Michael Burry on AI: Why He Wants the OpenAI and Anthropic IPOs Stopped

Michael Burry’s depreciation case is not that GPUs stop working in two or three years. It is that hyperscalers are recovering the capitalized cost of frontier Nvidia systems over five-to-six-year books while the product cycle that justifies those purchase prices is closer to two or three years—so reported earnings are higher than the economic consumption of the asset. He has repeated the same dollar estimate since November 2025 and has spent 2026 clarifying what the claim is not: rentability.

The mechanism he is actually arguing

Under GAAP, useful life is a management estimate. Stretch the life and annual straight-line depreciation falls, so operating income and EPS rise with no change in cash spent. Burry’s point is the mismatch: buyers are “massively ramping capex” on Nvidia chips and servers whose product cycle he puts at two to three years, while simultaneously lengthening the lives used to depreciate compute equipment. He treats that as earnings management of a familiar kind—“one of the more common frauds of the modern era”—not as a claim that the silicon fails.

He later sharpened the definition so critics could not answer him with rental contracts. Depreciation, in his words, “is not a bet on when a chip stops working. It is the recovery of capitalized sunk cost over a defined window of earnings at the frontier, set against the terminal value. A chip can rent and still depreciate very fast economically.” He has said explicitly that he never claimed A100s would stop functioning.

The filings contrast he uses as proof of discretion, not physics: in early 2025 Amazon shortened a subset of server lives from six years to five and cited the pace of AI development, taking extra expense; in the same period Meta extended most server lives to 5.5 years and reduced the depreciation hit. Same generation of hardware, opposite accounting conclusions.

Timeline of the claim

November 10, 2025 — the public launch (his words). On X, as Cassandra Unchained, Burry wrote:

“Understating depreciation by extending useful life of assets artificially boosts earnings -one of the more common frauds of the modern era. Massively ramping capex through purchase of Nvidia chips/servers on a 2-3 yr product cycle should not result in the extension of useful lives of compute equipment. Yet this is exactly what all the hyperscalers have done. By my estimates they will understate depreciation by $176 billion 2026-2028. By 2028, ORCL will overstate earnings 26.9%, META by 20.8%, etc. But it gets worse. More detail coming November 25th.”

The post included a chart. Coverage of that image—not the tweet text itself—describes lives moving roughly as follows: Meta network equipment from about 3 years toward 5.5 by 2025; Alphabet from about 3 to 6; Oracle from 5 to 6; Microsoft to 6; Amazon from 4 (2020) to 6 (2024), then a subset back to 5 in 2025. Those year-by-year labels are secondhand descriptions of his graphic.

CNBC the next day reported the same figures and noted it could not independently confirm the practice, and that GAAP gives companies leeway on useful-life estimates.

November 23–26, 2025 — Substack, promised detail (mostly paywalled). He launched Cassandra Unchained and the series The Heretic’s Guide to AI’s Stars. Part 1 (Nov. 23), “The Cardinal Sign of a Bubble: Supply-Side Gluttony,” frames Nvidia as the Cisco of the buildout and the five public spenders—Microsoft, Google, Meta, Amazon, and Oracle—as promising nearly $3 trillion of AI infrastructure over three years. Part 2 (dated Nov. 25/26), “Unicorns and Cockroaches… The Depreciation Problem,” opens on 10-K reading (“So, I found something…”) and is the piece he later cites as where he covered the depreciation argument. The body is paid; the $176 billion / 26.9% / 20.8% figures in wide circulation are the November 10 post, not a new public number from this essay.

Around the same days Nvidia circulated a seven-page analyst memo that named him. CNBC quotes his Substack reply: Nvidia emailed sell-side analysts “to push back on my arguments on [stock-based compensation] and Depreciation… I stand by my analysis. I am not claiming Nvidia is Enron. It is clearly Cisco.” Nvidia’s counter, as reported: customers depreciate GPUs over four to six years; 2020 A100s still run at high utilization and retain economic value beyond the two-to-three years critics claim. That is Nvidia’s position, not his.

July 8–9, 2026 — definition and the Amazon/Meta split (his words). In a Substack “Short Thoughts” post and a matching X post he rejected the rentability rebuttal and put filings numbers on the discretion point:

• Amazon shortened a subset of server lives from 6 years to 5, “explicitly cited the pace of development,” and took “additional expense across 9 months of $677 million.”
• Meta, “during the same quarter,” extended useful server lives to 5.5 years, “reducing the depreciation hit to net income by $2.9 billion.”
• “Depreciation cannot move two directions at once if it were a measurement of something. It is an economic lever, not a physical measurement.”
• He attributes to Satya Nadella (Dwarkesh Patel podcast): “I didn’t want to get stuck with four or five years of depreciation on one generation.”
• He attributes to Jensen Huang at GTC: when Blackwell ships in volume, “you couldn’t give Hoppers away.”
• He points readers back to Part 2 of the Heretic’s Guide for the fuller treatment.

(Filings coverage often states Meta’s 2025 change as about $2.92 billion less depreciation expense and about $2.59 billion more net income. Burry’s own phrasing is the $2.9 billion net-income figure above. Amazon’s own disclosures are widely summarized as roughly $700 million of operating-income impact plus accelerated depreciation; his stated figure is $677 million over nine months.)

September–October 2026 — residual-value slides, not a new dollar estimate. On October 1, 2026 he published “Don’t Believe Your Lyin’ Eyes, GPU Depreciation & Useful Lives” (Hystory Rhymes, Part II). The free portion is a response to Nvidia’s September 27, 2026 investor slide, “Nvidia AI Infrastructure Retains Value Beyond Accelerated Depreciation Schedules,” which plots purported retained value of A100, H100, and B200 against a five-year accelerated curve—implying, he writes, that companies are over-depreciating. He maps the dispute onto a 1968 scene in Adam Smith’s The Money Game, in which a speculator mocks anyone still asking “how fast these computers are written off,” while “Billy the Kid” pitches computer-leasing stocks on the claim that “the need for computers is practically infinite.”

Summaries of the paywalled remainder (secondhand) say he argues Nvidia’s residual values are discounted-cash-flow constructs, not secondary-market prices; that firmer rental rates on older GPUs reflect memory and power bottlenecks rather than durability; and that 1960s computer lessors depreciated over about eight years versus IBM’s about four, then saw rental income collapse when the next generation shipped, with leasing stocks down on the order of 80% by 1970. Those specifics are not in the free text retrieved here.

Through mid-October 2026, secondary coverage still cites the original $176 billion (2026–2028), Oracle ~26.9%, and Meta ~20.8%. No primary post retrieved for this reconstruction replaces those figures with a new industry total.

Table: claims, evidence, dates, attribution

Date	Claim	Figures / names	Evidence he points to	Attribution
Nov. 10, 2025	Extending useful lives while buying 2–3 year product-cycle Nvidia chips/servers understates depreciation and boosts earnings; “one of the more common frauds of the modern era”	Industry understatement $176 billion, 2026–2028; by 2028 ORCL earnings overstated 26.9%, META 20.8%; “etc.”; “all the hyperscalers”	His estimates; chart attached to the X post	His words (X)
Nov. 10, 2025 (chart, as described by outlets)	Useful lives of network/compute assets were lengthened after 2020	Meta ~3 → 5.5 years; Alphabet ~3 → 6; Oracle 5 → 6; Microsoft → 6; Amazon 4 → 6 then a 2025 cut toward 5	Graphic in the same post; outlets say it tracks 10-K lives	Secondhand description of his chart; year path not in the tweet text
Nov. 23, 2025	AI buildout is supply-side gluttony with Nvidia as Cisco	Five public “horsemen”: Microsoft, Google, Meta, Amazon, Oracle; nearly $3 trillion infrastructure spending over three years	Substack, Supply-Side Gluttony	His words as quoted by Reuters/CNBC; full essay paid
Nov. 25–26, 2025	Detailed depreciation case (“The Depreciation Problem”); stands by the analysis after Nvidia’s memo	No new public industry total in open text; “I am not claiming Nvidia is Enron. It is clearly Cisco.”	10-Ks; Substack Part 2; reply to Nvidia’s analyst memo	His words for the Cisco line (CNBC quoting Substack). Essay body not public. Nvidia’s “4–6 years / A100s still earning” line is Nvidia’s, reported by CNBC
July 8–9, 2026	Economic life ≠ physical life or rentability. Opposite life changes on similar hardware show depreciation is a lever	AMZN: subset 6 → 5 years; +$677 million expense over 9 months. META: to 5.5 years; −$2.9 billion depreciation hit to net income. Nadella: did not want 4–5 years of depreciation on one generation. Huang: Hoppers unsellable once Blackwell ships in volume	Company disclosures; podcast; GTC; cross-reference to Heretic’s Guide Part 2	His words (Substack and X). The Nadella and Huang lines are quotations he attributes
Oct. 1, 2026	Nvidia’s Sept. 27 residual-value slide (A100/H100/B200 vs a 5-year accelerated curve) does not show over-depreciation; the question rhymes with 1968 computer leasing	Slide title and chip names as he states them; historical parallel to The Money Game	Nvidia investor presentation; Adam Smith, 1968	His words in the free portion. DCF-vs-market, bottleneck, 8-year vs 4-year, and ~80% leasing-stock decline are secondhand summaries of the paid section
Nov. 2025–Oct. 2026 (press)	Same industry math, often rounded	“about $176 billion”; Oracle “roughly 27%” or 26.9%; Meta “about 21%” or 20.8%; sometimes Microsoft/Google/Amazon added as using 5–6 year lives	Restatements of the Nov. 10 post	Secondhand. A stray “$205 billion write-down” figure appears in at least one commentary and is not his stated $176 billion estimate

What is his, what is not

His own stated numbers, still the load-bearing ones: 2–3 year Nvidia product cycle versus extended compute lives; $176 billion understated depreciation from 2026 through 2028; Oracle 26.9% and Meta 20.8% earnings overstatement by 2028; Amazon $677 million / Meta $2.9 billion as the paired 2025 estimate changes; the definition that rentability is not the claim.

Secondhand or only partly visible: the year-by-year useful-life table implied by his November 10 chart; the arithmetic inside the $176 billion (asset base, which lives he resets to 2–3 years, tax rate); most of “Unicorns and Cockroaches”; and the October 2026 residual-value and 1968 leasing numerics behind the paywall. Press often rounds 26.9/20.8 to 27/21 and sometimes attributes 5–6 year lives to Microsoft, Google, Amazon, and Oracle as if he listed each in the tweet. He named Oracle and Meta in the text and “all the hyperscalers” as the set.

Not his claim: that chips are physically dead in three years; that A100s cannot be rented into the late 2020s; that Nvidia itself is capitalizing expenses the way WorldCom did. He has rejected the first and has said the Cisco analogy is about the capital cycle, not Enron-style fraud at Nvidia. Nvidia’s November 2025 memo and later residual-value slides are the other side of the argument: four-to-six-year customer lives and continued rental income on older GPUs. Burry’s July 2026 reply is that a booked rental contract does not measure economic depreciation of frontier cost.

---

Recent Findings Supplement (October 2026)

Michael Burry has continued and expanded his argument on hyperscaler accounting for AI hardware depreciation primarily through his Substack "Cassandra Unchained" (including the "Heretic’s Guide to AI’s Stars" series and "Hystory Rhymes" posts) and X activity in 2026. The core claim is that hyperscalers (notably Amazon, Meta, Alphabet/Google, Microsoft, and Oracle) extend the assumed useful lives of servers and GPUs (typically to 5–6+ years) beyond their realistic economic life (Burry argues ~2–3 years due to rapid Nvidia generational advances), thereby understating depreciation expense and overstating earnings.[1][1]

This builds on his earlier public statements (from late 2025) but includes new posts, responses to company actions, and refinements after April 16, 2026. Key new elements include explicit responses to Nvidia’s September 2026 investor materials, additional forensic analysis of filings (e.g., leases and construction-in-progress), and clarification that the issue is economic recovery of sunk costs at the technology frontier, not physical failure of chips.[2][3]

Recent developments (post-April 2026) center on Burry’s July–October 2026 Substack and X activity, company-specific accounting disclosures cited by him, Nvidia’s direct rebuttal, and secondary coverage of the debate. No major new regulatory changes or independent third-party audits of the estimates appear in recent sources; the discussion remains driven by Burry’s filings-based analysis and company responses.[4]

Timeline and Key Claims Table

Claims are drawn from Burry’s direct Substack/X output (flagged as “Burry own words”) or secondary reports citing his posts/filings analysis (flagged as “secondhand/reporting”). Only post-April 16, 2026 information is included. Figures like the $176 billion estimate originate in his 2025 work but are reiterated and contextualized in 2026 posts.

