The AI De-Rating: What the Filings Actually Show
Summary
Between 2 June and 29 July 2026, the S&P 500 fell at most 4.5% from its record close and finished the period just 1.6% below it. Over the same window, the equal-weight version of the same index — which strips out the influence of the largest companies — actually rose. What is being called an "AI stock market crash" was, on the evidence, something narrower and more precise: a violent repricing of roughly twenty AI-linked companies inside a market that was otherwise rotating, not collapsing.
We went through the primary filings and market data behind twelve of the most widely repeated claims about this episode — Alphabet's negative free cash flow, Michael Burry's depreciation thesis, Meta's off-balance-sheet financing structure, the "$1.65 trillion in AI debt" figure, and the Nvidia-OpenAI circular financing web. Four checked out exactly as stated. Four were wrong, in ways that matter. The rest were partially true or already stale by the time they reached wide circulation.
The setup, in seven numbers:
S&P 500 maximum drawdown, twelve months to 31 July 2026: 4.50%, against a 1.58% gap to the record close by month end.
Equal-weight S&P 500 (RSP), 2 June peak to 29 July trough: up 2.71%, while the Nasdaq 100 fell 11.32% over the identical window — a 14-point spread.
Semiconductor index (SOX) peak-to-trough decline: 28.6%, the deepest single-sector move in the episode.
Oracle's decline from its September 2025 peak: 65.0%, the worst of any AI-complex name we tracked.
Alphabet's Q2 2026 free cash flow: negative $5.9 billion — the first negative quarter on record, against reported net income of $112.2 billion the same quarter.
Investment-grade credit spread widening across the entire episode: 6 basis points. High yield: 15 basis points. Credit did not crack.
Fund flows into US equities during the single worst week of the selloff: positive $11.8 billion, including $4.9 billion into technology funds specifically.
A stock market that loses 4.5% at its worst is not crashing. A twenty-stock cohort inside it losing a third to two-thirds of its value is a de-rating, and a de-rating deserves a different diagnosis than a crash does.
The Index Never Broke
The story told about late July was of a market-wide AI panic. The tape does not support that framing. The S&P 500's worst peak-to-trough decline across the full twelve-month window was 4.50%, and the Dow's was 2.76%. Both are ordinary pullbacks by historical standards, not disorderly declines.
The damage that did occur was concentrated with unusual precision. The Nasdaq 100 fell 11.31% from its 2 June peak to its 29 July trough — a textbook correction. The semiconductor index fell 28.6% over a slightly longer window measured from its own later peak on 22 June. Meanwhile the equal-weight S&P 500 set a record high on 28 July, in the middle of what headlines were calling a market crash.
Measured from the 2 June cap-weighted peak to the 29 July trough, the equal-weight S&P 500 returned positive 2.71% while the Nasdaq 100 returned negative 11.32%. That is a fourteen-percentage-point spread inside the same market on the same days. Corroborating the same picture: the Russell 2000 was up 18.1% for the year through 31 July against 9.4% for the S&P 500, and the share of S&P 500 constituents beating the index year to date reached 47.3% — the widest breadth since 2022. Money did not leave the market in late July. It left one trade inside it.
Where the Damage Actually Landed
Inside the AI complex, the losses were real and severely uneven, and the pattern is legible. The further a company sat from paying customers and the closer it sat to borrowed money, the harder it fell.
Oracle lost 65.0% from its September 2025 peak. It is also the group's only sub-investment-adjacent name — rated BBB- — and it spent 82.6% of its FY2026 revenue on capital expenditure while its remaining performance obligations grew from $138 billion to $638 billion in a single year, roughly 9.5 times annual revenue. The GPU landlords fell further still: IREN down 61.6%, CoreWeave down 59.1%, Super Micro down 65.0%. By contrast, Microsoft — whose AI capital expenditure is producing measurable Azure revenue growth — ended July down only 14.3% from its own, much earlier, October 2025 peak.
Several of the sharpest declines predate the period most coverage focuses on entirely. Oracle peaked in September 2025. Meta and CoreWeave both peaked in August 2025. Palantir peaked in November. The framing that this began with Alphabet's 23 July earnings compresses a months-long de-rating into a single day, and in doing so misses most of the story.
We would flag one correction to a figure still circulating widely: reporting on "Nvidia's $279 billion wipeout, the biggest in US stock market history" describes an event from September 2024 and is unrelated to this episode. Nvidia's worst single day here was a 27 July decline of roughly $250 billion in market value, the day it ceded its position as the world's most valuable company to Apple.
