Is AI a bubble? A live, evidence-backed assessment
Version 0.9.2, 2026-10-25. Latest verified update: signals, 25 October. Built by AI agents from primary sources, steered by a human, on 26 cross-checked research notes, after three Codex red-team reviews, a Claude review and a Codex trim. Rules: Methodology. Glossary at the end.
The short answer
1. Are the signals that past bubbles showed present? Several are, and the financing ones are AI-specific. Forward purchases are at records; financing is shifting toward bonds, equity, SPVs and vendor credit. The Fed hiked in September; its minutes say AI-related investment is itself "contributing to inflation pressures", alongside an energy shock (Fed, A). CAPE is 40.6 against December 1999's 44.2, a market-wide signal (theory, A). These signals do not establish overpricing.
2. Is capital being deployed faster than returns can service it? Unproven; an illustrative payback scenario sets a high bar. At a 50% owner EBITDA margin and 10% discount rate, a $750B cohort needs about $295B annual owner revenue at a six-year equipment life, $468B at three years. Applied to a 60-100% AI share of Goldman's $3.4T capex forecast, the 2026-28 cohorts need $0.8-1.3T annually once operating. At 50% owner capture, that implies $1.6-2.7T annual end-user revenue. Reaching it by 2029-30 requires 75-150% annual growth (arithmetic B; inputs and sector application D). Estimated 2026 end-user revenue is ~$170B (range $130-210B; reported lab half C, allocated rest D). Revenue roughly tripled in 2026 and exits near a $250B run-rate; our base path assumes +76% then +60%. Replacement capex, advertising, internal savings, commissioning and residual values could change this scenario. Capex cuts (C1, C2) offer a self-correcting path; none is expected near term.
3. Where is the fragility, and who bears it? Failure risk sits in the middle of the stack; the largest mark-to-market exposure sits at the top. Oracle, neoclouds and both frontier labs face funding risk. No AI-related debt has defaulted, but the stress is priced: CoreWeave's 5-year CDS was 1,066bp on 21 October (off a 1,195bp intraday peak on 15 October) and its 2031 notes yielded 15.97% on 20 October; Oracle's CDS 308bp on 21 October after 315-322bp on 19 October; Meta's 115bp on 22 October (dated reports, B). Losses without defaults could be largest in Nvidia and hyperscaler equities, and on $330B of AI-related investment-grade bonds issued through Q2 2026 (KC Fed, A). A 100-200bp spread widening implies roughly $30-60B of paper losses (arithmetic B, ranking C), partly rates-driven and unrealised for hold-to-maturity investors. First realised credit losses are likelier in construction and sub-investment-grade GPU-backed loans. Stack
The four facts that carry the answer
- The incumbents generate cash but are now borrowing to build. Microsoft's FY2026 cash capex consumed 63% of operating cash flow. Amazon raised $67B of long-term-debt proceeds in H1, and about 62% of its Q2 pre-tax income was a preferred-stock repricing gain, mainly on its Anthropic stake. Meta's free cash flow turned negative in Q3 (−$2.1B); its narrowed $135-145B capex guide puts 2026 free cash flow between −$15B and +$1B on three sensitivities, short of its ~$22B of dividends and stock-tax withholding in every case, its $82B of cash and securities now sits below its $84B of bonds, and a new non-voting Class C share is set up for equity raises (filings, A; sensitivities B). S&P forecasts negative free operating cash flow at all six rated hyperscalers in 2026-27 (A for stated forecast).
- The intermediaries are cash-strained now. Oracle: FY2026 FCF −$23.7B, August borrowings $125B (filings, A); BBB− since 9 July (S&P, B). CoreWeave: net interest consumed 25% of Q2 revenue; H1 operating cash of $3.7B versus $5.2B principal repaid; the 30 June maturity schedule lists $4.4B for July-December 2026, $6.2B for 2027 and $4.4B for 2028. Its CEO and CFO found disclosure controls ineffective at 30 June because of material weaknesses (10-Q, A). Five-year CDS has stayed above 1,000bp since 15 October and its unrestricted equity-cure window on DDTL 3.0 ends 28 October (B). SpaceX's AI segment, by contrast, turned an operating profit of $431M on $7.56B of Q3 revenue against $16.4B of capex, with $12.9B of related-party equipment financing outstanding (10-Q, A).
- Demand is growing fast and is broad, while backlog is concentrated. Azure +43%, Google Cloud +82%; nearly 90% of Microsoft's FY2026 Cloud revenue came from outside frontier labs, but OpenAI represented about 45% of December 2025 commercial RPO (filings/calls, A). OpenAI's September net annualised revenue was roughly $50B and it is reported to be seeking up to $30B at a $1.25T pre-money valuation, unclosed (Reuters, C); Anthropic's July exit run-rate was above $65B on a gross basis, with an outside ~$74B October estimate, and it has committed 5% of its shares to a public-benefit fund (lab note). Anthropic's Q3 revenue was $16.8B at an 11% adjusted operating margin. These figures are unaudited, reported, unverified (C), on different bases. Adjusted profit is not cash flow; they do not rank cash generation.
