Michael Burry reveals "urgent messages" from a $573 billion AI‑funding project: What would happen if spending were to stop?
Zhitong Finance APP learned that Michael Burry, a well-known Wall Street investor and the real-life inspiration behind the film "The Big Short," had been preparing the next installment in his "AI Star Heresy Guide" series, focusing on where the money is actually coming from in this round of AI development. Then he came across Ares Management's Fall 2026 Alternative Credit Newsletter and, on Sunday evening (September 27), unexpectedly published a brief "thought piece." He described the study's findings as information "you need to know now" and "intelligence you can't afford to wait for," adding that the report signaled "mainstream Wall Street is shifting toward my view."
The Ares report singled out by Burry depicts an AI‑funding network: over the past 12 months, there were 26 disclosed financings totaling roughly $573 billion. Ares concludes with a cautionary note: every node in this network rests on the assumption that AI spending will continue to grow.
A $573 billion web: eight companies, five identities
According to this matrix from Ares, $573 billion spans five categories of instruments: corporate bonds, data-center construction loans, leases, GPU‑backed borrowings, and financial guarantees. The names listed on this network virtually encompass all the key players in this round of AI infrastructure development: Meta, Oracle, Microsoft, Amazon, Google, NVIDIA, OpenAI, and Anthropic.
What is truly noteworthy is that the roles these entities assume within the same network vary—borrower, client, chip supplier, tenant, or guarantor. A single company may serve as a lessor in one transaction and a guarantor in another. This means that any change in a company's spending pattern will ripple through the other firms connected to it. As Ares puts it: "Every transaction, every obligation, every backstop, every guarantee, and every agreement" in this network rests on the premise of sustained growth in AI‑related expenditures.
Ares further described a transmission pathway: if revenues fall short of expectations, as soon as the Directors of several companies begin reallocating capital elsewhere, the pressure will spread through interconnected transactions—collateral, in particular, will "be triggered precisely when the guarantor is at its weakest." He likened this potential chain reaction to the "vendor financing" that amplified the telecom sector's downturn in 2000.
Next question: Who holds these bonds?
Burry takes the issue one step further—namely, who ultimately ends up holding these debts. His answer is the insurance industry: U.S. and Bermuda‑based insurers together have aggregate float exceeding $10 trillion, and he notes that he has been gathering insurance‑industry data for months, which he plans to include in the sixth installment of his series. Ares offers a different figure: U.S. insurers' investment assets total $9 trillion.
Ares explicitly states that insurance balance sheets are one of the primary sources of funding for this round of digital infrastructure development. Its example is telling: Meta and BlackRock's $14 billion data center campus in El Paso, Texas, has, according to Ares, a fire‑insurance coverage limit of just $450 million—barely 3% of the project's value. Ares uses this figure as a point of comparison: insurers show limited interest in the physical assets of the campus, yet they are highly active in holding investment‑grade bonds linked to such projects.
This comparison is further linked to another set of figures: digital infrastructure accounts for 42% of investment-grade private placements issued so far this year, and at least 24 data-center‑related securitization transactions have been launched over the past 12 months.
British media previously disclosed an insurance structure more detailed than Ares's summary: the 1‑gigawatt campus is covered by a comprehensive property policy with a construction‑phase limit of roughly $427 million, rising to about $450 million once operational; terrorism coverage tops out at $645 million; rent‑abatement insurance for construction delays has a maximum of $218 million; and commercial general liability provides up to $50 million per occurrence—though the project is not insured against total loss. For these policies, the project pays approximately $5 million annually, with premiums increasing by 2% each year. S&P has assigned the project's debt an A+ rating, one notch below Meta's own corporate rating, citing, among other factors, that creditors cannot directly assert claims against the campus's physical assets. Moreover, should a severe incident cause delays exceeding 18 months, Meta could terminate the lease without penalty.
In other words, when the asset value of a single project becomes too large for the insurance market to absorb, risk shifts from the insurance policy to the lease, guarantees, and contractual provisions—whose enforceability, in turn, depends on the lessee's creditworthiness.
"Chip depreciation is 'very, very real'"
Burry turned his attention back to the chips themselves that underpin certain AI loans. He cited a chart from Ares: the residual value of NVIDIA's H100 has dropped by 51% over three years. The rapid depreciation of chips complicates loans collateralized by hardware—precisely why lenders require residual-value guarantees. According to Burry, manufacturers and cloud service providers offer such safeguards "not because they want to, but because they have no choice." He described chip depreciation as "very, very real."
