Renowned Wall Street investor Michael Burry, the real-life figure behind the film "The Big Short," was originally preparing the next installment of his "AI star heresy guide" series, focused on where the money for this AI buildout actually comes from. Then he read Ares Management's autumn 2026 alternative credit newsletter and promptly published an unscheduled "brief thoughts" post on Sunday night, September 27. He called the study's conclusions "information you need to know right now" and "intelligence you cannot afford to wait for," describing the report as evidence that "mainstream Wall Street is moving toward my view."

The Ares report Burry singled out draws a map of an AI financing network: 26 disclosed financings over the past 12 months, totaling roughly $573 billion. Ares's conclusion is a warning: every link in this web is premised on the assumption that AI spending keeps growing.

A $573 Billion Web: Eight Companies, Five Roles

According to Ares's matrix, the $573 billion spans five categories of instruments: corporate bonds, data center construction loans, leases, GPU-backed borrowing, and financial guarantees. The names caught in this web include nearly every protagonist of this AI buildout: Meta, Oracle, Microsoft, Amazon, Google, Nvidia, OpenAI, and Anthropic. What truly stands out is that their roles within the same web differ—borrower, customer, chip supplier, tenant, or guarantor. The same company can be a lessor in one deal and a guarantor in another. This means any change in one company's spending pace will transmit to the other companies connected to it.

Ares's wording is that "every transaction, every obligation, every backstop, every guarantee and agreement" in this web is premised on AI spending continuing to grow. Ares further describes a transmission path: if revenue falls short of expectations, and just a few companies' boards begin moving capital elsewhere, pressure will spread along the interconnected transactions—guarantees are precisely what get "triggered when the guarantor is at its weakest." It compares this potential chain reaction to the "vendor financing" that amplified the telecom industry's downturn in 2000.

The Next Question: Who Holds This Debt

Burry pushes the question one layer deeper—who ultimately ends up holding this debt. His answer is the insurance industry: U.S. and Bermuda insurers together hold float slightly exceeding $10 trillion, and he said he has been collecting insurance data for months and will write it up in the sixth part of his series. Ares uses a different measure: U.S. insurers' investment assets stand at $9 trillion. Ares explicitly writes that insurance balance sheets are one of the main funding sources for this digital infrastructure buildout.

Its example is telling: for Meta and BlackRock's $14 billion data center campus in El Paso, Texas, Ares says fire insurance coverage is capped at $450 million—just 3% of the project's value. Ares uses this figure as a comparison—insurers show limited interest in the campus's physical assets but are quite eager to hold investment-grade bonds tied to projects like this. That contrast connects to another set of numbers: digital infrastructure accounted for 42% of investment-grade private placements this year, and at least 24 data center securitizations have been issued in the past 12 months.

British media previously disclosed insurance structures more granular than Ares's summary: the 1-gigawatt campus's all-risk property insurance is capped at about $427 million during construction and rises to about $450 million once operational, terrorism insurance is capped at $645 million, rent-abatement insurance for construction delays goes up to $218 million, and commercial general liability is capped at $50 million per occurrence—the project is not insured for "total loss." For these policies, the project pays about $5 million a year, with premiums rising 2% annually. S&P assigns the project debt an A+ rating, one notch below Meta's own corporate rating, partly because creditors cannot directly assert claims on the campus's physical assets; if a serious accident delays the project by more than 18 months, Meta can also terminate the lease without penalty.

In other words, when a single project's asset value becomes too large for the insurance market to absorb, risk shifts from insurance policies to leases, guarantees, and contract terms—and the effectiveness of those terms depends on the tenant's own credit.

Chip Depreciation Is "Very, Very Real"

Burry then turns his gaze back to the chips themselves behind some AI loans. He cites an Ares chart: the residual value of Nvidia's H100 falls 51% over three years. Chip values change so quickly that it complicates hardware-backed lending—which is exactly why lenders demand residual-value guarantees. According to Burry, manufacturers and cloud service providers offer backstops "not because they want to, but because they have to." He calls chip depreciation "very, very real."