The Sequence the Popular Account Gets Wrong
The selloff is conventionally dated to Alphabet's 23 July earnings report. Cap-weighted indices actually peaked three weeks earlier, on 2 June. The semiconductor index peaked separately on 22 June. And the single worst trading session of the entire episode was not Alphabet's earnings day — it was 29 July, when the Federal Reserve held rates at 3.50% to 3.75% with three of twelve voting members dissenting in favor of a hike, and the 30-year Treasury yield touched 5.21%, its highest level since July 2007.
Two widely cited explanatory threads deserve correction on timing grounds specifically.
Oil is real, but mistimed. Brent crude rose from $69.56 to $105.32 in twelve trading sessions on an active US-Iran conflict, briefly crossing $100 for the first time since May. But Brent then fell 12.8% on 27 July, two full sessions before the worst equity session of the episode, on news of a ceasefire. Oil cannot be the explanation for the 29 July low; the timing runs the wrong way.
Credit markets, contrary to a lot of the framing, never actually cracked. Investment-grade option-adjusted spreads widened only 6 basis points across the entire episode and high-yield spreads by 15 basis points — both consistent with an orderly market, not a credit event. The visible stress was concentrated in single names: Nvidia's five-year credit default swap spread moved from roughly 40 basis points to a record 82 basis points on 27 July, the largest single-day jump since that contract began trading actively in November 2025. Oracle's moved from 144 to roughly 215 basis points. CoreWeave's reached approximately 855 basis points, implying close to a 50% five-year default probability. These are meaningful single-name signals. They are not evidence of a systemic credit event, which the broad indices would have shown and did not.
Nor was there any sign of capitulation. During the single worst week of the episode, technology sector funds took in $4.9 billion of net inflows and US equity funds overall took in $11.8 billion, according to Lipper fund-flow data. The VIX peaked at only 20.66 — elevated, not panicked. This was reallocation inside a market that, on the whole, kept buying.
The Profit That Was Mostly an Illusion
Alphabet's second-quarter 2026 results were reported almost everywhere as a record: $112.2 billion of net income, the largest quarterly profit the company has ever posted. Both of the following facts are also true and independently verified against the 10-Q. Operating cash flow for the quarter was $39.1 billion. Capital expenditure was $44.9 billion. Free cash flow was negative $5.9 billion — the first negative quarterly free cash flow figure Alphabet has reported since at least 2007, the earliest year for which the underlying XBRL data exists. We reconstructed all eighty intervening quarters. None of them were negative before this one.
The profit itself, however, does not survive close inspection. "Other income (expense), net" for the quarter was $98.0 billion, against $2.7 billion a year earlier. Of the $99.0 billion equity-securities gain inside that line, $98.75 billion was unrealised — a paper mark-to-market adjustment — and only $278 million, or roughly 0.3%, was actual realised gain. Strip the unrealised portion out and Alphabet's pre-tax income for the quarter was approximately $39.7 billion, not $138.8 billion, against $44.9 billion spent on capital equipment in the same three months.
The source of that gain is where the story becomes structural rather than merely a matter of accounting presentation. The mark came primarily from Alphabet's minority stake in Anthropic, whose carrying value rose from $64.1 billion to $124.3 billion over six months, with a further contribution from its SpaceX holding. Anthropic has separately committed on the order of $200 billion back to Google Cloud and Google's own TPU infrastructure. By our estimate, the Anthropic revaluation alone accounted for roughly 71% of Alphabet's reported pre-tax income for the quarter — meaning the single largest contributor to the period's headline profit was an upward revaluation of a company that is simultaneously one of Alphabet's largest customers. Circular financing in this cycle has now reached audited GAAP earnings at the largest scale we have found in the market.
How is the spending being funded? Alphabet raised approximately $101.4 billion over the preceding six months: $51.8 billion of notes and $49.6 billion of equity and mandatory convertible preferred stock, explicitly earmarked for AI-related capital expenditure. Long-term debt rose from $12 billion at the end of 2024 to $46.5 billion at the end of 2025 to $98.2 billion by June 2026. Meta's own second quarter told a related story from a different angle: revenue beat estimates, earnings per share missed, 2026 capital expenditure guidance was raised to $135 to $145 billion, and free cash flow collapsed to $784 million. Meta shares fell 7.95% the next trading day, against a Nasdaq 100 that rose 3.36% the same session.
The Depreciation Question
The most frequently cited structural bear argument, associated with Michael Burry, holds that the major cloud operators have extended the assumed useful life of their AI servers, which mechanically suppresses reported depreciation expense and flatters earnings. The filings confirm the extensions. They also contain one widely repeated fact that turns out to be wrong.