This isn't the first time he's discussed depreciation. Since November 2025, Burry has argued that hyperscale vendors are stretching the useful life of NVIDIA chips far beyond the realistic 2–3-year replacement cycle, suggesting this could cause the entire industry to understate depreciation by roughly $176 billion between 2026 and 2028. According to his estimates, by 2028 Oracle's earnings could be overstated by about 27%, and Meta's by around 21%. NVIDIA struck back with a seven-page memo sent to Wall Street analysts, contending that a four- to six-year lifespan is more realistic, while also correcting a figure cited by Burry: the company's share repurchases since 2018 total $91 billion, not $112.5 billion.
On the hardware side, there is also evidence to the contrary. The CEO of CoreWeave stated that the H100 capacity freed up by expiring contracts has been re‑secured at nearly 95% of the original contract price; meanwhile, NVIDIA CEO Jensen Huang noted that H100 rental rates once surged by 22% in a single month, reaching $3.28 per hour, and cited a CoreWeave contract that leases A100 chips through 2029—nine years after the chip's initial release.
The two metrics being considered are not the same: residual value answers the question "How much is this chip worth today?", while rental income addresses "How much can it still generate in revenue today?" A chip may depreciate sharply yet still generate substantial rental income; therefore, neither figure alone is sufficient to determine an appropriate depreciation period.
The market provided a footnote last week.
Burry's post did not appear out of thin air. 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 fails to come online as scheduled in 2028, Oracle hopes to defer some payments. The trigger is the associated natural gas pipeline—New Mexico's Land Office has twice rejected the pipeline plan, pushing its commissioning date back nearly six months to February 1, 2027. Oracle emphasized that "Project Jupiter remains on our planned timeline," while Blue Owl declined to comment. The $18 billion construction loan provided by roughly 20 banks has fallen to 89–91 cents on the dollar; on the same day, Oracle and Blue Owl shares each dropped about 4%, and Bloom Energy fell 6%. It's important to clarify: trading at a discount does not equate to default—Oracle has not missed any payments, nor has the project been canceled; according to insiders, this is simply a "risk-allocation move."
Put it in the bigger picture, and the pressure becomes clearer. Morgan Stanley's data show that the overall leverage ratio of hyperscale vendors has risen from 0.9 times in the third quarter of 2025 to 1.8 times—more than doubling in just over two quarters—and now exceeds the level of the entire energy sector. Amazon, Meta, Google, Microsoft, and Oracle together account for 4% of the U.S. investment-grade bond index; when weighted by duration, the six major hyperscale vendors plus three leading chip companies make up 9%.
On the financing cost front, the yield on the 10-year U.S. Treasury note has approached 5.17%, rising by roughly one percentage point since the start of the year, forcing AI companies that have re-entered the bond market to offer more attractive coupon rates. JPMorgan estimates that by 2030, AI-related firms will have cumulatively issued $4.1 trillion in debt. Meanwhile, on the equity side, the AI computing‑power sector in both A‑shares and Hong Kong stocks experienced a sharp correction on September 28, with the optical‑module, semiconductor, and copper‑clad laminate sectors leading the decline.
Burry's roadmap, and "the person across from you"
According to Burry's own roadmap, the sixth installment of the series will be released in a few weeks, covering this financing network along with his insurance research; the seventh installment will examine "new cloud" vendors such as CoreWeave and Nebius. His established positions include short positions in Oracle (around $145), Nebius ($211.77), Micron, and Palantir, as well as a short bet on the Philadelphia Semiconductor Index via put options expiring in January 2027.
The opposition is far from quiet. NVIDIA's seven-page memo and the statement by Micron's chief commercial officer constitute the most direct rebuttal— the latter told investors that demand for memory chips will exceed the company's supply capacity at least through 2028, directly countering the view that "current AI demand is overstated." CoreWeave's decision to renew its old‑generation chip contracts at near‑full price provides the strongest counterexample to the "residual value collapse" thesis. Meanwhile, the bearish camp is also expanding: GMO co‑founder Jeremy Grantham calls current AI valuations a bubble, while DoubleLine's Jeffrey Gundlach expects this race to produce clear losers. And an observation metric proposed by former Kynikos analyst Jonathan Weil serves as a reminder: the key term to watch is the "funding gap"—the last time it swept through financial‑industry lexicons alongside the "burn rate" was in 2000.
Burry concluded his post with a single sentence: "The bubble is indeed painfully obvious. Painful."
Editor/Deng