• July 8/9, 2026 (Burry own words, Short Thoughts Substack post): Clarifies “Depreciation is not a bet on when a chip stops working. It is the recovery of capitalized sunk cost over a defined window of earnings at the frontier, set against the terminal value.” Cites Amazon shortening subset of server lives to 5 years from 6 (explicitly due to pace of development; +$677 million additional expense over 9 months) vs. Meta extending most servers/network assets to 5.5 years (reducing 2025 depreciation expense by ~$2.9 billion / increasing net income by ~$2.59 billion). Quotes Satya Nadella (Microsoft): “I didn’t want to get stuck with four or five years of depreciation on one generation.” References Part 2 of Heretic’s Guide on the depreciation problem.[3][3]

• September 19–26, 2026 (Burry own words, Heretic’s Guide to AI’s Stars Parts IV & V Substack posts): Forensic review of Big 5 hyperscaler (MSFT, AMZN, ORCL, META, GOOG) 10-K/10-Q footnotes. Part IV focuses on “the Missing $3 Trillion” (off-balance-sheet commitments, leases, construction-in-progress). Part V includes rankings and maps.[5]

• September 24, 2026 (Burry own words, referenced in reporting): Expands warning to construction-in-progress assets across Amazon, Meta, Alphabet, Microsoft, and Oracle (these do not yet generate depreciation under GAAP).[1]

• October 1, 2026 (Burry own words, “Don’t Believe Your Lyin’ Eyes, GPU Depreciation & Useful Lives” / Hystory Rhymes Part II Substack post): Responds to Nvidia’s September 27, 2026 investor presentation slide (“Nvidia AI Infrastructure Retains Value Beyond Accelerated Depreciation Schedules”), which plotted A100/H100/B200 retained value against a 5-year accelerated curve to argue companies may be over-depreciating. Burry uses a 1968 analogy from Adam Smith’s The Money Game (the “Great Winfield” dismissing depreciation concerns in the computer leasing boom) to highlight recurring patterns. Notes the slide appears aimed at critics like himself.[2][6]

• October 2, 2026 (Burry own words, X post): Shares/promotes the October 1 Substack post.[6]

• Ongoing/October 2026 (Burry own words and secondhand): Reiterates the $176 billion industry-wide understatement of depreciation (2026–2028) and specific impacts (Oracle earnings overstated by ~27%, Meta by ~21% by 2028). Cites hyperscaler policy changes (Meta to 5.5 years from 2025; Oracle servers/networking from 5 to 6 years; Microsoft from 4 to 6 years; Google to 6 years). Nvidia sent a 7-page memo to analysts (naming Burry) arguing 4–6 years is realistic because A100s (2020) remain in use and rental demand persists.[1][4]

Supporting context from company actions and pushback (new or highlighted in 2026 coverage): Meta’s 5.5-year extension cut 2025 depreciation by ~$2.9 billion. Microsoft’s fiscal 2027 change extending data center/office building lives from 15 to 25 years reduces reported capex (e.g., 2026 forecast drop of ~$15 billion via lease classification shifts) and future depreciation. Counter-examples cited against Burry include CoreWeave A100 contracts through 2029 and Nvidia/analyst arguments for longer economic utility.[4][7]

Implications for competitors/entrants: Burry’s filings-driven approach highlights the value of scrutinizing depreciation policy footnotes, lease commitments, and capex classifications in hyperscaler 10-Ks/10-Qs, especially amid rapid tech cycles. The debate underscores accounting discretion as a potential earnings lever but also the difficulty of proving economic obsolescence versus physical utility in real time. Secondary sources note the argument remains contested, with demand-side rebuttals (e.g., from Steve Eisman) emphasizing utilization over schedules.[8]

All monetary figures are in USD as reported. Sources are limited to post-April 16, 2026 publications or references to Burry’s 2026 activity; full Substack text is often paywalled, so details rely on consistent reporting across multiple outlets. No new independent verification of the $176 billion aggregate or exact 2028 overstatement percentages beyond Burry’s estimates appears in the results.

Michael Burry’s late-September and early-October 2026 posts treat OpenAI and Anthropic not as pure technology bets but as companies trying to buy seats inside an AI cloud oligopoly that Big Tech already expects government to protect. He wants equity markets to fall hard enough that those IPOs cannot clear, because public capital would lock in spending commitments whose economics, in his view, have not been proven and may never pay back.

The oligopoly thesis, in his words

On October 6, 2026, posting as Cassandra Unchained, Burry wrote: “The only reason the CEOs of these tech monopolies are spending wildly on data centers like they are is because they assume they will be granted an oligopoly that is too big too fail.” The next day, in the Substack note “Short Thoughts: The AI Cloud Oligopoly,” he extended the argument to the labs and to Oracle. Monopolies, he said, “need not spend wildly as if their lives depended on it.” The hyperscalers (excluding Apple) are doing so because they expect government to “tolerate their oligopoly on AI.” OpenAI, Anthropic, and Oracle are not monopolies, but “they intend to be part of the oligopoly. None can imagine breaking the dominance of a Microsoft, Amazon, Google, or Meta within each silo.” He grants them only “the craven desire to expand a monopoly, to join an oligopoly.” That is “a want, not a need,” and “human thought is far too redundant for what is being built.” The end state he expects is compression: “much of what is being spent will be just so much sunk cost.”

The mechanism he is describing is a land grab for durable rents, not a response to proven unit economics. Once public shareholders finance the labs, the same capital that funds the build-out also makes the resulting capacity politically harder to strand. Ed Zitron’s public reply, reported the same week, pushed the opposite concentration point: cloud growth is already heavily dependent on OpenAI and Anthropic demand, so the oligopoly bet may rest on two customers rather than a broad market.

Funding, compute commitments, and the circular loop

Burry’s objection is not that the labs lack money. It is that the money is largely supplier-financed and then spent back on the suppliers.

Anthropic’s confidential IPO prospectus, reviewed by Reuters in late September 2026, shows at least $518 billion of cloud, computing, and infrastructure commitments over roughly a decade with six partners, about 80 percent non-cancelable or payable regardless of usage. Of that, Reuters reported at least $111.1 billion with Google (April 2026–July 2033), $110 billion with Amazon (May 2026–April 2036), and $31.4 billion with Microsoft (November 2026–May 2033). The filing language, as reported, is take-or-pay: “If our actual spend falls short, we must pay Google the difference,” with similar terms for Amazon; the Microsoft commitment is non-cancelable except for uncured material breach. Equipment leases tied to Broadcom are also described as largely non-cancelable. Against that, the same prospectus showed 2025 revenue of about $4.6 billion, a net loss of about $42 billion (much of it an accounting charge on financing that could convert to equity rather than cash operating spend), and an operating loss that Reuters put near $8.06 billion. Compute and infrastructure spend in 2025 was about $7.33 billion.

Private capital has already scaled the valuations. Anthropic’s May 2026 Series H raised $65 billion at a $965 billion post-money valuation. OpenAI’s March 2026 round raised $122 billion at an $852 billion valuation; by early October it was reported in talks for a further private round of at least $30 billion at valuations discussed around $1.2–1.4 trillion or higher. Those rounds include the same strategic partners that sell them compute.

The circular piece Burry has highlighted most sharply is Nvidia’s role as investor, guarantor, and chip supplier. In late July 2026 he responded to reports of a large Nvidia financing backstop for an OpenAI-linked Ohio campus with “Around and around we go.” After Nvidia’s subsequent 10-Q disclosed guarantees “capped at a total of $105 billion” for credit support on land, power, and shell for an OpenAI-leased campus at SoftBank’s SB Energy PORTS site in Ohio—about 4.25 gigawatts of IT load that would “exclusively host NVIDIA AI infrastructure”—he called it “Whistling past the graveyard.” Nvidia’s CFO has acknowledged that some will call the support circular financing and said the company sees it differently. Public reporting also places Nvidia’s equity contribution to OpenAI’s large 2026 round at about $30 billion (well short of an earlier letter-of-intent figure of up to $100 billion) and up to $10 billion toward Anthropic alongside Microsoft.

The wider web, which Burry has circulated via a Bloomberg diagram, runs in both directions. Reported multi-year purchase commitments include roughly $250 billion from OpenAI to Microsoft Azure and on the order of $300 billion tied to Oracle. Oracle’s remaining performance obligations have been cited by Burry in the hundreds of billions, with his claim that “the customer is now the lender” because incremental capital for the build-out comes from the customer side of the contract. Amazon and Alphabet are both large Anthropic shareholders and large counterparties on Anthropic’s locked compute spend; Microsoft is both a major OpenAI owner and a major Azure counterparty. The loop Burry and others describe is simple: a hyperscaler or chipmaker funds or guarantees a lab, the lab commits to buy that firm’s compute or chips, and the supplier books revenue (and often marks up its equity stake) even if end-user cash profits have not caught up.

The “tank hard” call: quotes, date, and reasoning

On September 29, 2026, Burry posted: “For the benefit of humanity, the markets should tank hard and prevent the OpenAI and Anthropic IPOs.” In the same thread he wrote: “These are companies that are going suck up and then destrop TRILLIONS of dollars of capital, and that will be the least of the damage they do.” (The misspelling “destrop” is in the original post.) When a user replied that the market should tank so “Skynet can’t IPO,” Burry answered: “Along those lines.”

The reasoning is capital-rationing, not a prediction that models will become Terminator. An IPO, in his framing, is the mechanism that lets the labs raise sums private markets cannot permanently supply and that converts today’s circular commitments into public-market liabilities. If markets fall hard enough that the window closes, the industry is forced to confront spending before the commitments get larger and before “too big to fail” becomes a political fact. He had already called September safety-slowdown essays from OpenAI and Anthropic executives “self-serving,” writing that “IPOs need hype & puffery” and that “we are so awesome it could become dangerous” is itself hype. The September 29 posts landed as the White House hosted AI executives on voluntary self-regulation, which he treated as further evidence that political cover and capital access are moving together.

Critics in the replies and in later coverage noted that Burry has held put exposure tied to AI-linked names, so a crash would also serve those positions. That does not change the stated mechanism: stop the IPOs so the capital sink cannot be socialized through public equity.

IPO status as of mid-October 2026

OpenAI. The company confidentially submitted a draft S-1 on June 8, 2026, with Goldman Sachs and Morgan Stanley reported as lead banks, after earlier talk of a late-2026 listing. On August 19, CFO Sarah Friar told employees the company would be public in 2027, or sooner only if the business “continues to inflect.” On September 12, Sam Altman told Fortune that “right now would be an ill-advised moment to go public” given safety work, and when pressed said “I would say not 2026.” As of mid-October there is no public prospectus, no price range, and no confirmed quarter inside 2027. The company continues to raise privately; reports in late September and early October described early talks for at least $30 billion more at valuations discussed from roughly $1.2 trillion to $1.4 trillion or higher. Annualized revenue figures circulating among investors have been disputed in the press, with one Financial Times account putting the end-of-September run rate near $50 billion versus higher figures previously given to backers.

Anthropic. It confidentially submitted a draft S-1 on June 1, 2026. A public S-1 had not appeared on EDGAR by mid-October. On October 14 it held a pre-IPO investor day in San Francisco with CEO Dario Amodei and CFO Krishna Rao. Attendees were told third-quarter revenue was $16.8 billion, up about 45 percent from $11.6 billion in the second quarter (and from about $4.7 billion in the first), and that the company was profitable on an adjusted basis, with some accounts also citing positive free cash flow and roughly $1.70 of revenue per dollar of compute spend. Sequential growth had slowed sharply from the prior quarter’s more-than-doubling, and by October 15 Reuters reported that some current and prospective investors were discussing a valuation around $1.5 trillion, down from the roughly $2 trillion figure that had been in circulation for a November sale. Formal marketing has been reported as potentially starting the week of November 9, with a listing aimed after the November midterms and, in some accounts, before Thanksgiving—subject to the required public filing window before a roadshow. Goldman Sachs and Morgan Stanley have been reported as leading the deal. The May private mark remains $965 billion. The $518 billion commitment stack and the 2025 loss profile are still the numbers the leaked prospectus put in front of the market.

What this implies for anyone on the other side of the trade

Burry’s position is that the binding constraint is not model quality but the financing structure. Take-or-pay contracts with the same firms that own equity in the labs turn variable compute cost into something closer to fixed debt, while an IPO would refinance that structure with public equity at valuations that assume the oligopoly rents arrive before compression does. A competitor or a public-market investor does not need to share his “benefit of humanity” framing to track the same variables he is using: how much of Anthropic’s $518 billion is truly non-cancelable, whether OpenAI’s 2027 window stays open if private rounds keep clearing above $1 trillion, whether Nvidia’s guarantee book and hyperscaler remaining performance obligations keep rising faster than cash returns, and whether the October 14 deceleration in Anthropic’s sequential growth is a one-quarter pause or the start of the compression he expects to turn much of the spend into sunk cost.

