Is the AI Bubble About to Pop?
Will the gap between AI infrastructure spending and AI revenue trigger a major market correction (>20% decline in AI-weighted indices) by end of 2027?
Your retirement portfolio probably holds NVIDIA, Microsoft, and Google. This is what they're betting on — and what they need to deliver.
Your Prediction
Where do you think this lands?
Join others who've weighed in
Scenarios
Current value: 2025 CapEx: $443B actual (+73% YoY). 2026 projected: $602B. OpenAI: $20B ARR (up from $2B). CapEx-to-revenue ratio narrowing.
S-curve position: CapEx in steep buildout phase. Revenue trailing significantly. Classic early-stage infrastructure overshoot pattern.
30%+ crash — dot-com style correction when Q4 2026 earnings disappoint
10-15% pullback — a 'healthy correction' but not a crash, as underlying technology value is real
No correction — AI revenue catches up, CapEx spend justified within 3 years
How We'll Know
- What we measure
- Whether AI-weighted stock indices (NASDAQ AI Index, S&P AI Select) experience a sustained >20% decline from peak, with AI investment concerns cited as primary driver
- Confirmed if
- AI-weighted indices decline >20% from peak with AI spending/returns gap cited as primary catalyst
- Refuted if
- AI-weighted indices remain within 15% of peak through 2027, OR AI revenue growth closes the gap with CapEx
- Data sources
- NASDAQ AI Index
- S&P 500 AI-weighted components
- Goldman Sachs AI infrastructure tracker
- Hyperscaler earnings reports
- Financial Times / Bloomberg market analysis
Evidence Trail
Evidence For
- Mar 7, 2026
OpenAI expects $74B operating losses through 2028 despite $20B ARR. S&P 500 top 5 companies = 30% of index — greatest concentration since dot-com. Case-Shiller P/E exceeded 40 (first since dot-com crash). Altman himself said investors are 'overexcited' about AI. Acemoglu warns of 'circular financing' structures (CoreWeave/OpenAI stock-for-compute deals). Open source closing MMLU gap from 17.5 to 0.3pp in one year, threatening pricing power.→ Probability: 25%
- Jun 7, 2026
Broadcom's June 3 guidance miss (AI-semiconductor outlook below consensus, -14%, dragging AMD/Intel/Micron) is the first major AI-chip disappointment in eight quarters; Ray Dalio publicly called the AI bubble bursting on liquidity grounds; data-center debt issuance ($200B+ in 2025, $250-300B projected 2026) carries a one-year-GPU-depreciation vs 7-15-year-debt mismatch the Fed has flagged as a systemic risk. The bear channel is the debt structure, not equity revenue.→ Probability: 22%
- Jun 27, 2026
The closest the market came to a correction all year: a June 23-25 semiconductor selloff wiped ~$1.3T (Korea's KOSPI tripped its circuit breaker twice, Micron -13% intraday, Nasdaq -2.21%), and a cluster of Chinese hedge funds — including one whose founder called the 2007 top — declared a "super bubble." Software now trades at a discount to the S&P 500 for the first time on record (IGV -21% YTD, ~$2T erased). But it recovered within days (Micron +15% on earnings, >$1T 2026 capex reaffirmed); the bear case strengthened at the margin without crossing the 20% AI-index threshold.→ Probability: 24%
- Jul 18, 2026
W29: For the first time the index leg of the resolution criteria actually printed. The Philadelphia Semiconductor Index fell as much as 5.7% on July 17 after Moonshot released Kimi K3, pushing its drawdown past 20% from the June record — roughly $3.3T of global semiconductor market capitalisation erased since June 22, with Marvell, ARM and Intel each down more than 30% over that span. The financing channel deteriorated in parallel and on hard data: Apollo figures show hyperscaler bond cover ratios falling from 5x in February to 2x in July, Amazon's $25B eight-tranche issuance on July 7 drew the weakest demand of any hyperscaler deal this year (2.5x covered versus 3.2x in March) despite dangling extra yield, and Amazon told underwriters it would not issue further debt in 2026. Full-year AI-linked debt issuance is now projected at $570B — roughly double the $250-300B Morgan Stanley modelled six weeks earlier — with the five hyperscalers having already issued $159B through May, more than their combined issuance over the prior five years.→ Probability: 27%
- Jul 18, 2026
New tracked prediction P-049 attached: a hyperscaler depreciation-schedule change or AI-infrastructure impairment by end of 2027, priced at 0.45 — the accounting mechanism behind Burry's thesis already fired once (Amazon's 2025 useful-life shortening), and hyperscaler bond cover ratios fell from 5x to 2x between February and July.
