The “big bears” target the $573 billion AI financing network: The next question we need to ask is, what will happen if AI spending stops? Who will take over these debts?

The Zhitong Finance App learned that Michael Burry, a well-known Wall Street investor and prototype of the movie “The Big Short,” was originally preparing the next installment of his “AI Star Heretic Guide” series. The theme was “Where did the money actually come from” for this round of AI construction. He then read Ares Management's Fall 2026 Alternative Credit Newsletter and immediately posted an unplanned “short thought” post on Sunday night (September 27). He called the findings of the study “information we need to know now” and “intelligence we can't wait,” and called the report “mainstream Wall Street is shifting to my point of view.”

The Ares report, named by Burry, is a map of the AI financing network: a total of 26 disclosed financings over the past 12 months, amounting to approximately US$573 billion. Ares's conclusion is a caveat — every link on this network assumes “AI spending continues to grow.”

A $573 billion network: eight companies, five identities

According to Ares's matrix, $573 billion spans five types of instruments: corporate bonds, data center construction loans, leases, GPU collateral loans, and financial guarantees. The names on this website include almost all the protagonists in this round of AI construction: Meta, Oracle, Microsoft, Amazon, Google, Nvidia, OpenAI, and Anthropic.

What's really remarkable is that they all have different identities within the same network — borrowers, customers, chip vendors, tenants, or guarantors. The same company can be a lessor in A transaction and a guarantor in B This means that any change in the spending pace of any company will be transmitted to other companies linked to it. Ares worded this: “every transaction, every obligation, every guarantee and agreement” in this network, all provided that AI spending continues to grow.

Ares further described a transmission path: if revenue falls short of expectations, as soon as several companies' boards of directors start moving capital elsewhere, the pressure spreads along interconnected transactions — and guarantees are “triggered when the guarantors are at their weakest.” It likens this potential ripple effect to “vendor financing” (vendor financing), which amplified the telecom industry's decline in 2000.

Next question: Who holds these debts

Burry took the problem one level further — who would end up with these debts. The answer he gave was the insurance industry: the combined floating funds of US and Bermuda insurance companies were slightly over 10 trillion US dollars, and he said he had collected several months of insurance data for this, which will be included in the sixth part of the series. Ares uses a different caliber: US insurance companies have $9 trillion in investment assets.

Ares clearly stated that insurance balance sheets are one of the main sources of funding for this round of digital infrastructure construction. The example it gives is quite illustrative: Meta and BlackRock are located at the $14 billion data center campus in El Paso, Texas, and their fire insurance coverage is capped at $450 million, which is only 3% of the project value, according to Ares. Ares compares this figure — insurance companies have limited interest in the park's physical assets, but are quite active in holding investment-grade bonds linked to such projects.

This comparison is linked to another set of figures: digital infrastructure accounts for 42% of investment-grade private placement since this year, and at least 24 data center securitization products have been issued in the past 12 months.

The insurance structure previously disclosed by the British media is more detailed than Ares's summary: the upper limit of the one-gigawatt park's full-insurance property insurance was about US$427 million during the construction period, raised to about US$450 million after operation, the upper limit of terrorism insurance was US$645 million, rent relief insurance due to construction delays was up to US$218 million, and the commercial general liability insurance was capped at $50 million per accident — the project was not covered by “total loss.” For these policies, the project pays about $5 million a year, and premiums increase by 2% every year. S&P gave the project debt an A+ rating, which is one level lower than Meta's own corporate rating. One reason is that creditors cannot directly claim rights over the park's physical assets; if a serious accident causes the project to be extended for more than 18 months, Meta can also terminate the lease without penalty.

In other words, when the asset value of a single item is too high for the insurance market, the risk will shift from the policy to the terms of leases, guarantees, and contracts — and the effectiveness of these terms depends on the lessee's own credit.

Chip depreciation is “very, very real”

Burry turned his gaze back to the chip itself behind some of the AI loans. He quoted a chart from Ares: The residual value of Nvidia's H100 fell 51% in three years. Chip values are changing too fast, making hardware-backed loans complicated — this is why lenders require a residual value guarantee. According to Burry, the reason manufacturers and cloud service providers are providing underwriting is “not because they want to do it, but because they have to.” He called chip depreciation “very, very real.”