---

Recent Findings Supplement (October 2026)

Michael Burry has argued that OpenAI and Anthropic seek to join (rather than disrupt) the existing Big Tech AI oligopoly dominated by Microsoft, Amazon, Google, and Meta, viewing their massive spending and IPO ambitions as unsustainable bets that could destroy trillions in capital.[1]

His commentary, primarily via X posts and Substack in September–October 2026, ties directly to their funding rounds, enormous compute/infrastructure commitments, and plans to access public markets. He has explicitly called for markets to decline sharply enough to block their IPOs.[2]

Burry’s View on Oligopoly Ambitions

Burry frames OpenAI and Anthropic not as challengers but as aspiring members of a protected oligopoly. In an October 2026 Substack post, he stated that these companies “intend to be part of the oligopoly. None can imagine breaking the dominance of a Microsoft, Amazon, Google, or Meta within each silo.” He described their drive as a “craven desire to expand a monopoly, to join an oligopoly,” but dismissed it as a “want, not a need,” arguing human thought is “far too redundant for what is being built.”[1]

He extended this to Big Tech hyperscalers, claiming their data center spending spree occurs because CEOs “assume they will be granted an oligopoly that is too big [to] fail.”[3]

This perspective underpins his skepticism about the durability of their economics and capital needs.

Burry’s Warnings on IPOs and Call for Market Decline

In late September 2026 (around Sept 29–30), Burry posted on X under @michaeljburry (Cassandra Unchained): “For the benefit of humanity, the markets should tank hard and prevent the OpenAI and Anthropic IPOs.”[2]

In follow-ups, he added: “These are companies that are going suck up and then destroy TRILLIONS of dollars of capital, and that will be the least of the damage they do.” He agreed with a user’s “Skynet can’t IPO” joke.[4]

Earlier in September 2026 (around Sept 14), he called AI leaders’ public warnings about slowing development “self-serving,” writing that “IPOs need hype & puffery; ‘we are so awesome it could become dangerous’ is hype & puffery.” He argued LLMs are “not AI and won’t be AGI,” so there is “nothing AI to slow down,” and that slowdown talk benefits incumbents while competition rises.[5]

In early October 2026 Substack commentary, he benchmarked Anthropic’s valuation against historical precedents (e.g., UPS’s inflation-adjusted pre-IPO record of ~$119 billion) and noted that Anthropic’s figure could buy “78” profitable S&P 500 companies (examples: Domino’s, Clorox, lululemon). He described stocks as in a “denial stage” expected to last 6–9 months.[6]

Implication for entrants/competitors: Burry sees IPOs as the mechanism that would lock in unsustainable capital flows; blocking them via a market decline could force a reckoning on unproven economics before more damage occurs.

Funding Rounds, Compute Commitments, and Infrastructure Scale

Anthropic’s May 2026 Series H round raised $65 billion at a $965 billion post-money valuation, led by investors including Altimeter Capital, Dragoneer, Greenoaks, and Sequoia, with participation from strategic infrastructure partners.[7]

A leaked draft prospectus (reported by Reuters in late September 2026) detailed:
- 2025 revenue of ~$4.6 billion (roughly 12x growth).
- Net loss of ~$42 billion (much of it an accounting charge, per reports; operating loss cited around $8 billion in some coverage).
- $7.33 billion spent on compute and infrastructure in 2025.
- ~$518 billion in future cloud, computing, and infrastructure commitments (with reports noting a high percentage non-cancelable).[2]

More recent updates (October 2026) showed Anthropic reporting Q3 2026 revenue of $16.8 billion to prospective investors.[8]

OpenAI has pursued large private raises (including a reported $122 billion round in March 2026) and additional funding talks (at least $30 billion more at ~$1.4 trillion valuation), alongside major compute deals.[9]

Burry has highlighted the scale of these commitments as evidence that IPO proceeds would fuel further expansion before returns are proven, rather than resolving underlying economics.[7]

Circular Financing and Ties to Nvidia, Hyperscalers

Burry has criticized accounting practices around AI infrastructure depreciation by hyperscalers (Amazon, Meta, Alphabet/Google, Microsoft, Oracle), arguing useful lives are overstated (e.g., extending server life to 5+ years vs. a more realistic 2–3 years for chips), potentially understating depreciation industry-wide by tens to hundreds of billions through 2028.[10]

Public reporting notes complex interdependencies: OpenAI has deals spanning Microsoft (primary backer), Amazon, Google, and Oracle (including a reported $300 billion compute access deal with Oracle); Nvidia has engaged in vendor financing-style arrangements (e.g., investments or support tied to chip purchases); and Anthropic has cloud partnerships with multiple providers.[11]

Burry links these to oligopoly protection and questions sustainability, noting compression advances could render much spending “sunk cost.”[1]

IPO Plans and Status as of Mid-October 2026

Anthropic: Confidential S-1 filing in June 2026. Targeting a November 2026 Nasdaq IPO (after U.S. midterms, potentially before Thanksgiving), with marketing/roadshow activity eyed for the week of Nov. 9. Pre-IPO investor events pitched at $1.8–2 trillion valuation (roughly double the May 2026 private mark); some reports note $100 billion potential raise. As of early-mid October 2026, investor meetings were underway or imminent, with some prospective investors reportedly tempering expectations amid revenue growth concerns.[12]

OpenAI: Explicitly ruled out a 2026 IPO. CEO Sam Altman stated in September 2026 that it would be “ill-advised” due to safety and alignment priorities; the company is focusing on those issues and pursuing further private capital instead of public listing (targeting 2027 at earliest).[13]

These timelines represent shifts from earlier 2026 expectations, enabled in part by continued private market access for the largest players.[9]

No major new regulatory or policy developments directly altering these plans appear in the most recent reporting. Burry’s bearish positioning (including puts on indices like NDX/SMH and names like Oracle/Palantir) has evolved with “moving timelines up” on his AI thesis.[14]

Overall, recent developments center on Burry’s sharpened public critique timed with Anthropic’s prospectus disclosures and IPO momentum, contrasted with OpenAI’s delay—highlighting the tension between private capital scale and public market scrutiny.

Michael Burry does not price Anthropic on a sales or earnings multiple. In an October 8, 2026 post, he prices it as a pile of cash large enough to buy 78 profitable S&P 500 companies, then contrasts that pile with UPS’s 1999 IPO, which he puts at 26 times earnings and 2.4 times sales. He publishes no Anthropic revenue, margin, or growth assumption. The only explicit multiples in the post belong to UPS.

What Burry actually did

The mechanism is a market-cap shopping list, not a valuation model. On X as Cassandra Unchained, he wrote that “a fun game in times like these” is to count how many profitable S&P 500 firms “the bubble private company valuation” can buy. His count for Anthropic is 78. Named examples are Domino’s, Clorox, J.M. Smucker, Stanley Black & Decker, Deckers, Lululemon, McCormick, Tractor Supply, NVR, Albemarle, FedEx Freight, News Corp, Alliant Energy, Kimco, Hormel, Weyerhaeuser, DaVita, Zimmer Biomet, Lennox, Masco, A.O. Smith, MGM, Wynn, Brown-Forman, Norwegian Cruise Line, and Huntington Ingalls, “and many more.” He did not publish the full 78, the market-cap source, or the Anthropic figure used in the sum. Secondary coverage maps the exercise to the May private round at $965 billion. [1] [2]

The only multiples he states are historical, and they are UPS’s, not Anthropic’s: the largest inflation-adjusted pre-IPO valuation of 1990–2000 was UPS at $119 billion, then 92 years old, “26 times earnings at a net margin of about 8.6%, 2.4x sales.” [1] In a June Substack comment after the May round, he wrote there was “no guarantee, and not even a strong likelihood,” that Anthropic is long-term worth anywhere near $1 trillion. [2] That is a judgment about durability, not a published forecast.

Independent check of his inputs

The $965 billion mark is company-confirmed, not a Burry invention. On May 28, 2026, Anthropic said it raised $65 billion in Series H at a $965 billion post-money valuation, led by Altimeter, Dragoneer, Greenoaks, and Sequoia, including $15 billion of previously committed hyperscaler money. The same announcement said run-rate revenue had crossed $47 billion earlier that month, up from $14 billion in February and about $9 billion at the end of 2025. [3] [4] [5] Reported IPO talk since then is $1.5–2 trillion, not a priced deal. Reuters has described a sought valuation of more than $2 trillion; an October 16 secondary report said investor-day discussion had drifted toward about $1.5 trillion. Treat both as reported targets. [6] [7] [8]

The UPS math is in the right neighborhood, not exact. At the November 1999 offer price of $50, contemporaneous reports put equity value a little above $60 billion on a sale of about 9% of the company. [9] [10] First-day close was about $81–82 billion. [11] [12] CNN reported trailing net income of $1.95 billion through September 30, 1999, a P/E of 28 at the offer and almost 38 at the close, and trailing revenue of $26.3 billion. [13] Offer-price value over that sales figure is about 2.3 times, close to Burry’s 2.4 times. His 26 times earnings is near CNN’s 28 times, not a match to 1998 net income of about $1.7 billion, which would have implied a mid-30s multiple at the offer. [11] CPI-U was 336.940 in September 2026. [14] Roughly doubling a ~$60 billion 1999 offer value lands near his $119 billion. He appears to be inflating the offer-price equity value, not the first-day close. Age (92) matches the 1907 founding cited at the time. [15]

The 78-company average is the bottom of the index, which is the tell. $965 billion divided by 78 is about $12.4 billion per name. That lines up with the companies he bothered to name, not with a typical S&P 500 firm:

• Domino’s: about $9.9–10.0 billion market cap, trailing P/S about 2.0, trailing P/E about 16.7, on about $5.1 billion of revenue and about $609 million of net income (mid-October 2026). [16]
• Clorox: about $10.1–10.4 billion. [17] [18]
• Hormel: about $11.1 billion. [19]
• J.M. Smucker: about $12.8 billion; McCormick: about $12.3 billion, in the same peer table. [19]

He listed Clorox twice and misspelled several names. The sum is not auditable from the post. “FedEx Freight” is also an odd index name relative to FedEx itself. The game works because he is shopping the small profitable tail.

Revenue base versus the price he is attacking

Burry’s post contains no revenue assumption. Public estimates, which should stay labeled as such because the prospectus is still confidential, move the multiple a lot depending on which base you use.

Item	Figure	Status	Implied multiple
May 28, 2026 post-money	$965 billion; $65 billion raised	Company-confirmed [3] [4]	—
Run rate at that round	Crossed $47 billion in May	Company-stated [4]	About 20.5x that run rate at $965 billion
2025 revenue	Nearly $4.6 billion, ~12x 2024	Leaked draft prospectus via Reuters [6]	About 210x at $965 billion; PitchBook puts $2 trillion at about 435x 2025 sales [20]
2025 operating loss	$8.06 billion; compute spend $7.33 billion	Reuters on the draft filing [6]	Not meaningful as a P/E
2025 GAAP net loss	About $42 billion, of which about $34 billion was a non-cash charge on financing that can convert into shares	Reuters [6]	Do not use as cash burn
Q1 / Q2 2026 revenue	About $4.7 billion, then about $11.5 billion	Draft filing / subsequent reporting [20] [21]	$2 trillion is under 44x annualized Q2, per PitchBook [20]
Late-July run rate	$65 billion annualized	Reuters, cited again Oct 15 [7]	About 14.8x at $965 billion; just under 31x at $2 trillion [20]
Q3 2026 revenue	$16.8 billion, up about 45% from Q2, vs about 145% the prior quarter	Investor meeting, people familiar, Reuters Breakingviews, Oct 15 [7] [22]	Flat annualization ≈ $67 billion; about 14.4x at $965 billion, about 30x at $2 trillion
Adjusted profit / gross margin	Second straight adjusted profit; GAAP gross margin 85% in the year through September, vs 40% in 2025; $1.70 of revenue per compute dollar	Same investor-meeting reporting [7] [23]	Unaudited; not a public filing
Compute obligations	About $518 billion over coming years, ~80% non-cancellable or payable regardless of use	Reuters on the draft filing [6] [24]	The number Burry’s shopping list does not touch
Sought IPO value	Up to about $2 trillion; some Oct 14 discussion nearer $1.5 trillion	Reported targets, not a price [6] [8]	~22–30x a ~$67 billion Q3 annualization

S&P 500 context, also as of mid-October 2026: FactSet’s forward 12-month P/E was 19.1 on October 11, equal to the 10-year average and below the 5-year average of 19.8. [25] Multpl’s price-to-sales ratio was 3.82 on October 15, but that series is estimated off sales through September 2025, so it lags the 2026 earnings boom. [26] Nvidia, the obvious public AI comp Burry did not use, was around 17–18 times trailing sales, with trailing P/E near 27–29 and forward P/E in the high teens, on about $303 billion of trailing revenue. [27] [28] Median public SaaS/cloud is a different planet: about 4.2 times revenue on the SaaS Capital and Clouded Judgement reads in late September 2026, with only the fastest growers in the high teens and a handful (Palantir, CrowdStrike, Cloudflare) still above 30–40 times forward sales. [29] [30] [31]

Where the framing is vulnerable

The comparison is built to maximize the gap. Staples and industrials at roughly 2 times sales look like UPS in 1999. They are not the peer set a growth investor is using. Against Nvidia’s high-teens sales multiple, a $965 billion price on a $65–67 billion run rate is in the same zip code, not eight times “the market.” Against median software at ~4 times, or the S&P at ~3.8 times sales and ~19 times forward earnings, it is still a large premium. Burry never makes that distinction, so the post reads as if every dollar of S&P equity is a Clorox dollar.