- Aug 22, 2026
W34: Institutional voices joined the bear chorus in one week. The ECB warned in a blog post that the AI rally is set for a correction under either scenario - overoptimism unwinding, or AI succeeding but missing embedded profit expectations - quantifying ~EUR440B of Magnificent-Seven exposure held by European households, insurers and pension funds; ex-NY Fed president Bill Dudley made the same duration-mismatch argument in a Bloomberg op-ed. On hard numbers, a WSJ analysis found nine large tech firms carry roughly $3T in off-balance-sheet AI commitments (long-dated leases, GPU financing, take-or-pay) against ~$600B of trailing reported capex, and the financing channel visibly strained - CoreWeave now carries $72B liabilities against $5B equity, Nebius priced a second convertible raise in five weeks, Meta's new $12.5B data-center bond priced above its 2025 deal, and hyperscaler bond issuance hit $194B YTD (+79%).→ Probability: 29%
- Sep 12, 2026
The spend-to-revenue gap this claim hinges on was pressed again this week. Gartner puts the 2026 AI-infrastructure spending increment alone at $401B and Goldman Sachs put 2025 hyperscaler AI capex at $400-423B, against enterprise AI revenue that every independent estimate places at a fraction of either — alongside documented circularity in the financing: J.P. Morgan Asset Management research describes Nvidia's roughly $100B investment in OpenAI, which then buys Nvidia GPUs, with Harvard's Andy Wu characterising it as Nvidia effectively paying its customers to buy its products. OpenAI is separately reported to be seeking a valuation above $1T while projecting a $14B 2026 operating loss and no profitability before 2029-2030.
- Sep 18, 2026
Nscale's S-1, the first public filing by a pure-play AI datacenter operator, is the cleanest audited look inside a neocloud to date and the circularity is on the page. Against H1 2026 revenue of $140.6M it discloses $103.4B of active-and-contracted TCV as of August 31 against only $2.6B active, a net loss of $1,020.1M, a net loss margin of 726 percent, and total liabilities of $12.36B against $4.79B of equity. Its counterparties are its own suppliers and customers: Microsoft statements of work worth up to $43.8B through 2033, an Anthropic services agreement worth up to roughly $44.6B, a $1.2B NVIDIA master cloud services agreement, and a Figure AI deal under which Nscale will itself become a Figure shareholder. https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm→ Probability: 31%
Evidence Against
- Mar 7, 2026
AI revenue growing 100-200% YoY at leaders — the gap is closing. OpenAI raised $110B at $730-840B valuation — smart money still bullish. Inference costs falling 99%+ means AI becomes embedded everywhere. Unlike dot-com, AI is already delivering measurable value (coding, customer service, drug discovery). Hyperscalers can afford the CapEx from existing cash flows.