This is not the first time he has discussed depreciation. Since November 2025, Burry has argued that hyperscalers are stretching Nvidia chips' useful lives far beyond a realistic 2-3 year replacement cycle, which he says could cause the industry to understate roughly $176 billion in depreciation in 2026-2028; by his estimate, Oracle's earnings could be overstated by about 27% by 2028 and Meta's by about 21%. Nvidia's counter is a seven-page memo to Wall Street analysts arguing that 4-6 years is a more realistic lifespan, and it also corrects a figure Burry cited: the company's buybacks since 2018 total $91 billion, not $112.5 billion.

There is counter-evidence on the hardware side too. CoreWeave's CEO said H100 capacity released from expiring contracts is being re-signed at close to 95% of original contract prices; Nvidia CEO Jensen Huang mentioned that H100 rental rates once rose 22% in a month to $3.28 per hour, and a CoreWeave contract that rents A100s until 2029—nine years after the chip's initial launch.

The two sides are actually looking at different metrics: residual value answers "what is this chip worth today," while rental rate answers "how much can it still earn today." A chip can depreciate sharply yet still generate substantial rent, so neither figure alone is enough to determine how long the depreciation period should be.

The Market Gave a Footnote Last Week

Burry's post did not appear in a vacuum. Around September 24, Oracle sent 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 begin operations on schedule in 2028, Oracle wants to defer some payments. The trigger was the supporting natural gas pipeline—the New Mexico State Land Office has twice rejected the pipeline plan, pushing the pipeline's in-service date back nearly six months to February 1, 2027. Oracle stressed that "Project Jupiter remains on our planned timeline," while Blue Owl declined to comment.

The $18 billion construction loan provided by about 20 banks has fallen to 89-91 cents on the dollar; on the day of the news, Oracle and Blue Owl shares each fell about 4%, and Bloom Energy dropped 6%. A distinction is needed here: trading at a discount is not the same as default—Oracle has not missed any payments, the project has not been canceled, and in the words of people familiar with the matter, this is a "risk-allocation move."

Put it in a broader picture, and the pressure becomes clearer. Morgan Stanley data shows hyperscalers' overall leverage has risen from 0.9 times in Q3 2025 to 1.8 times, doubling in just over two quarters and exceeding the level of the entire energy industry; Amazon, Meta, Google, Microsoft, and Oracle already account for 4% of the dollar investment-grade bond index, and on a duration-weighted basis, the six major hyperscalers plus three major chip companies already account for 9%.

On the funding-cost side, the 10-year U.S. Treasury yield is near 5.17%, up about 1 percentage point since the start of the year, meaning AI companies returning to the bond market must offer more attractive coupons; JPMorgan expects AI-related companies to issue a cumulative $4.1 trillion in debt by 2030. On the equity side, A-share and Hong Kong AI computing power chains saw deep corrections on September 28, with optical modules, semiconductors, and copper-clad laminate sectors leading the declines.

Burry's Roadmap, and "The People on the Other Side"

According to Burry's own roadmap, the sixth part of the series will be published within weeks, covering this financing web plus his insurance research; the seventh part will examine "neocloud" players such as CoreWeave and Nebius. Positions he has already established include shorts on Oracle (around $145), Nebius ($211.77), Micron, and Palantir, as well as a short on the Philadelphia Semiconductor Index through put options expiring in January 2027.

The other side is not quiet. Nvidia's seven-page memo and Micron's chief commercial officer's remarks provide the most direct rebuttal—the latter told investors that memory chip demand will exceed the company's supply capacity at least through 2028, directly offsetting the judgment that "current AI demand is exaggerated." CoreWeave re-signing old-generation chip contracts at close to full price offers the hardest counterexample to the "residual value collapse" thesis.

Meanwhile, the bearish camp is expanding: GMO co-founder Jeremy Grantham calls current AI valuations a bubble, and DoubleLine's Jeffrey Gundlach expects this race to produce clear losers. And a metric raised by former Kynikos analyst Jonathan Weil feels more like a reminder: the word to really watch is "funding gap"—the last time it swept through the financial lexicon alongside "burn rate" was in 2000.

Burry ended his post with one line: "The bubble is painfully obvious. Painful."