Meta extended its assumed server and network life from three years to 5.5 years. Alphabet moved from three years to six. Microsoft moved from three to six. Oracle's baseline, contrary to how it is usually described, was four years rather than three, and it moved to six. Each of these disclosed a specific earnings benefit: Meta's most recent extension alone added roughly $2.9 billion to reported operating income and $2.6 billion to net income, worth about a dollar of earnings per share.
Amazon did not follow this pattern, and reporting that groups it with the others is incorrect. Effective 1 January 2025, Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five. The company's own language, taken directly from its FY2025 10-K, is worth quoting exactly: the change was made because of "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." The effect was an increase in depreciation expense of $1.4 billion and a reduction in net income of $1.0 billion, or ten cents a share. The single largest operator of AI infrastructure in the world looked directly at the question Burry raises and moved in the opposite direction to its peers, disclosing the AI-driven obsolescence rationale by name and absorbing the earnings cost of doing so.
Burry's specific forward projection — that by 2028 Oracle will be overstating earnings by 26.9% and Meta by 20.8% — is a published forecast on his own Substack from November 2025, not a disclosure in any filing, and should be treated as exactly that: one analyst's projection, not an established fact. We rebuilt his sensitivity analysis independently from disclosed gross asset balances and depreciation figures. His Meta figure corresponds to an implied economic asset life of roughly 3.2 years; his Oracle figure to roughly 3.7 years. Both sit comfortably inside the range our own reconstruction produces. The arithmetic holds up. The contested input is not the maths — it is the underlying assumption that the true economic life of AI server hardware is somewhere between two and four years rather than the five to six years currently being assumed for accounting purposes.
Both sides of this argument are correct about different variables. Six-year-old A100 GPUs genuinely remain in productive service today, typically shifted from model training into inference workloads once newer chips take over the frontier. At the same time, A100 resale and rental values have fallen an estimated 50% to 70% over three to four years. Utilisation and market price are independent facts: AI hardware does not become useless with age, but it does become cheap, and that is a genuine problem for any business model — like the GPU-rental "neoclouds" — whose economics assume a longer useful life than the resale market is willing to price in.
The Debt You Cannot See on Any Balance Sheet
The most structurally interesting part of the current buildout is that a substantial share of it does not appear as debt on any single company's balance sheet.
Meta's Louisiana data centre campus, internally referred to as the Hyperion project, is owned by a special-purpose vehicle called Beignet Investor LLC, which issued $27.294 billion of senior secured notes at a 6.581% coupon, maturing in 2049. Blue Owl-managed funds hold 80% of the vehicle; Meta holds the remaining 20% and accounts for its position under the equity method, meaning the debt itself is not consolidated onto Meta's own balance sheet. Meta's 10-K states this plainly: the company is "not the primary beneficiary and do[es] not consolidate the variable interest entity." The commonly quoted project cost of roughly $29 billion is slightly off; Meta's own filing states approximately $27 billion.
The figure worth focusing on is one that is usually omitted from popular accounts of this structure entirely. Meta's disclosed maximum exposure to loss related to the Hyperion venture is $45.95 billion — a number that includes a residual value guarantee with an aggregate threshold of approximately $28 billion, for which no liability is currently recorded on the balance sheet. Meta's own auditor designated the non-consolidation judgment a Critical Audit Matter, the formal signal auditors use when a conclusion required unusually difficult or subjective judgment.
The figure of $1.65 trillion in AI-related debt, attributed to Nikkei Asia, has circulated widely as if it were the whole picture. It is real, and it understates the true position by roughly half. That $1.65 trillion figure captures only the off-balance-sheet portion of the buildout — the SPV structures, lease guarantees and similar arrangements. It sits on top of approximately $1.35 trillion of separately reported, on-balance-sheet obligations across the same companies. Rebuilding the combined figure directly from filings puts the true total closer to $3.0 trillion.
The Circular Financing Web, Reported Accurately
The structure that unsettled credit markets on 27 July is the one in which a chip manufacturer effectively guarantees the financing its own customer uses to buy that manufacturer's chips. The amounts involved are genuinely large. Several of the most widely repeated figures describing them are also, at the time of writing, not yet signed.
Nvidia was reported by the Wall Street Journal on 27 July, and separately corroborated by Bloomberg and Reuters, to be in talks for a roughly $250 billion lease-financing guarantee tied to OpenAI's data centre lease in Ohio, alongside a further approximately $350 billion of chip-procurement financing — a combined exposure that would approach $600 billion against Nvidia's own FY2026 revenue of exactly $215.9 billion. All of this, as reported, is negotiation rather than signed commitment. An earlier $100 billion figure was a non-binding letter of intent that Nvidia's own chief executive has since publicly described as "never a commitment."