- Mar 7, 2026
OpenAI: $2B→$20B ARR in 2 years (10x). Cursor: $2B ARR with 12 people. Harvey: $195M ARR. Revenue growth is the fastest in tech history, closing the CapEx gap. Inference cost declining 200x/year makes AI viable for increasingly marginal use cases. The CapEx-to-revenue ratio is a snapshot of a rapidly converging trajectory, not a stable imbalance.→ Probability: 20%
- Apr 10, 2026
Nasdaq entered correction (~10%) in March/April 2026, but the cited drivers are tariff policy uncertainty and Iran war concerns — not the AI CapEx/revenue gap. Hyperscaler 2026 CapEx now projected at $660-690B vs ~$51B direct AI revenue, a widening gap, but markets are not pricing this as bubble-bursting. Resolution criteria requires the correction to cite AI-specific driver — currently not met.→ Probability: 20%
- Jul 18, 2026
W29: The mechanism does not match the claim. This topic asks whether the gap between AI infrastructure spending and AI revenue triggers the correction; the July 17 selloff was triggered by a competitive shock — a Chinese open-weight model topping an independent coding leaderboard — which threatens pricing power rather than showing the revenue failed to arrive. The pattern also rhymes with the June 23-25 scare that reversed inside a week. Capital formation continued straight through the drawdown: Anthropic began institutional pre-IPO investor meetings on July 15 targeting an October listing at its $965B valuation, Databricks signed a Coatue-led term sheet at $188B (up roughly 40% from December), and Fireworks AI closed a $1.5B Series D at $17.5B with ARR past $1B and daily token volume tripled to 40 trillion. Application-layer names moved opposite to the chips, with ServiceNow raising its Now Assist target to $1.5B and Workday drawing an upgrade on agentic progress — not the pattern of a demand-side bust.→ Probability: 27%
- Sep 12, 2026
The strongest counter this year landed the same week. Bloomberg reported on September 9 that analysts raised 2026 S&P 500 earnings-growth forecasts to 32% from 24% before the Q2 season, with 86% of reporters beating estimates — the highest beat rate since 2021 — and Bloomberg Intelligence attributing roughly half of that growth to AI-infrastructure beneficiaries. Earnings actually printing is the distinction between a multiple- expansion bubble and a capex cycle. Cognition points the same way at the private level: its valuation went from $26B to $48B while its revenue multiple stayed roughly flat near 53x, meaning the mark tracked revenue growth rather than outrunning it. https://www.bloomberg.com/news/articles/2026-09-09/ai-boom-set-to-push- s-p-500-earnings-growth-to-32-this-year
- Sep 18, 2026
The same filing cuts against a bubble read in three specific ways. The going-concern doubt it discloses was explicitly alleviated - management concluded its plan to defer, reduce or cancel capital expenditure is probable of implementation and would provide sufficient liquidity - so there is no unresolved qualification. Customer concentration is improving rather than worsening, from substantially all of revenue in 2024 to 73 percent in 2025 to 52 percent in H1 2026. And the contracts are take-or-pay with a 5.7-year weighted average life against counterparties including Microsoft, which is not a fragile structure.
- Oct 10, 2026
W41 hold. The Nasdaq-100 closed at a record on October 6 and Reuters reported software stocks at their 2026 high the same day, while the Philadelphia semiconductor index closed October 9 about 14% below its June record, and the Fed's September minutes record its markets desk attributing this year's equity gains entirely to actual and expected earnings, with price-to-earnings multiples down. Neither leg of the criterion - a decline of more than 20% with the capex-revenue gap named as its main cause - moved. https://www.aol.com/articles/us-software-stocks-scale-fresh-141121000.html→ Probability: 31%
How Our View Evolved
- Oct 10, 202631%→31%
W41 rotation, HOLD at 0.31. The Nasdaq-100 hit a record and software stocks their 2026 high after our last update, and the Fed's markets desk attributed this year's equity gains entirely to earnings, with price-to-earnings multiples down; neither leg of the criterion moved.
- Sep 19, 202629%↑31%
W38, +0.02 to 0.31. Nscale filed the first public S-1 by a pure-play AI datacenter operator on September 18, giving the first audited view inside a neocloud: $103.4B of contracted TCV against $2.6B active and $140.6M of H1 revenue, a 726 percent net loss margin, and a counterparty list that is simultaneously its supplier and its customer base - Microsoft, Anthropic, NVIDIA, and a Figure AI deal in which Nscale will become a shareholder in its own customer. That is genuine vendor-financing circularity evidence and it is now in a filing rather than in commentary. Held to +0.02 because the same document cuts the other way: the going-concern doubt was explicitly alleviated, customer concentration is improving from 73 to 52 percent, and the contracts are take-or-pay with a 5.7-year average life against investment-grade counterparties.