This isn't his first time talking about depreciation. Since November 2025, Burry has advocated that hyperscale manufacturers extend the lifespan of Nvidia chips to a far surreal 2-3 year replacement cycle, believing that this may reduce the entire industry's depreciation by about $176 billion in 2026-2028; according to his estimates, Oracle's profit by 2028 may be overestimated by about 27% and Meta by 21%. Nvidia's counterattack was a seven-page memo sent to Wall Street analysts, arguing that 4-6 years is a more realistic time frame, and incidentally corrects a figure quoted by Burry: the company's repurchases since 2018 were $91 billion, not $112.5 billion.

There is also counterevidence on the hardware side. The CEO of CoreWeave said that the H100 production capacity released from the expired contract was re-signed at a level close to 95% of the original contract price; Nvidia CEO Wong In-hoon mentioned that H100 rent had risen 22% to $3.28 per hour in a month, and a CoreWeave contract to rent the A100 to 2029 — nine years after the chip was first released.

The two sides don't actually look at the same indicators: the residual value answers “how much is this chip worth today,” and the rent answers “how much money can it still make today.” Chips can depreciate drastically but still bring considerable rent, so no single figure is sufficient to determine how long the depreciation period should be.

Market Week gave a footnote

Burry's post didn't appear in a vacuum. Around September 24, Oracle issued a force majeure notice to Stack Infrastructure, a developer under Blue Owl Capital: if its 2.45 gigawatt data center campus in New Mexico cannot be put into operation as scheduled in 2028, Oracle hopes to delay part of the payment. The trigger point was the supporting gas pipeline — the New Mexico Land Office has twice vetoed the pipeline plan, and the pipeline's commissioning date was delayed by nearly six months until February 1, 2027. Oracle emphasized that “Project Jupiter is still on our planning timeline,” and Blue Owl declined to comment. The $18 billion construction loans provided by about 20 banks have fallen to a face value between 89-91 US; Oracle and Blue Owl shares each fell about 4% on the same day, while Bloom Energy fell 6%. It needs to be clarified here: a discount transaction does not equal a default — Oracle did not miss any payments, and the project was not cancelled. In the words of people familiar with the matter, this was a “risk allocation operation.”

Put it into the bigger picture, the pressure is more clear. According to Morgan Stanley's data, the overall leverage ratio of hyperscale manufacturers has risen from 0.9 times in the third quarter of 2025 to 1.8 times, doubling in more than two quarters, and surpassing the level of the entire energy industry; Amazon, Meta, Google, Microsoft, and Oracle already account for 4% of the US dollar investment-grade bond index. If weighted over time, the six largest hyperscale manufacturers plus three major chip companies already account for 9%.

On the financing cost side, the yield on 10-year US bonds is close to 5.17%, up about 1 percentage point since the beginning of the year. AI companies returning to the bond market must offer more attractive dividends; J.P. Morgan expects AI-related companies to issue a total of 4.1 trillion US dollars in debt by 2030. On the stock side, the AI computing power chain for A-shares and Hong Kong stocks underwent deep adjustments on September 28 during the same period, leading the decline in the optical module, semiconductor, and copper-clad plate sectors.

Burry's road map, with “the person opposite”

According to Burry's own roadmap, the sixth part of the series will be released within a few weeks. The content is about this financing network plus his insurance research; the seventh part will examine “new cloud” vendors such as CoreWeave and Nebius. His established positions include short positions on Oracle (approximately $145), Nebius ($211.77), Micron and Palantir, and shorting the Philadelphia Semiconductor Index through a put option due in January 2027.

It's not quiet across the street. Nvidia's seven-page memo forms the most direct rebuttal to Micron's chief commercial officer's statement — the latter tells investors that demand for memory chips will exceed the company's supply capacity until at least 2028, directly hedging the judgment that “current AI demand is being exaggerated.” CoreWeave renewed the older-generation chip contract at close to full price, providing the hardest counterexample to the “collapse of residual value theory.” At the same time, the bearish camp is expanding: GMO co-founder Jeremy Grantham says the current AI valuation is a bubble, and DoubleLine's Jeffrey Gundlach expects the race to run out as a clear loser. The observation indicator proposed by Jonathan Weil, a former analyst at Kynikos, is more like a reminder: the real word to focus on is “funding gap” (funding gap) — the last time it took the financial world dictionary by storm along with “money burning rate” was in 2000.

Burry concluded his post with one sentence: “The bubble is really painfully obvious. Pain.”