He also freezes the private mark and ignores the revenue base that moved under it. The $965 billion round was struck when the company itself was citing a $47 billion run rate. By the time he posted on October 8, Reuters had already reported a $65 billion late-July run rate. A week later, investor-day figures put the quarter at $16.8 billion. Using 2025’s $4.6 billion makes Anthropic look absurd (hundreds of times sales). Using the current run rate makes the May mark about 15 times sales. Both can be true. A valuation argument that never states the denominator cannot be stress-tested.

The profitability contrast is partly real and partly a snapshot. UPS was an 8–9% net-margin compounder. Anthropic’s 2025 GAAP loss of ~$42 billion is a bad headline for this purpose: Reuters says about $34 billion was a non-cash revaluation of convertible financing, while the operating loss was $8.06 billion on $4.6 billion of revenue. [6] By the second and third quarters of 2026, the same reporting trail describes a second adjusted operating profit and an 85% gross margin. Those claims are unaudited and came from a roadshow. They do mean “zero earnings forever, therefore worth the staples basket” is no longer the clean argument. The cleaner one is the $518 billion of mostly non-cancellable compute obligations against customers who, on the draft filing, can often cut spending. [6] [24] Burry’s game does not price that liability.

Three structural mismatches remain even if the 78-count is arithmetically fair:

• Last-trade market caps are not takeover prices. Buying 78 public companies would require control premiums. The $965 billion figure is a negotiated private round, not a bid for the whole company in the open market. Neither side of the “buys” verb is a real transaction.
• The named basket is the cheap tail. An equal-weighted grab of ~$12 billion names says nothing about how the S&P 500’s earnings and sales are actually distributed. You can “buy” 78 profitable firms and still own a rounding error of index earnings.
• The AI capex loop is inside the index he is using as the alternative. Reported commitments include on the order of $110 billion to Amazon, $111 billion to Alphabet, and $31 billion to Microsoft. [24] Anthropic’s private mark and the mega-caps’ public marks are not independent claims on separate economies.

The vulnerability that has shown up since the post is growth, not the shopping list. Reuters Breakingviews notes quarter-on-quarter growth slowed from about 145% to about 45%, with August and September averaging about $5.7 billion a month versus a July pace implied by the $65 billion run rate of roughly $5.4 billion. Usage can rise while token prices fall. A $2 trillion ask is about 30 times a run rate that is no longer compounding at the rate that justified the May round’s reputation. [7] That is a harder critique than “this equals 78 staples companies,” and it is not the one Burry wrote down.

Bottom line for anyone using his frame: treat $965 billion / 78 as a rhetorical equal-weighted tail comparison, internally consistent with ~$10–13 billion names, and treat 26x / 2.4x / 8.6% as a 1999 UPS benchmark that is roughly right on sales and inflation if you start from the ~$60 billion offer-price equity value. Do not treat it as an Anthropic multiple, a growth model, or a statement about the S&P 500’s own valuation. On publicly estimated current run-rate math, the May mark is about 15 times sales and a $2 trillion IPO hope is about 30 times — rich versus the index and versus median software, ordinary versus the dearest AI and security names, and still untested against $518 billion of fixed compute.

---

Recent Findings Supplement (October 2026)

Michael Burry (as @michaeljburry / Cassandra Unchained on X, early October 2026) highlighted Anthropic’s ~$965 billion May 2026 private valuation as equivalent to the aggregate market value of 78 profitable S&P 500 companies, framing it as a bubble indicator via a “fun game” of comparing private valuations to established public firms.[1][2]

He listed examples including Domino’s, Clorox, J.M. Smucker, Stanley Black & Decker, Deckers, Lululemon, McCormick, Tractor Supply, NVR, Albemarle, News Corp, Alliant Energy, Kimco Realty, Hormel, Weyerhaeuser, DaVita, Zimmer Biomet, Lennox, Masco, A.O. Smith, MGM, Wynn, Brown-Forman, Norwegian Cruise Line, and Huntington Ingalls (among others). He benchmarked against the largest inflation-adjusted pre-IPO valuation from 1990–2000: UPS at $119 billion (company age 92; 26x earnings at ~8.6% net margin; 2.4x sales). Anthropic (founded 2021) received no explicit revenue, growth, or per-company multiple assumptions in the post—only the aggregate valuation comparison.[3][4]

This October 2026 commentary is the primary recent development matching the query; no earlier 2026 Burry statements on Anthropic pricing appear in results.

Anthropic’s latest reported/estimated financials (mid-to-late 2026 updates shared with investors) show rapid but decelerating revenue growth amid high infrastructure commitments and a shift toward adjusted profitability.[5]

• 2025 full-year revenue: ~$4.6 billion (roughly 12x growth from ~$400 million in 2024); operating loss ~$8 billion; net loss ~$42 billion (majority from accounting charges).[6]
• Q2 2026 revenue: ~$11.5–11.6 billion. Late July 2026 annualized run rate: >$65 billion.[7]
• Q3 2026 (Jul–Sep): $16.8 billion revenue (up ~45–46% QoQ); adjusted operating margin ~11%; positive cash flow; high gross margins (reported ~85% in one period). Run-rate estimates now ~$67–74 billion annualized (some investor targets previously >$100 billion by year-end now seen as short).[8][9]
• IPO context: Confidential filing; targeting November 2026 listing (post-midterms) at $1.8–2 trillion valuation (vs. $965 billion May private round). Compute/infrastructure commitments: $518 billion over coming years (large non-cancellable portion to cloud providers).[1]

These figures are publicly estimated/reported via investor briefings and media (e.g., Reuters, FT).

S&P 500 valuation metrics (October 2026 data) show median trailing P/E around 23x and forward P/E near 19x, with variation by source and equal-weighted vs. cap-weighted views.[10]

• Recent snapshots: forward P/E ~19.1–19.2x (down from earlier 2026 levels); median constituent trailing P/E ~23.2x; tech sector forward multiples compressed.[11]
• Broader context includes elevated long-term metrics (e.g., P/E10) but earnings growth supporting prices amid multiple contraction.

Comparison table (Burry framing vs. independent public estimates; all Anthropic/S&P figures labeled as publicly estimated/reported):

• Valuation basis (Burry): $965B (May 2026 private round) buys 78 profitable S&P 500 firms (examples listed above). UPS historical: $119B pre-IPO (2.4x sales, 26x earnings).[3]
• Anthropic valuation (public est.): $965B private; targeting $1.8–2T IPO (~30x recent ~$67B run-rate sales).[7]
• Anthropic revenue/growth (public est.): 2025: $4.6B; Q3 2026: $16.8B (~45% QoQ, implying ~$67B+ annualized run-rate); growth decelerating from prior quarters.[8]
• Anthropic multiples/profitability (public est.): Not directly addressed by Burry. Implied high sales multiples at IPO target; recent adjusted profitability (op. margin ~11%) but prior GAAP losses and heavy capex.[5]
• S&P 500 context (public est.): Median ~23x trailing P/E; ~19x forward P/E. Burry’s 78 firms are profitable established businesses (implied lower growth/margin stability vs. Anthropic).[10]

Burry’s framing is vulnerable on several points (my analysis of the data): It equates a high-growth, early-stage AI firm’s private valuation to the summed market caps of mature, profitable, lower-growth S&P companies without adjusting for growth differentials, margins, or risk profiles—Anthropic’s revenue trajectory (even if slowing) and adjusted profitability path contrast with the stable but slower UPS benchmark or typical listed peers.[12] However, it highlights real risks: massive non-cancellable compute commitments, customer concentration potential, AI price competition eroding margins, and a $2T IPO target implying multiples that assume sustained hyper-growth and high profitability not yet fully demonstrated in GAAP terms. S&P multiples have also compressed recently, but earnings support remains tied to broader economic factors rather than AI-specific execution. No direct rebuttal or updated Burry details post his post appear in results.

Scion’s last regulatory snapshot is the Q3 2025 13F. Everything after November 2025 is voluntary Substack and X disclosure, not SEC filings. Burry’s 2026 book is a personal barbell: a shrinking set of long-dated AI/semiconductor puts, plus value longs, with the short book consolidated in mid-October into Nasdaq-100 and semiconductor-ETF puts.

Scion’s status: deregistered, no 2026 13Fs

Scion Asset Management’s registration was terminated as of November 10, 2025. In a letter to investors dated October 27, 2025, Burry said he would liquidate the funds and return capital (aside from a small audit/tax holdback) by year-end, writing that his estimation of value “is not now, and has not been for some time, in sync with the markets.” Reuters reported the fund had about $155 million across four accounts as of late March 2025 on its most recent Form ADV. He posted the terminated status on X and pointed to “much better things” on November 25. He then launched the paid Substack Cassandra Unchained (around early November 2025) as his public channel.

There is no Scion 13F for Q4 2025 or any quarter of 2026. Trackers that still show a “latest 13F” are recycling the November 3, 2025 filing. Post-deregistration positions are what he chooses to say on Substack (often paywalled, same-day) and on X (@michaeljburry / Cassandra Unchained). Those statements are not audited, not sized in dollars in public summaries, and can change without a filing lag—or without any public update at all.

Separately, on September 15, 2026, Burry announced on X that he had taken a senior advisor role at Minerva Investment Management, working with Lakshmi Ganapathi and Unicus Research on a short-biased fund expected to launch soon. Reuters confirmed the advisory role. He is not the portfolio manager of that fund, and his Substack trade notes describe his own book, not Minerva’s.

Baseline: the final 13F (period ended September 30, 2025)

Filed November 3, 2025, the last 13F reported about $1.381 billion of holdings. Options are reported at the market value of the underlying shares, not the premium paid, so the headline notional is not capital at risk.

Position	Type	Reported value	Underlying shares
Palantir	Put	~$912.1M	5,000,000
Nvidia	Put	~$186.6M	1,000,000
Pfizer	Call	~$152.9M	6,000,000
Halliburton	Call	~$61.5M	2,500,000
Molina Healthcare	Common	~$23.9M	125,000
Lululemon	Common	~$17.8M	100,000
SLM Corp	Common	~$13.3M	480,054
Bruker 6.375% preferred	Preferred	~$13.1M	48,334

PLTR and NVDA puts together were roughly 80% of reported value. Versus Q2 2025, the fund had exited a slate of calls (UnitedHealth, Regeneron, Meta, Estée Lauder, JD, Alibaba, ASML, and others) and replaced them with the two large AI puts plus the Pfizer and Halliburton calls and a small equity book. Media coverage of “$1 billion short Nvidia/Palantir” routinely conflates notional with premium. That distinction matters for every later options comment as well.

Dated timeline of 2026 position changes

Dates below are the dates of Burry’s own posts (or the market close immediately before them), as quoted by CNBC, Bloomberg, Business Insider, and contemporaneous X summaries of the Substack. Secondary aggregators fill in color; where they go beyond his quoted words, that is flagged.

January 9, 2026 — Oracle disclosed. In a Substack post after the close, Burry said he owned Oracle puts and had directly shorted the stock during the prior six months. Stated reason: he did not like how Oracle was positioned or the investments it was making in a debt-funded data-center buildout—“It did not need to do what it is doing… Maybe ego.” He did not publish strike or size. Bloomberg and later Motley Fool coverage treat this as the first public Oracle disclosure of this cycle, distinct from the 2025 Nvidia and Palantir 13F puts.

May 2026 — bubble language, no new sizing. He described equities as feeling like the last months of the 1999–2000 bubble. That is commentary, not a position change.

August 2026 — adding, not yet the big roll. Secondary accounts of his Substack say he added to his largest shorts, including Nvidia puts (December 2026 and later expirations, strikes in the low $100s) and Palantir puts plus short stock around $175, and opened or added Nebius (described as opened near $212). These details come from trackers reconstructing paywalled posts, not from a filing. Treat sizes as unconfirmed.