- Sep 12, 2026Initial assessment: 29%
W37 HOLD at 0.29. The week produced strong evidence in both directions and they very nearly cancel. Against the claim: S&P 500 2026 earnings-growth consensus rose to 32% from 24% with the highest beat rate since 2021, AI-infrastructure names credited with about half of it, and Cognition's revenue multiple held roughly flat near 53x across a $26B-to-$48B revaluation — both signs the numbers are tracking fundamentals. For the claim: the spend-to-revenue gap persists, with Gartner's 2026 AI-infrastructure increment alone at $401B against enterprise AI revenue every independent estimate places at a fraction of it, and documented circular financing behind part of it. No >20% drawdown in AI-weighted indices printed, so the resolution bar is untouched and the honest answer is no move rather than a number split between two good arguments.
- Aug 22, 202627%↑29%
+0.02 to 0.29. The bear case's financing channel moved up a credibility tier: the ECB (a central bank) and ex-NY Fed president Dudley both warned of a likely correction the same week the WSJ quantified ~$3T of off-balance-sheet AI commitments across nine tech firms against ~$600B trailing capex, with neocloud balance-sheet stress (CoreWeave $72B liabilities/$5B equity, Nebius' second raise in five weeks) and richer hyperscaler debt pricing. Held to +0.02 because no >20% AI-index decline printed this week and Anthropic's disclosed $65B ARR shows revenue still closing the gap the claim hinges on.
- Jul 18, 202624%↑27%
+0.03 → 0.27, the highest this topic has been. The >20% AI-weighted index decline named in the resolution criteria actually occurred: SOX fell 5.7% on July 17 on the Kimi K3 release, past a 20% drawdown from its June record, with ~$3.3T of semiconductor value erased since June 22. Financing stress is now measured rather than argued — Apollo puts hyperscaler bond cover ratios at 2x in July against 5x in February, Amazon's $25B July 7 issuance was the weakest-covered hyperscaler deal of the year at 2.5x, and 2026 AI debt issuance is tracking to $570B versus Morgan Stanley's $250-300B estimate six weeks ago. Kept to +0.03 because the trigger was a competitive shock to pricing power rather than the spending-versus-revenue gap the claim specifies, and because capital formation continued straight through it — Anthropic pre-IPO meetings at $965B, Databricks at $188B, Fireworks $1.5B at $17.5B.
- Jun 27, 202622%↑24%
+0.02 to 0.24. The June 23-25 chip selloff (~$1.3T wiped, KOSPI circuit breaker twice, Micron -13% intraday) plus Chinese hedge funds calling a "super bubble" were the closest the market came to a correction this cycle, and software trading at a discount to the S&P 500 for the first time is a real structural re-rating. But the selloff reversed within days (Micron +15%, >$1T 2026 capex reaffirmed) and never crossed the 20% AI-index threshold — a marginal bear-case strengthening, not the predicted correction.
- Jun 7, 202620%↑22%
+0.02 to 0.22. First genuine bear signals of the cycle: Broadcom's June 3 AI-chip guidance miss (-14%, dragging the semis), Ray Dalio's explicit bubble warning, and the data-center-debt mismatch the Fed has flagged. But AI-weighted infrastructure indices are still up (NVIDIA +85%, BOTZ +12%) and the SaaS rout is AI eating SaaS, not the AI trade breaking; P-012 resolves refuted June 30. Small move acknowledges the bear case strengthening at the margin, not a correction.
- Apr 10, 2026Initial assessment: 20%
Rotating refresh — HOLD at 0.20. Nasdaq ~10% correction is tariff-driven, not AI-specific. CapEx/revenue gap widened but not pricing as bubble. Resolution trigger (AI-cited driver) not met.
- Mar 8, 2026Initial assessment: 20%
Baseline — initial published assessment
What Experts Say
Yann LeCun
Chief AI Scientist, Meta (former); Founder, AMI Labs
“AI investment boom is a bubble likely to burst by 2026”
Gary Marcus
AI Researcher, NYU Professor Emeritus, AI critic
“Without world models, you cannot achieve reliability, and without reliability, profits are limited”
ARK Invest (Cathie Wood)
Investment Management, Disruptive Innovation Research
“Annual data center CapEx will grow from $500B (2025) to $1.4T (2030)”
ARK Invest (Cathie Wood)
Investment Management, Disruptive Innovation Research
“AI inference costs have fallen 99%+ in a single year, driving explosive demand growth”