OpenAI's much-cited $1.4 trillion of infrastructure commitments against roughly $25 billion of annual recurring revenue is also stale: that figure was superseded by a reset to approximately $600 billion by 2030, announced in February 2026, and OpenAI's actual annual recurring revenue run-rate is closer to $20 billion than $25 billion.
Set against that, the Alphabet-Anthropic loop described above is fully verified and, in our view, the least-covered part of the entire structure: Alphabet's equity stake in Anthropic was marked up by tens of billions of dollars in a single quarter, becoming the majority of Alphabet's reported profit, while Anthropic simultaneously commits roughly $200 billion back to Google's own cloud and chip infrastructure. That loop is closed, audited, and already inside reported earnings — which makes it more consequential than the larger, unsigned Nvidia-OpenAI figures that receive most of the attention.
The Bull Case's Load-Bearing Assumption Just Broke
Every optimistic version of the AI capital expenditure story rests on a single mechanism: that the cost of running inference falls fast enough, for long enough, that demand and eventual margin expansion arrive before the capital comes due. It is the assumption the entire buildout leans on. The evidence available in mid-2026 suggests it stopped holding at the frontier some time in the past year.
There are now two separate token-pricing curves, not one. At the commodity tier, prices continue to fall sharply — one major provider cut its lowest tier by 80% in a single move on 30 July 2026. At the frontier tier, pricing has reversed. The lowest frontier input price available in August 2025 was $1.25 per million tokens; the comparable frontier tier roughly a year later prices at $5.00 per million tokens, a fourfold increase. One major provider has an already-scheduled price increase on its mid-tier model effective 1 September 2026. Layer on top of that the fact that agentic AI tasks now consume anywhere from 100,000 to over a million tokens per completed task, at an estimated cost of four to five dollars per frontier-tier task, and the honest conclusion is that cost per completed task is not clearly falling at all — which is the specific metric the entire bull case depends on, not cost per token in isolation.
A widely repeated figure describing a "26 times" gap between required and actual AI revenue is also a comparison error: it measures a 2030 revenue requirement against a single company's 2025 revenue figure. Using de-duplicated, cross-provider AI revenue for mid-2026, which we estimate at $150 to $250 billion, the real gap is closer to three to thirteen times — still very large, but not the number in wide circulation. Worth noting separately: J.P. Morgan's frequently cited figure of $650 billion in required annual AI revenue is, on inspection, the mildest of the serious bear estimates rather than the most alarming one. It represents only 13% of the roughly $5 trillion asset base against which it is measured; a standard cost-of-capital recovery calculation at a 10% hurdle rate and a six-year asset life implies a figure closer to 24%.
To reach the revenue levels required to earn a 10% return on current capital expenditure, our reconstruction implies AI-specific revenue of $940 billion to $1.88 trillion by 2028, or $1.66 trillion to $3.32 trillion by 2030 — compound annual growth rates in the 60% to 145% range depending on the year and the assumption set. For context, the fastest any company has ever grown from $10 billion to $100 billion in annual revenue is roughly a 39% compound annual rate, achieved by both Meta and Tesla over seven years. Nothing in the historical record has grown at anywhere close to the pace this buildout now requires.
The strongest honest version of the bull rebuttal is this: the comparison above prices the buildout against the roughly $900 billion global software market, when the more relevant comparison is against the roughly $50 trillion global wage bill, against which $1.7 trillion is only 3.3%. That is not an absurd claim. It is, at present, simply unevidenced — and the bulls do hold real evidence of their own. Approximately $1.69 trillion of contracted, disclosed backlog sits across Alphabet, Microsoft and Amazon combined. AWS just posted its fastest growth rate in eighteen quarters. Available power capacity is acting as a hard physical constraint that limits how large the buildout can get — which cuts against the more extreme bear scenarios that assume unconstrained overbuilding. And Meta's own advertising business shows a roughly 12% increase in price per ad that management attributes to AI-driven targeting improvements — value that is being captured invisibly inside an existing business line, which is a genuine weakness in every version of the revenue-gap calculation above, since all of them count only explicitly labelled AI revenue.