September 9, 2026 — de-risk, do not roll the near puts. Trading post: sold the entire December 2026 Nvidia puts and December 2026 Palantir puts and did not replace or roll them, explicitly to avoid rapid time decay now that September had arrived. Trimmed every long and every short. Kept 2027 puts on Palantir and QQQ. Raised cash he was “happy to sit on.” “This will be an interesting market this fall.”

Short book order he gave that day, after the trim: Oracle, Palantir, Nebius, Nvidia, SOXX, Micron, Caterpillar, CoreWeave. Largest longs, order unchanged: Lululemon, Molina, MercadoLibre, then Temple & Webster, Zoetis, Sprouts, JD, Birkenstock, Adobe, HCA, Fiserv, Flutter, Build-A-Bear, Veeva, Fannie common, Freddie common, PayPal. This is his own positioning list, widely republished the same day. Rationale was mechanical (theta) plus a portfolio-wide risk cut, not a change of thesis.

September 15, 2026 — advisory role, not a trade. Senior advisor to Minerva/Unicus short-biased fund. He said he would continue posting his own trades.

September 28, 2026 — shorts converted to puts; timeline pulled forward. “Fundamentally, I am moving timelines up… I want more leverage in my short positions… Nothing says leverage like options… put options, which are relatively cheap due to exceptionally tight volatility.” He said he was “more confident than ever” the thesis would play out over the next year, versus an earlier base case of 2028, and that “the bubble in AI may burst sooner than later.” Most of the move was research-driven (he cited an Ares Management note on AI capex depending on unproven revenue); some was tax.

Quoted conversions (CNBC, Business Insider):

• Micron short → June puts, strike around $500
• Nebius short → June puts, double-digit strikes
• SOXX short → September 2027 puts, low $400s
• Palantir short plus existing puts → enlarged September 2027 puts, low $100s
• Nvidia short → September 2027 puts, mid-$100s (Business Insider; stock cited near $229 that day)
• Oracle, Caterpillar, and Nasdaq-100 shorts also swapped into puts expiring between June and December 2027 (Oracle later described with a December 2027 expiry in some accounts)
• CoreWeave short covered and not replaced; he said he had not found puts at a price he liked
• New MetLife puts (strikes in the $70s in one secondary account)

He wrote that as of that day, puts had completely replaced his short-stock positions. Rationale for the structure: a shorter expected window makes paying for leverage acceptable, and compressed VIX made the puts cheap relative to the move he expected.

October 1, 2026 — Nvidia accounting, not a new trade. “Don’t Believe Your Lyin’ Eyes, GPU Depreciation & Useful Lives.” He argued Nvidia’s depreciation math and residual values rest on modeled rental income rather than actual resale prices. This is the public rationale that has stuck to the Nvidia leg all year; it is not itself a position disclosure.

October 8, 2026 — financing thesis, still not a trade ticket. “Short Thoughts: The AI Cloud Oligopoly, Stages of Grief & a Flashing Bear Market Indicator.” Argument: hyperscalers and would-be oligopoly entrants (including Oracle, OpenAI, Anthropic) are spending as if survival depended on it, financed through increasingly leveraged structures (private equity, private credit, insurers). Market breadth and structure, in his view, already flash late-cycle. He frames the tape as still in a denial phase.

October 11, 2026 — consolidation into two index puts, plus the GSE switch. “Consolidating the Big Short” / trading post:

• Rolled most options exposure into two vehicles. Nasdaq-100 puts: April 2027 expiry, roughly 20% out of the money. VanEck Semiconductor ETF (SMH) puts: June 2027 expiry, roughly 20% out of the money. “At the moment, almost my entire short position is expressed through SMH and NDX puts.”
• Significant market top expected within zero to nine months; the final stretch of the rally “could get wild.” Low volatility is why the options are inexpensive. He expects the puts to gain before the underlyings fall a full 20%.
• Still holds individual puts on Oracle, Palantir, MetLife, and Caterpillar as side bets. Other single-name shorts were closed into the index puts.
• Palantir stock short: half covered in June, the rest closed when he moved fully into September 2027 puts. Shares had doubled since June to a record high. He does not rule out another sharp rally, but says falling volume on the rally is exhaustion.
• SMH rationale, in his words: Nvidia 19.6%, TSMC 9.3%, AMD 5.6%, with Broadcom, Micron, Marvell, and Applied Materials each near 5%—“very much in my wheelhouse of individual names I would like to be short.”
• Fannie Mae and Freddie Mac: sold all common; bought several series of junior preferreds; raised total “Twins” exposure to a “very full” position, about 8–10% of the portfolio across five series. Also realized the tax loss on the common. Rationale: both stocks had given back essentially all post-election gains and printed a large topping pattern; Wall Street’s base case (he cites KBW) is shifting toward Treasury converting its senior preferred liquidation preference—he puts that claim near $388 billion, about twice combined book—into common, which he says dilutes common holders at least 90% and grows every quarter. Junior preferred ranks above common, so dividends cannot reach common until preferred dividends are reinstated. He still expects conservatorship to end within about two years, before warrants expire, and has described potential upside of roughly four to five times from then-current preferred prices if senior claims are resolved and dividends return.
• Other same-day notes: added to BYD; wrote November 2026 calls on the full Veeva position, roughly at the money; still not back in Alibaba.

Later secondary write-ups (Motley Fool, October 14) say he remains in Nvidia, Palantir, and Oracle puts with 2027 expirations (Nvidia rolled to September 2027; Palantir September 2027; Oracle December 2027) and that Oracle has been the one of the three that has worked, with the stock down sharply, while Nvidia and Palantir have not. That is consistent with residual single-name puts sitting alongside the SMH/NDX consolidation, not with a full exit from Nvidia exposure—SMH itself is a large Nvidia proxy.

What the book looks like as of mid-October 2026

Confirmed in his October 11 words: almost the entire short is NDX April 2027 puts and SMH June 2027 puts, both about 20% out of the money, sized for a top inside nine months. Residual puts he explicitly said he still holds: Oracle, Palantir, MetLife, Caterpillar. GSE exposure is junior preferreds, not common, at 8–10% of the portfolio.

Still described as open in September lists and not contradicted on October 11: a value long book whose largest names have included Lululemon, Molina, and MercadoLibre, plus a broader set (Sprouts, JD, Birkenstock, Adobe, Build-A-Bear, Veeva, PayPal, and others). Aggregators also list Hong Kong names (JD, Tencent, Meituan, Haidilao, and an October BYD add). The BYD add is in an X summary of the October 11 post; some China holdings in tracker tables are not restated in the widely quoted portion of that post, so treat the full HK roster as lower confidence than the index puts and the GSE switch.

Closed or replaced, per his posts: December 2026 NVDA and PLTR puts (September 9, not rolled); outright shorts in Micron, Nebius, SOXX, Nvidia, Palantir, Oracle, Caterpillar, and Nasdaq-100 (September 28, replaced with puts); CoreWeave short (September 28, not replaced); most remaining single-name option shorts (October 11, rolled into NDX and SMH); Fannie and Freddie common (October 11). Tesla and Applied Materials shorts were reportedly covered for gains in early September; that detail is tracker-level, not in the major wire quotes.

Filings vs. his posts vs. rumors, and the lags

Confirmed regulatory record. Only the Q3 2025 13F (filed November 3, 2025) and the November 10, 2025 deregistration. 13Fs lag up to 45 days, omit shorts and most non-U.S. holdings, and mark options at underlying notional. There is nothing to “update” from the SEC in 2026.

Confirmed by Burry. The January Oracle disclosure, the September 9 risk cut, the September 15 advisory announcement, the September 28 short-to-put conversion and timeline change, and the October 11 consolidation and GSE rotation. These are his sentences, republished by CNBC, Bloomberg, Business Insider, Reuters, and his own X account. Strike ranges and expirations in those articles are quotations from the Substack, not independent position checks.

Secondary reconstruction. Position-size rankings, exact entry prices ($175 Palantir, ~$152 Oracle, Nebius near $212), and the long list of value adds (QXO, Ero Copper, Sprouts adds, Flutter doubles) come from Substack trackers that say they read the paywalled posts. Useful as a map, not as a filing. A few tracker lines explicitly note that a reported sale was never confirmed in his words.

Not his book. Minerva/Unicus recommendations (Carvana, Faraday Future, Snowflake, Moderna, Robinhood, and others) are that firm’s research history. His advisor seat does not mean those are Scion or personal positions.

How to read a “Burry sold X” headline. After November 2025 there is no 45-day lag and no completeness requirement. A headline can be same-day and still omit size, remaining exposure (a covered short can sit next to a larger put), or a roll that happened two weeks later. The September 9 “sold all Nvidia and Palantir puts” headlines were true of the December 2026 contracts and false as a description of the bearish book, which still held 2027 puts and the stock shorts. The October headlines that he “consolidated into two ETFs” are his framing, with four named single-name puts left on as side bets and SMH itself a concentrated Nvidia vehicle.

What the sequence implies

The mechanism is consistent even as the instruments change. He is not trying to be short every AI name at once with borrowed stock. He wants convexity into a window he has compressed from “2028 base case” (August) to “over the next year / by next summer” (September 28) to “a significant top within zero to nine months” (October 11). When near-dated puts bled theta (September 9), he cut them rather than roll. When the timeline shortened and VIX was low (September 28), he replaced stock shorts with puts because the option was the cheaper way to get leverage. When he decided the expression should be the complex rather than a dozen single names (October 11), he collapsed most of that into NDX and SMH, 20% out of the money, and kept only the names he still wants idiosyncratic exposure to—Oracle (debt-funded buildout), Palantir (volume divergence and earlier accounting critiques), MetLife (insurers in the AI financing web), and Caterpillar (data-center construction cycle).

The GSE trade is the clearest non-AI example of the same habit: he did not abandon the thesis (end of conservatorship), he changed the security once dilution math on the common looked worse than the preferred. Tax-loss harvesting is acknowledged as a partial motive on both the September options roll and the Fannie/Freddie common sale; he presents research, not taxes, as the main driver.

For anyone treating this as a signal, the binding constraints are structural. There is no 13F to confirm size, cost, or whether a disclosed put is 1% or 15% of the book. Notional language from 2025 does not carry forward. The advisory role is not a second disclosed portfolio. And the October 11 book can already have been trimmed again—the only update mechanism is the next Substack post.

---

Recent Findings Supplement (October 2026)

Scion Asset Management deregistered as an investment adviser in November 2025 (terminated November 10, 2025), ending its SEC registration and 13F filing obligations.[1][2][3] Burry now manages his personal capital and discloses positions and trades primarily via his paid Substack newsletter ("Cassandra Unchained") and X posts, rather than regulatory filings. His last 13F (Q3 2025, as of September 30, 2025, filed November 3, 2025) showed a ~$1.38 billion portfolio heavily concentrated in derivatives.[4][5]

No 13F filings exist after the Q3 2025 report due to deregistration. Trackers and news confirm Scion no longer files quarterly holdings reports.[3][3]

Final confirmed 13F positions (Q3 2025, filed Nov. 3, 2025):[4][5]
- Palantir Technologies (PLTR) put options: ~$912.1 million (largest position, ~66% of reported value).
- Nvidia (NVDA) put options: ~$186.6 million (~13.5%).
- Pfizer (PFE) call options: ~$152.9 million.
- Halliburton (HAL) call options: ~$61.5 million.
- Molina Healthcare (MOH) common: ~$23.9 million.
- Lululemon (LULU) common: ~$17.8 million.
- SLM Corp (SLM) common: ~$13.3 million.
- Bruker Corp 6.375% preferred: ~$13.1 million.

