Meta earn plenty from their core businesses; customers like OpenAI are themselves buying compute with outside money.

Lenders, compute factories, AI companies and investors are all linked by each other's promises. If one link jams, the whole chain gets questioned. This week's sell-off looked like a rehearsal for that.

The Japanese stocks that fell hardest were those supplying parts and equipment to AI compute factories, like Advantest and Tokyo Electron. SoftBank Group is channeling bond money into OpenAI, and Japanese banks are arranging loans.

Three numbers are worth watching from here: the yields on CoreWeave's and Oracle's bonds, OpenAI's fundraising, and the price of renting GPUs. The real test will come if all three worsen at once. Conversely, as long as there's no shortage of companies wanting to rent GPUs, AI's compute factories will keep going up.