The railway and fibre-optic analogies that bulls sometimes invoke cut in the opposite direction from how they are usually deployed. In both historical cases, the infrastructure that got built ultimately proved essential and society captured enormous value from it — and the investors who financed the original buildout were largely wiped out. Roughly a third of all authorised railway mileage in nineteenth-century Britain was never built at all, and American long-haul fibre capacity later sold for pennies on the dollar relative to its construction cost. "The technology is real" and "the equity that financed it is money-good" are two entirely separate claims, and history does not guarantee that the second follows from the first. It took electricity roughly forty years to show up clearly in national productivity statistics. A thesis can be completely correct about the technology and still be badly early relative to the five- or six-year accounting life of the asset that was bought to build it.
Twelve Claims, Checked Against the Record
We treated the popular account of this episode as a set of falsifiable claims and checked each one against a primary source.
Alphabet's first negative free cash flow on record, at negative $5.9 billion, is fully verified: every underlying figure matches the 10-Q within rounding error, and our own reconstruction of eighty prior quarters found no other negative reading going back to the earliest available data.
The claim that major cloud operators extended server useful lives to flatter earnings is largely verified for Meta, Alphabet, Microsoft and Oracle, each with a specific, disclosed dollar impact — though Oracle's starting point was four years, not the three years often cited.
The claim that Amazon extended its useful lives alongside its peers is not correct. Amazon shortened a subset of asset lives in 2025, explicitly citing AI-driven technological obsolescence, at a disclosed $1.0 billion cost to net income.
Michael Burry's specific 2028 overstatement projections for Oracle and Meta are best understood as a forecast, not a filed fact — arithmetically sound, resting on a contestable assumption about hardware economic life.
The claim that Meta financed a data centre through a special-purpose vehicle that borrowed $27.3 billion is verified, with the project cost corrected from the commonly cited $29 billion to the filed figure of approximately $27 billion, and with the more important figure — Meta's $45.95 billion maximum exposure to loss — largely missing from popular coverage entirely.
The claim that the industry carries $1.65 trillion of AI-related debt is real but understated: that figure captures only the off-balance-sheet portion, sitting on top of roughly $1.35 trillion reported separately, for a combined total closer to $3.0 trillion.
The claim that Nvidia has committed roughly $600 billion of financing to OpenAI is accurately reported as being in active negotiation, not yet signed.
The claim that OpenAI carries $1.4 trillion in commitments against $25 billion of revenue is stale, superseded by a reset to roughly $600 billion by 2030 announced in February 2026, with actual revenue closer to $20 billion.
The claim that Nvidia's credit default swap spread recorded its largest jump on record as credit investors grew concerned is verified, moving from roughly 40 to a record 82 basis points on 27 July.
The claim that $100 oil helped drive the selloff is mistimed: oil did cross $100, but it then fell sharply two full sessions before the actual worst trading day, which was driven by the Federal Reserve's decision.
The claim that the US stock market was crashing is not supported by the index-level data: the S&P 500's worst drawdown was 4.5%, and the equal-weight index rose through the worst of the episode.
And the claim that falling token costs will eventually justify the buildout is now in serious doubt at the frontier tier specifically, where pricing has risen roughly fourfold from its 2025 low, even as commodity-tier pricing continues to fall.
What Would Actually Settle This, and By When
Almost everything written about the AI capital expenditure cycle is currently unfalsifiable, and it should not be. Two figures would resolve most of the disagreement, and neither has ever been disclosed under audit: the true gross margin on inference at scale, net of all subsidised pricing, and the real cost per completed agentic task rather than cost per token in isolation.
Several dated events over the coming months will move this materially. Nvidia's next quarterly results, due in late August 2026, will be checked directly against its own $91 billion guidance. Anthropic's reported registration filing, expected sometime in September or October 2026, is likely the single most informative event on the calendar — it would be the first time a major private AI company's revenue claims are examined by outside auditors and the Securities and Exchange Commission rather than presented on the company's own terms. Michael Burry's 2028 depreciation projections will either be borne out or refuted by that year's filings. And a specific, dated prediction from short-seller Jim Chanos regarding AI-related equity issuance resolves on 31 December 2026.
The strongest version of the bear case, in our view, does not require any of the contested or disputed figures above. Alphabet's disclosed negative free cash flow, Amazon's disclosed decision to shorten rather than extend its asset lives, Meta's disclosed $45.95 billion maximum exposure to its own data centre financing structure, and Oracle's disclosed obligations at 9.5 times annual revenue are each, individually, audited facts already sitting in public filings. None of them require believing an unverified number to be concerning.
This article is for educational purposes only and does not constitute financial advice. All figures are drawn from SEC filings, FRED, EIA, FactSet, and named news sources as of 3 August 2026, and are cited as verified, partially verified, or reported/unsigned throughout. The Financial View may hold positions in the securities discussed.