This filing reflected a bearish stance on AI names via large PLTR and NVDA puts (notional value of underlying shares per 13F rules; actual premiums paid are lower). It marked the fund’s final regulatory snapshot before deregistration.[5]

Post-deregistration position updates come exclusively from Burry’s Substack and X disclosures (confirmed, not rumors), with no material reporting lags beyond his posting cadence. These show evolution from the final 13F, including consolidation of bearish bets, tax-loss harvesting trades in October 2026, and additions to longs.[3][6]

Dated timeline of key changes and rationale (2026 Substack/X updates; all post-April 2026):

• September 2026 (and earlier 2026 moves referenced in later posts): Burry converted outright shorts on AI-related names into put options and shifted some positions (e.g., Micron, Nebius, SOXX/iShares Semiconductor ETF, Palantir) into puts with June or September 2027 expirations. He sold all December 2026-dated NVDA and PLTR puts (citing time decay) while maintaining or adding longer-dated exposure. Rationale: Increasing confidence in an AI bubble burst (initial base case shifted earlier, toward 2027); seeking greater leverage via options while trimming risk.[7][7]

• Early October 2026 (around Oct. 5 post): Portfolio reshuffle amid tax-loss harvesting season. Sold LULU common shares and bought Deckers (DECK) as a proxy (plan to potentially swap back after 30-day wash-sale period). Sold FNMA common and shifted exposure into FMCC junior preferred shares (increasing overall Fannie/Freddie allocation to a “very full” 8-10% position). Replaced SFM and ZTS common with long-dated LEAP calls. Added to BYD common shares. Wrote November 2026 calls on the full VEEV position (roughly at-the-money). Bought more MET 2029 puts and JD 2029 calls. Rationale: Tax-loss harvesting opportunities in out-of-favor names; maintaining or adjusting exposure without net reduction in key theses (e.g., mortgage giants pending conservatorship resolution).[8][7]

• October 11–12, 2026 (“Consolidating the Big Short” post and related updates): Consolidated nearly all bearish exposure into NDX (Nasdaq-100) puts expiring April 2027 (~20% OTM) and SMH (VanEck Semiconductor ETF) puts expiring June 2027 (~20% OTM). “Almost my entire short position is now expressed through” these two. Retained select individual puts (e.g., PLTR Sept. 2027, ORCL Dec. 2027, MET, CAT). Closed remaining Palantir share short (half covered in June; rest converted to puts). Noted Palantir’s rally (doubled since June) but cited declining volume as a sign of exhaustion. Added that he sees a “significant market top within zero to nine months” and that “this last few months of the bull could be wild.” SMH described as a concentrated proxy for overextended semis (NVDA ~19.6%, TSM ~9.3%, AMD ~5.6%, others near 5%). Rationale: Efficiency/leverage on AI/semiconductor bear thesis amid unsustainable capex and data-center spending; moving timeline up due to higher confidence in a 2027 reckoning (comparisons to 2000/2008 denial phases). Long book includes MELI (~12%), DECK, MOH, TPW, and various others (JD, BIRK, ADBE, HCA, FISV, BBW, VEEV, PYPL, QXO, ERO, BYDDF, FMCC, etc.), plus select calls.[9][6][10]

No confirmed new additions or exits to NVDA/PLTR puts beyond the described rollovers/consolidations; bearish stance on AI complex persists but is now more index/ETF-focused. Long positions emphasize value, healthcare, and select international/China exposure (e.g., repeated BYD adds; no re-entry to BABA yet).[11]

Implications for competitors/observers: Substack/X provide higher-frequency (near real-time) transparency than lagged 13Fs, but positions can change rapidly via options rolls or harvesting trades. The shift to index puts reduces single-stock granularity while amplifying macro AI-bubble exposure. Track his Substack/X directly for updates, as regulatory filings have ceased. All quantitative details (e.g., specific expirations, percentages) derive from his disclosures rather than third-party verification.

Michael Burry’s core claim is accounting, not physics: hyperscalers book Nvidia-class servers over roughly five to six years, while he argues the economic replacement cycle is closer to two to three years, so industry depreciation is understated by about $176 billion from 2026 to 2028, with Oracle’s earnings potentially overstated by roughly 27% and Meta’s by about 21% by 2028. The strongest counters do not deny that newer chips are far more efficient. They show that older chips still clear cash costs, still get multi-year contracts, and still sit inside a GAAP framework that keys off expected use by that company, not Nvidia’s product cadence.

1. Residual value and utilization: the hardest empirical counter

The cleanest attack on a two-to-three-year economic life is the market for the chips themselves. A 2020 A100 is still a contracted asset in 2026, not scrap.

CoreWeave disclosed on its Q2 2026 call an A100 contract running into 2029—nine years after the May 2020 launch—and management said prior-generation pricing is at or above levels from years earlier, with an expiring H100 batch rebooked at about 95% of the prior rate. Jensen Huang amplified that the A100 fleet is “mission-capable from 2020 through 2029,” and Nvidia CFO Colette Kress has said A100s shipped six years earlier were still running at full utilization because CUDA keeps the software stack current. Silicon Data’s residual-value estimate for the A100 has held near $5,000 (about $4,956 in one September 2026 reading), built from projected rental income rather than a fire-sale print. Marketplace and index prints in early October 2026 still show A100 hours around $1.50–$1.80 at neoclouds and higher at hyperscalers, while H100 on-demand has been firming—Ornn recorded a 22% monthly rise to about $3.28/hour in September. Nvidia’s own September investor slide, using Silicon Data, plotted A100 residual value at 25% of an assumed $20,000 cost, H100 at 58% of $35,000, and B200 above assumed cost. Bernstein’s Stacy Rasgon (November 2025) put the operating math under that: cash costs on an older GPU are far below rental prices, so even a five-year-old A100 can still earn a comfortable contribution margin (one summary of the note put A100 hourly revenue near $0.93 against cash costs near $0.28, implying roughly a 70% contribution margin), and only something as old as Volta approaches cash breakeven. Lambda’s Matt Rowe has told Bernstein the effective life can stretch to seven or eight years via warranty extensions and redeployment, with H100 utilization above 85% and public pricing unchanged for more than a year at the time of that comment.

What this means for the bear case: a chip can lose frontier-training relevance and still be a positive-NPV asset. Burry’s own reply—that rental demand does not prove the book value is right—is fair, but it shifts the argument from “these chips die in 24–36 months” to “residual value is lower than a six-year straight-line schedule implies.” That is a narrower claim, and the market data has not yet produced the collapse in old-chip rents his original framing required.

2. The waterfall, not a single obsolescence curve

Operators and independent bulls treat useful life as stacked markets, not one curve. Frontier training ages out fast—Hoppers are weak for new frontier jobs once Blackwell is plentiful—but inference, fine-tuning, batch jobs, and smaller models do not need the latest FLOPS-per-watt. Dave Friedman’s framing is the mechanism: if you price only the training curve, you get a one-to-three-year life and conclude the books are cooked; if you stack inference and long-tail workloads on top, the same silicon keeps earning after it leaves the frontier. Power and cooling reinforce that. A100 racks draw far less than liquid-cooled Blackwell racks, so legacy halls that cannot take 100 kW+ racks keep older silicon productive rather than idle. CUDA fungibility is what makes the cascade financeable: the same software stack means a customer can move a workload down a generation without a rewrite, which is why CoreWeave can still clear “attractive” prices on 2020 silicon.

For anyone shorting the depreciation gap, the test is not “does Blackwell obsolete Hopper for training?” It is “does the residual rental curve, after power and colocation, stay above cash cost for years four through six?” Current prints say yes for A100 and H100. That can reverse if inference supply floods or if power becomes the binding constraint everywhere—but it has not reversed yet.

3. Accounting defenses: estimates, not fraud

GAAP and IFRS define useful life as the period the reporting entity expects to use the asset, including residual value, not as Nvidia’s shipping cadence. Straight-line over four to six years is an industry convention, not a mandate; companies can accelerate if consumption of benefits is front-loaded. Extending a life is a change in estimate, applied prospectively, not an error that forces a restatement. Auditors have continued to sign the hyperscaler filings.

The filings also cut against a one-way earnings-management story. Meta’s move to 5.5 years for most servers and network assets, effective 2025, cut 2025 depreciation by about $2.92 billion and lifted net income by about $2.59 billion. Oracle moved servers from five to six years. Amazon went the other way: effective January 1, 2025 it shortened a subset of servers and networking from six years to five, citing the pace of AI hardware, and that change increased 2025 depreciation by about $1.4 billion. Microsoft’s range remains roughly two to six years; Alphabet has held around six. Nvidia’s seven-page memo to analysts, which named Burry, argued customers depreciate over four to six years based on observed longevity, and that two to three years is simply wrong. Ed Yardeni sided with the hyperscalers on the same grounds: older data-center fleets have historically stayed in service, and the cascade story is how management justifies the estimate. The Footnotes Analyst view is similar—different lives on similar hardware are not evidence of misreporting if expected use and residual values differ.

The limit of this defense is real. A legal estimate can still be optimistic. Amazon’s partial shortening shows managements do not all believe six years is conservative for every GPU cohort. The accounting counter defeats “fraud”; it does not prove six years is the right economic life for the newest, most expensive cohort.

4. Eisman: the case is “too academic,” and the real risk sits elsewhere

On the New Money podcast in early October 2026, Steve Eisman—the other Big Short figure—accepted the accounting logic and rejected its weight.

His words, as reported across Business Insider, Yahoo Finance, Benzinga, and IBTimes:

• “With all due respect to Michael, I think his argument is too academic.”
• “If AI succeeds because Anthropic and OpenAI grow like crazy, and the hyperscalers do well, it won’t matter whether the depreciation schedule changed from three or four years to five or six years.”
• “Where I think he is wrong, for the moment, is that there is such demand for chips right now that there is still huge demand for the older chips.”

Where he agrees: stretching lives flatters near-term earnings, and the dollar amounts Burry cites (more than $175 billion of understated depreciation in 2026–2028) are large enough to matter on a standalone P&L. Where he disagrees: current scarcity means older chips are fully utilized, so the economic write-down Burry models is not showing up in rents or contracts. The binding risk, in his view, is concentration, not depreciation. He has called OpenAI and Anthropic the “Achilles’ heel” of the AI trade and said that if OpenAI fails and “the whole chain goes in reverse,” the result is a “massive correction that has nothing to do with the depreciation schedule.” He has also partially hedged his own AI book rather than dismissing the sector. A related fragility point attributed to him—that a large share of hyperscaler AI revenue traces to those two labs—is directionally consistent with public disclosures such as Microsoft’s OpenAI-related revenue, but the precise “70% of AI revenue” figure is his estimate, not a filed number.

Eisman’s argument is strategically coherent and empirically anchored in today’s tight GPU market. It is weaker as a refutation of Burry’s multi-year accounting math, because it is conditional on AI demand continuing to absorb both new and old capacity. He is explicit that the condition can fail.

5. Lab revenue: the condition Eisman needs is, for now, being met—unevenly

Anthropic told prospective investors on October 14, 2026 that Q3 revenue was $16.8 billion, up about 45% from $11.6 billion in Q2 and from roughly $4.7–$4.8 billion in Q1, and that it was profitable on an adjusted basis (one account cites an 11% adjusted operating margin and positive free cash flow; another notes the company was cash-flow positive in the quarter). Full-year 2025 revenue was about $4.6 billion, so the 2026 ramp is an order of magnitude. It is preparing an IPO, with reported valuation talk in the $1.8–$2 trillion range. OpenAI’s picture is messier and slower. It told investors annualized revenue was approaching $50 billion at the end of September, about $20 billion below a ~$70 billion figure that had circulated; the Financial Times and Bloomberg attribute much of the gap to methodology—Anthropic books gross sales through cloud partners, OpenAI books its net share. Bloomberg has also reported an expectation of $70 billion or more in annualized revenue by year-end 2026, while recognized full-year 2026 revenue has been put nearer $35–$36 billion in investor-material summaries. OpenAI remains loss-making; one report put 2025 revenue at $13.1 billion against large operating losses, and materials summarized by the FT have projected deeply negative cumulative free cash flow through 2030 even as revenue scales. IPO timing has slipped toward 2027, with a reported raise of about $30 billion at a $1.2–$1.5 trillion valuation range discussed in October coverage.

These trajectories support the bull path Eisman describes: the labs are growing fast enough that hyperscalers can point to real AI demand, and Anthropic has crossed into adjusted profit. They do not close Burry’s depreciation math, and they sharpen his circularity concern—much of the “demand” that keeps old GPUs rented is demand from two private companies whose own capex is funded by the same ecosystem. Methodology differences mean headline run rates should not be stacked as if they were GAAP revenue.

6. Capex returns: the weakest pure counter, and partly a bear point

Sell-side work on returns does not cleanly refute Burry. Goldman Sachs (Ryan Hammond and colleagues, notes in late September and mid-October 2026) estimates the largest U.S. hyperscalers need on the order of $300 billion in annual AI revenue to break even on average 2026–27 capex, and roughly $417–$526 billion for a 10–20% return, against cloud revenue only about $70 billion above the pre-AI trend as of Q2 2026. The same work has hyperscalers on track for about $800 billion of capex in 2026, with consensus nearer $1.1 trillion in 2027 and Goldman higher still. Company guidance after the July 2026 prints clusters around Amazon ~$220 billion cash capex, Alphabet $195–$205 billion, Meta $130–$145 billion, and Microsoft on the order of $175 billion including leases for calendar 2026—roughly $745 billion for the four, before Oracle. Realized ROIC has already compressed (Goldman has cited a drop from about 40% to 25%). The bull offset is the growth rate, not the level: year-over-year cloud revenue growth for Amazon, Google, Microsoft, and Oracle accelerated from about 25% in 2024 to about 48% in Q2 2026, and managements still describe themselves as capacity-constrained. An independent tally from Exponential View put annualized AI revenue near $276 billion covering more than the cost of capital already in service, while warning that capital committed over the next 12–24 months requires infrastructure revenue to keep compounding at roughly 27–42% a year for the life of the equipment. Hyperscaler forward multiples have also de-rated to multi-year lows, so the equity market is no longer paying peak-bubble multiples for the spend.

Capex ROI is therefore a partial agreement with Burry, not a rebuttal. The schedule debate changes the timing of expense; it does not create the revenue required to earn a return on $800 billion-plus of annual spend. Bulls win this round only if cloud and lab revenue keep compounding faster than depreciation catches up. Bears win it if the $230 billion-scale gap Goldman describes persists once the current capacity comes online and starts depreciating.

Ranking by strength and evidence quality

Rank	Counterargument	Strength vs. Burry	Evidence quality	Why it ranks here
1	Old GPUs still rent, still contract, still clear cash cost	Strongest direct hit on the 2–3 year life	High: contracts, rental indexes, Bernstein unit economics, CFO utilization statements	Market prices are harder to dismiss than narrative. Does not prove six-year book value is exact.
2	Waterfall: training obsolescence ≠ economic death	Strong mechanism	Medium-high: operator testimony, power constraints, price-per-token math on A100 vs H100	Explains why residual value holds. Still a model, not a guarantee if inference oversupply arrives.
3	GAAP useful life is entity-specific; Amazon shortened a subset	Strong vs. “fraud,” moderate vs. “too long”	High on standards and filings; medium on whether 5–6 years is right	Kills the fraud framing. Leaves the estimate open to challenge, which Amazon’s own shortening illustrates.
4	Eisman: too academic while demand is scarce; labs are the real risk	Strong as a ranking of risks, weak as a multi-year accounting rebuttal	Medium: clear quotes, consistent with rental tightness; conditional on AI success	Best statement of why the P&L distortion may not be the trade. Explicitly agrees the accounting point is real.
5	Anthropic/OpenAI revenue ramps	Supports the “if AI succeeds” branch	Medium: multiple outlets on Anthropic’s $16.8B quarter; OpenAI figures conflict by ~$20B on run-rate definition	Shows demand is not fictional. Private, non-GAAP, and circular with hyperscaler capex.
6	Capex will earn its cost of capital	Weak as a counter; partly supports caution	Medium-high on the gap (Goldman), lower on the path to closing it	The $300B breakeven vs ~$70B above-trend is the number bulls have not yet beaten. Growth rates are the only live rebuttal.

Bottom line. The evidence-weighted counter to Burry is not that depreciation schedules are irrelevant. It is that the two-to-three-year economic death he uses to generate the $176 billion gap is not what rental markets, nine-year A100 contracts, or Bernstein’s cash-cost math currently show. GAAP gives companies room to book five to six years if they genuinely expect to use the assets that long, and at least one hyperscaler has shortened lives when it disagreed. Eisman’s October 2026 intervention is the clearest investor-level version of that view: the accounting is real, the scarcity of compute makes it second-order for now, and the trade breaks if OpenAI and Anthropic stop growing—not if Meta’s server life is 5.5 years instead of 3. The open flank in the bull case is returns on the capex itself. Goldman’s roughly $300 billion annual AI-revenue hurdle versus about $70 billion of cloud revenue above trend is a larger, better-documented problem than the depreciation schedule, and it is the one Burry’s broader short does not need the accounting argument to win.

---

Recent Findings Supplement (October 2026)

Steve Eisman's October 2026 rebuttal on the New Money podcast stands as the most prominent and directly targeted counter to Michael Burry's AI depreciation and overinvestment thesis. Eisman calls Burry's focus "too academic," arguing that overwhelming chip demand—including for older generations—renders changes in depreciation schedules (e.g., from 3-4 to 5-6 years) immaterial if Anthropic and OpenAI scale rapidly and hyperscalers thrive. He agrees the accounting concern has some validity but views it as secondary. The real vulnerability, per Eisman, is extreme concentration: OpenAI and Anthropic reportedly account for ~70% of AI-related revenue at Microsoft, Amazon, Alphabet, and Oracle. A failure there triggers a "massive correction" unrelated to depreciation. Eisman has partially hedged his own AI positions while acknowledging he has participated in the trade.[1][2][3]

This directly addresses Burry's core claim (hyperscalers understating depreciation by >$175 billion across 2026-2028 by extending asset lives) while shifting emphasis to demand sustainability and customer concentration.[4]

Empirical data on older GPU useful life and utilization strongly challenges assumptions of rapid obsolescence. CoreWeave disclosed an A100 (launched 2020) rental contract extending through 2029—nine years post-launch—at attractive pricing, with CEO Michael Intrator and CFO Nitin Agrawal noting prior-generation ASPs holding at or above prior levels and renewals providing incremental returns. Occupancy for tracked A100s rose from 61% average in 2024 to 66% in 2025 and ~77-86% in 2026; rental rates stabilized around $0.92/hour after earlier declines. Nvidia's Jensen Huang highlighted this as evidence of durability via CUDA compatibility and fungibility, with residual-value estimates (e.g., Silicon Data) placing six-year-old A100s near $5,000 based on projected rental income. H100s show similar retention patterns in secondary/resale tracking.[5][6][7]

These real-world contracts and utilization trends (post-April 2026 disclosures) provide concrete evidence that economic life exceeds Burry's implied 2-3 year horizon for many workloads, particularly in power-constrained or legacy infrastructure environments.

Anthropic and OpenAI revenue trajectories demonstrate accelerating monetization that supports ongoing capex. Anthropic reported Q3 2026 revenue of $16.8 billion (up ~46% QoQ from $11.5-11.6 billion in Q2), implying an annualized run rate of ~$67 billion; it achieved adjusted profitability for a second straight quarter. Full-year 2025 revenue reached $4.6 billion (roughly 12x prior levels per its IPO prospectus). OpenAI's annualized run rate approached $50 billion by end-September 2026 (below some prior $70 billion circulating estimates, partly due to gross-vs-net partner sales accounting differences), with full-year 2026 revenue projected around $35 billion. Both show explosive growth tied to cloud partnerships, bolstering hyperscaler AI demand.[8][9][10]

This growth trajectory counters narratives of insufficient AI revenue to justify spend, though concentration risks (noted by Eisman) remain.

Hyperscaler executives and sell-side analyses highlight monetization speed and capex coverage. Amazon's Andy Jassy stated the company is "monetizing capacity as fast as we can install it" amid $200 billion 2026 capex guidance (heavily AI-focused), with AWS showing continued strong growth. Broader analyses (e.g., one investment brief) estimate annualized AI/infrastructure revenue at $276 billion covering 182% of in-service AI capital costs (depreciation + OpEx + 15% ROIC target) for deployed assets. Nvidia's investor materials and memos defend 4-6 year lives as realistic, citing retained values (A100 ~25%, H100 ~58% in their framing) beyond accelerated schedules. Meta disclosed a specific $2.9 billion depreciation reduction from extending certain server lives to 5.5 years.[11][12]

These point to returns materializing faster than bearish timelines assume for operational fleets.

Depreciation accounting defenses emphasize economic vs. book life distinctions and actual performance. Companies like Meta, Microsoft, and Alphabet extended useful lives (to 5-6 years in some cases), with executives noting capacity is often pre-sold. Realized rental economics and renewals (as with A100/H200 examples) support longer horizons than strict 2-3 year product-cycle assumptions. Resale/residual trackers show meaningful retained value (e.g., H100s 60-75% after 2-3 years in some data), and power/cooling constraints in legacy data centers sustain demand for lower-power older GPUs.[13][14]

Ranking of counterarguments by strength and evidence quality (post-April 2026 sources prioritized):
- Strongest: Eisman's demand-and-concentration argument (direct, reasoned rebuttal with qualitative agreement on risks; multiple corroborating reports).
- Very strong: A100/CoreWeave utilization/resale data (specific contracts, occupancy trends, residual estimates from earnings calls and trackers).
- Strong: Anthropic/OpenAI revenue figures (recent quarterly disclosures and prospectus data showing scale and growth).
- Moderate-to-strong: Hyperscaler monetization statements and capex-coverage analyses (executive commentary + independent modeling).
- Supporting: Accounting defenses tied to realized economics (less novel but reinforced by new data).

These elements collectively suggest Burry's thesis underweights current demand dynamics and realized asset performance, though concentration and future return validation remain acknowledged risks even among partial bulls like Eisman. Implications for competitors or entrants: Focus on diversified demand sources, proven utilization economics, and power-efficient older hardware plays could differentiate from pure new-GPU capex bets. Data remains fluid with ongoing IPO filings and earnings.

Michael Burry’s AI short is a timing-and-accounting bet, not a claim that GPUs stop working. As of mid-October 2026 he has rolled most of the exposure into puts on the Nasdaq-100 and the VanEck Semiconductor ETF (SMH), with those contracts expiring in June 2027, while still holding some single-name puts including Oracle and Palantir. He has said he sees a “significant top” within about nine months and that the last stretch of the bull run could be “wild.” The case that he is wrong is that the same pattern—directionally interesting, operationally early, and expensive to hold—has already defined much of his post-2008 record, and that observed GPU economics are longer and more demand-driven than the two-to-three-year life his depreciation math assumes.

Timing is the product, and the product has often failed

Burry’s reputation still rests on one trade: the housing short that paid in 2007–09. Since then, the public record is a sequence of crash calls that were early, reversed, or simply wrong while equities compounded.

In a September 2019 Bloomberg interview he compared index-fund flows to pre-crisis CDOs and warned that the longer the bubble ran, the worse the crash would be. Seven years later that call has failed as a reason to sell: a 24/7 Wall St. reconstruction notes that investors who stayed in broad index funds roughly tripled their money over that span while the warning stayed open. In late 2020 and early 2021 he disclosed a large Tesla put position (13F notional around $530 million on about 800,100 shares) and tweeted that the stock would implode the way the housing bubble had. Tesla kept rising for months; the stock later fell hard from its late-2021 peak, which is the classic “right eventually, ruined by the path” outcome for a short. On 31 January 2023 he tweeted one word, “Sell.” On 30 March 2023 he walked it back: “I was wrong to say sell,” after the Nasdaq-100 had entered a bull market, up more than 20% from its December 2022 low. In August 2023, filings showed roughly $1.6 billion of puts on the S&P 500 and Nasdaq-100; those positions expired as tech led a large advance. Secondary tallies of his post-2017 public calls put the miss rate well above half. The mechanism is consistent: he identifies real excess, then the excess gets larger, liquidity and earnings extend the cycle, and a dated bearish instrument dies before the narrative does.

For anyone copying the AI trade, the implication is structural. Being early on a multi-year capex cycle is not a smaller version of being right. It is a different bet, and his own history shows he has repeatedly paid the difference.

Puts cap the loss and still bleed if the top is late

He has already lived part of that cost on AI. By his own account he has been short pieces of this complex since the fall of 2025 while indexes made new highs, and he now compares the tape to the last months of 1999–2000—an admission that he does not know the top, only that he thinks it is ahead. In late September 2026 he covered outright shorts in names including Micron, Nebius, Nvidia, Palantir, Oracle and a semiconductor ETF and replaced them with puts, writing that he was pulling the timeline forward from a prior 2028 base case, that shorter timelines make leverage more acceptable, and that puts looked cheap because the VIX was compressed. Micron puts were struck around $500 with a June expiry (implying a drop of more than half from prices then near $1,000-plus). Nebius puts were double-digit strikes, also June. Other contracts ran to September 2027.

That structure limits loss to premium, which is better than a naked short, but it does not remove the cost of a continued rally. Time decay accelerates as June 2027 approaches. A further “wild” melt-up—the scenario he himself flagged on 11 October—raises the underlying, crushes implied volatility if the move is orderly, and can zero the contracts even if a crash arrives in the second half of 2027. He has already sold some nearer-dated Nvidia puts rather than roll them, which is the behavior of someone managing theta, not of someone indifferent to calendar risk. The competitive implication is simple: a thesis that needs a 40–50% drawdown inside a fixed window is a different claim from “AI capex is excessive.” The second can be true and the first can still expire worthless.

Observed GPU lives and residual demand are longer than two to three years

The load-bearing assumption in his earnings critique is that frontier GPUs have an economic life of about two to three years, so five- and six-year books understate depreciation by roughly $176 billion across 2026–2028, with Oracle’s earnings overstated by close to 27% and Meta’s by about 21% by 2028 on his figures. The operating record is pushing the other way.

CoreWeave, on its August 2026 earnings call, disclosed an Nvidia A100 contract running into 2029. The A100 launched in May 2020, so a paying customer is booking 2020 silicon into its ninth year. Management said pricing on prior-generation SKUs was at or above levels from years earlier, and that a batch of H100s coming off contract was rebooked at about 95% of the original rate. Jensen Huang amplified the point: the A100 fleet is “mission-capable from 2020 through 2029,” with CUDA keeping older architectures rentable. Bernstein’s Stacy Rasgon, in a November 2025 note, concluded that GPUs can profitably run for about six years and that hyperscaler depreciation accounting is “much more fair than bears fear.” The mechanism he cited is the gap between rental price and cash cost: even five-year-old A100s still cleared comfortable contribution margins (one Bernstein illustration put A100 hourly revenue near $0.93 against cash operating cost near $0.28). Value drops faster in year one—often 20–30%—then flattens as chips move from training to inference and other less latency-sensitive work. Lambda’s Matt Rowe has said effective lives of seven to eight years are plausible once warranties and redeployment are counted. A Wing Venture Capital review of fleet history put physical service at seven to nine years, book life at six, and cash payback at two to four—three numbers that can all be true at once. Microsoft’s CFO has said two-thirds of capex is short-lived assets, mostly GPUs, depreciated over six years, and that the capacity is already sold for the entirety of that life. Meta’s extension of certain servers to 5.5 years cut 2025 depreciation by about $2.9 billion. None of this proves residuals stay high forever. It does show that scarcity, inference demand, power and cooling constraints on new liquid-cooled racks, and a software stack that keeps old silicon useful have already stretched commercial life past the replacement cycle Burry treats as economic death.

The methodology critiques that actually bite

Accounting treatment is an estimate, not a hidden restatement. Under ASC 250, a revision to useful life is a change in estimate, applied prospectively. Prior periods are not reopened, and the change is not an error correction unless management ignored information that was already available when the original life was set. Amazon and Meta reached opposite conclusions in the same technology cycle: in early 2025 Amazon shortened a subset of servers from six years to five and cited the pace of AI development (hundreds of millions of extra expense); Meta lengthened lives and booked billions of lower expense. That divergence is what a judgment call looks like, not what a coordinated fraud looks like. Burry is right that longer lives raise reported EPS without creating cash—the cash left when the servers were bought. He is not right that GAAP will force a catch-up write-off merely because Nvidia ships a new architecture. Straight-line over six years can still mis-time the earnings (front-loaded economic decay, back-loaded book expense), which is a quality-of-earnings point, not a balance-sheet time bomb by itself.

The Cisco analogy imports the wrong balance sheet. Burry has been explicit: he is not calling Nvidia Enron; he is calling it Cisco, the picks-and-shovels supplier at the center of a supply-side glut. Cisco did lose most of its value after 2000 even though the internet was real, because orders were pulled forward and then vanished. The differences that matter now are profits, balance sheets, and multiples. Hyperscalers funding this build are among the most cash-generative companies ever built, not late-1990s CLECs financing gear with vendor loans and equity issuance. Nvidia’s own margins and backlog are orders of magnitude more real than Cisco’s 2000 order book, which was stuffed with financed demand that later cancelled. Nasdaq-100 multiples in the low-to-mid 30s are stretched versus history; they are not the 70-plus trailing multiple of the 2000 peak. The analogy can still win if orders are cancelled. It does not win just because both eras had a dominant infrastructure vendor.

A market fall does not stop the IPOs, and hoping it will is not a trade. In late September 2026 Burry wrote that, “for the benefit of humanity, the markets should tank hard and prevent the OpenAI and Anthropic IPOs,” arguing those companies would absorb and then destroy trillions of capital. That is a preference, not a mechanism. Listings can be delayed, downsized, or completed into a weaker tape if sponsors want liquidity; private capital, sovereigns, and the hyperscalers themselves have already funded the labs without a public exit. Circular financing is a real monitoring item—UBS estimated in October 2025 that the OpenAI–Nvidia arrangement could represent up to 13% of Nvidia’s then-consensus 2026 revenue, and Nvidia’s CFO has since described balance-sheet support for AI-lab demand as contributing toward roughly a quarter of next year’s business, a broader and softer claim than “a quarter of sales are fake.” Equity stakes, capped guarantees that only bite if a tenant defaults, and take-or-pay contracts are not the same instrument as 1999 vendor loans that booked revenue on gear that sat in warehouses. Steve Eisman’s rebuttal is the cleanest one-line critique: if Anthropic, OpenAI, and the hyperscalers actually grow, “it won’t matter whether the depreciation schedule changed from three or four years to five or six.” Demand either covers the accounting or it doesn’t. The schedule is a second-order variable.

When the thesis is falsified

The thesis, as currently structured, fails if the following hold through the life of the June 2027 (and September 2027) puts:

• Calendar. The Nasdaq-100 and SMH do not fall far enough, soon enough, for those puts to finish in the money. A 2028 or 2029 air-pocket that he no longer has contracts for is a narrative win and a P&L loss.
• Residuals. A100-class and H100-class capacity keeps clearing cash operating cost by a wide margin into 2028–29, and re-contracting prices do not collapse toward power-and-colocation cost. CoreWeave-style nine-year bookings, if they become a pattern rather than one disclosed deal, directly contradict a two-to-three-year economic life.
• Books. Hyperscalers do not shorten lives or take large impairments. Amazon’s 2025 shortening was a warning shot; a broad reversal toward three-year lives, with the earnings hit Burry sized near $176 billion, would confirm him. Continued six-year lives that auditors accept, backed by sold-out capacity, falsifies the “hidden write-down” path.
• Spend. There is no “season” in which AI revenue disappoints the capex that underwrote it and boards redirect capital—the Ares point he has cited. Sustained backlog conversion at Nvidia and peers, plus enterprise and lab revenue that keeps rising faster than the incremental depreciation, removes the reason for the orders to vanish the way Cisco’s did.
• Supply. Memory and accelerator supply from Micron, Samsung, SK Hynix, and custom silicon is absorbed rather than producing the classic down-cycle he is positioned for in Micron. If Chinese and new-fab capacity arrives and rental rates and memory prices hold, the cyclical half of the short is wrong even if valuations stay rich.
• Capital formation. OpenAI and Anthropic list, or continue to raise, without a crash, and the circular slice of Nvidia revenue stays a minority while end-demand grows. That outcome separates “I dislike these companies” from “the market structure cannot finance them.”

The non-obvious risk for bears is that Burry can be right about earnings quality and still be wrong about the trade. Longer depreciation flatters EPS; it does not, by itself, force a 1987-style break or a Cisco-style 80% drawdown inside eight months. The falsifier is not a blog post. It is June 2027 with old GPUs still rented, capex still funded by cash-rich buyers, and his puts expired.

---

Recent Findings Supplement (October 2026)

Michael Burry has actively adjusted and consolidated his bearish AI positions in September–October 2026, shifting from individual stock shorts to leveraged put options on indexes and ETFs while compressing his expected timeline for a bubble burst.[1][2]

In an October 11, 2026, Substack trading post, he described rolling most options positions into puts on the Nasdaq-100 (NDX) and VanEck Semiconductor ETF (SMH), citing a “significant top” within nine months and a potentially “wild” final bull-run stretch. Earlier moves (late September) replaced shorts in names like Nvidia, Palantir, Micron, Nebius, and Oracle with puts expiring mostly June–September 2027 (e.g., Micron puts struck near $500; NDX puts in the $24,000s). He cited new research (including an Ares Management report) as making him “more confident than ever” that the AI bubble could burst “sooner than later,” moving his base case forward from 2028.[3][4]

This evolution reflects both tactical leverage (cheaper volatility) and a view that the trade’s momentum may peak imminently, with remaining longer-dated Palantir/Oracle puts retained. For competitors or entrants, the signal is that even a high-conviction bear is managing duration risk aggressively rather than riding shorts through potential rallies.

Burry’s history shows repeated early macro calls that incurred carrying costs before (or without) full validation, a pattern relevant to sustained AI rallies.[5]

Seven years prior to October 2026, he warned that index funds represented a bubble whose “longer it goes on, the worse the crash will be”; markets continued higher in the interim. His AI-related shorts (initiated or expanded in 2025–2026) reportedly lagged during summer 2026 rallies in chip and AI names before later adjustments. Being early on depreciation or cyclical reversal arguments has meant ongoing premium/interest costs or mark-to-market pressure while demand and pricing remained robust.[6]

Implication: Short-side or skeptical strategies in momentum-driven sectors like AI face asymmetric timing risk; even directionally correct views can require substantial capital or hedging buffers to survive extended rallies.

Recent usage, rental, and resale data on GPUs (particularly older generations) provide concrete counter-evidence to assumptions of rapid 2–3 year economic obsolescence.[7][8]

Nvidia’s September 27, 2026, investor materials and related analyses (Ornn, Sprout, Silicon Data, Barkr) highlight that A100 chips (launched 2020) remain in production use years later, with six-year-old units retaining ~25% value or trading at rental rates ($1.02–$1.58/hour on AWS equivalents) well above operating costs (~$0.40/hour). Rental pricing for prior-generation Hopper chips has risen in select periods (e.g., 22% month-over-month in one tracker), and hyperscalers report capacity sold out with demand exceeding supply. CoreWeave holds A100 contracts extending to 2029 (~9-year economic life in that case).[9]

Nvidia and CEO Jensen Huang have publicly emphasized fungibility, durability, and rentability, arguing 4–6 year (or longer) windows better reflect reality than 2–3 years. This directly challenges Burry’s view of overstated asset lives and understated depreciation. For market participants, it underscores that inference/training workloads can extend hardware utility beyond initial training-phase assumptions, supporting continued capex justification if utilization holds.

Hyperscaler accounting changes and financing interconnections (detailed in a Fall 2026 Ares Management report cited by Burry) represent the core methodological critique, but counter-data shows measurable earnings benefits from extensions and questions the immediacy of circularity risks.[10]

Burry estimates industry-wide understatement of depreciation at ~$176 billion for 2026–2028 from extending server/GPU lives to 5–6 years (vs. his 2–3 year view), with specific impacts like Oracle earnings overstated ~27% and Meta ~21% by 2028. Meta’s January 2025 extension to 5.5 years for certain servers cut depreciation expense by ~$2.9 billion in one year; similar moves by Microsoft and Alphabet added billions to operating income.[6][8]

Nvidia responded with a seven-page memo directly addressing Burry, defending longer lives with usage data. Ares mapped ~$573 billion across 26 recent financings (bonds, leases, GPU-backed loans, guarantees) linking players including Meta, Oracle, Microsoft, Amazon, Google, Nvidia, OpenAI, and Anthropic—each transaction assuming sustained capex. Burry called this “mainstream Wall Street coming around” to circular-financing concerns.[10]

However, a Gartner October 15, 2026, analysis flags broader unsustainability ($3.5 trillion potential impairment risk by 2030 if buildout outpaces the ~$3.6 trillion projected attributable AI revenue). This supports skepticism but also highlights that any reversal may unfold gradually rather than via immediate accounting shocks. Competitors should monitor actual utilization and lease/CIP accounting disclosures, as extensions have delivered verifiable (if one-time) earnings support.

Burry’s dot-com/1960s parallels and warnings on IPOs highlight methodology risks, including that market corrections need not halt listings or hype cycles.[11][12]

He has repeatedly compared current conditions to the late-1999/early-2000 run-up (chip indexes), 2000/2008 denial phases (6–9 months), and a 1968 “Money Game” passage mocking concerns over computer write-offs—framing Nvidia’s retained-value slide as history repeating. He has also warned against OpenAI/Anthropic IPOs as early as 2026, citing misaligned incentives, unsustainable valuations, and safety rhetoric as potential hype.[13]

A market decline (or even his anticipated peak) does not automatically preclude IPOs if private valuations remain supported by narratives or strategic buyers; Anthropic reportedly eyed a 2026 Nasdaq listing near $2 trillion despite broader caution. This decouples equity-market corrections from company-level financing events. For entrants, it means regulatory or sentiment shifts around listings could proceed independently of broader AI capex or valuation resets.

Burry’s thesis would be falsified by sustained or accelerating AI-related revenues and utilization that validate longer depreciation schedules, continued hyperscaler capex without impairment signals, and resolution of financing circularity without defaults or pullbacks—particularly if older GPUs maintain rental economics and new capacity monetizes faster than projected.[8]

Specific falsifiers include: hyperscaler disclosures showing AI workloads driving incremental revenue exceeding incremental depreciation/costs; resale/rental data confirming 5–6+ year economic lives across generations; Ares-style networks unwinding without triggering guarantees; and no major capex pauses or write-downs through mid-2027 (his nearer-term window). Gartner’s revenue-vs.-spend gap narrowing or reversal would also undermine the unsustainable-buildout premise.[14]

In practice, monitor quarterly utilization metrics, lease accounting footnotes, and any shifts in forward capex guidance from the big five hyperscalers—these provide the clearest empirical tests beyond narrative comparisons.