- Investment returns are unknown. No company discloses cohort returns. OpenAI's 2026-30 cash-flow forecasts drifted from +$14B to −$179B across three plans (reported, unverified, C), with a disputed fourth at −$278B (D). End-2025 cash plus the $122B round funds February's burn plan only into 2028 (reported, unverified, C). Reuters reports Anthropic commitments of at least $518B over a decade, ~80% non-cancellable or payable regardless of use (reported, unverified, C).
Forecasts
Full register: 6 / 12 / 24 months from 17 October 2026.
| Event | 6m | 12m | 24m |
|---|---|---|---|
| M1: equal-weighted basket of NVDA, AVGO, TSM, ORCL, CRWV, AMD, MU, VRT falls 40% from its 16 October level | 10% | 22% | 30% |
| M1a: the five chip names alone fall 40% | 6% | 14% | 25% |
| M1b: ORCL, CRWV, VRT alone fall 40% | 20% | 35% | 50% |
| F1: CoreWeave distressed credit event, or an acquisition closing below $40 a share | 8% | 20% | 35% |
| F2: Oracle rated below investment grade by any major agency | 8% | 24% | 40% |
| F3: OpenAI cancels, cuts, or defers by a year or more, a binding compute commitment of $10B+ | 15% | 32% | 52% |
| F3b: the same for Anthropic | 10% | 22% | 38% |
| Credit-led unwind: M1 plus distressed credit events at two distinct obligor groups | 3% | 9% | 16% |
| Equity-led unwind: M1 with at most one such event (the 2000 pattern) | 7% | 13% | 14% |
Markets: Polymarket's “AI bubble burst” requires three of six triggers from all-time highs: 6.9% by end-2026, 16.5% by June 2027, closest to M1a (market snapshot, A). CoreWeave CDS implies risk-neutral default probabilities of 8-10% at 6 months and 28-33% at 2 years, bracketing F1. Goldman's capex path fits C1. No market prices an Oracle downgrade or lab commitment cut.
Strongest arguments each way
| Topic | Bull (source, grade) | Bear (source, grade) |
|---|---|---|
| Valuation | S&P technology ~21x earnings versus 70x in 2000 (Timmer; attribution B, estimates C) | GPU buyers' capex is suppliers' revenue; earnings can be the bubble despite ordinary P/Es (Druckenmiller via FT, B) |
| Demand path | OpenAI plus Anthropic revenue $600-700B by December 2027 (SemiAnalysis, C) | Scenario: $8.5T cumulative owner revenue needed through 2030 at a 12% return; at most $3.6T attributable to AI datacenters (Gartner; attribution A, outcomes C) |
| Supply | TSMC says demand exceeds supply (Moody's upgraded it to Aa2 on 21 October); Lam beat and put 2027 wafer-fab equipment spending above $200B, led by DRAM, yet equipment stocks fell 11-13% on 22 October on a supply-capped guide; OCPI B200 monthly mean rose from $6.14 (Aug) to $7.25 (Sep) to $7.74 (Oct to date), with H100 $3.02 and B200 $9.57 on 24 October and Silicon Data H100 $2.88 (A/B) | TSMC also calls consumer segments "challenged" and is "prudent" on 2027; Nvidia guides gross margin to 71-73% (A) |
| Credit | Buybacks could be cut; Microsoft liquidity $77B (A) | IMF: ~$300B equity into weaker firms, ~$650B additional debt across investors/investees, ~$1.4T future orders returning spending to backers (staff estimates, A) |
| Demand quality | 19.8% of employer businesses use AI (Census, May 2026, A) | Token spending roughly flat August-September despite rising usage (Ramp sample, A) |
| Intermediaries | CoreWeave claims initial-term recovery and renewals (C) | DDTL 5.5 lends five years against three-year contracts (A/C) |
What we do not know, and what would tell us
| Unknown | The observable test |
|---|---|
| GPU economic life | H100 monthly index (P1, age test by 2028); used prices; depreciation changes |
| Labs' audited numbers and commitment terms | Public S-1; comparable run-rates (D1, D2); down round (F5) |
| Contract firmness | Deferral/cancellation (F3, F3b); CoreWeave principal paid from operations; amendments after DDTL 3.0's unlimited equity-cure window closes, 28 October 2026 |
| Project-bond and GPU-loan holders and marks | $1B+ project/SPV distress (F4); waivers; November lender filings |
| Demand without supplier/investor/customer financing | Backstop usage; warrant vesting; Nvidia receivables |
| Hyperscaler cuts (C1, C2) | Calendar-2027 capex guidance with Q4 2026 results |
| Oracle customer funding | Q2 FY2027 advances/hardware against assumed $20-25B annual capex offsets; Jupiter permit hearing, reportedly moved to 7 December |
| AI versus rates/oil repricing (in the week of 9-16 October yields, real yields and oil all fell while the AI basket dropped 12.7% and ORCL/CRWV/VRT fell 16.5% against 10.5% for chips: AI-specific, via credit) | Daily M1 return minus the equal-weight S&P, regressed on real 10-year, 2-year and Brent changes; AI versus non-AI high-yield spreads at the same rating |
Glossary
RPO: contracted revenue not yet recognised. FCF: operating cash flow less capex. CDS: cost in basis points (bp, 0.01%) of insuring debt. DDTL: delayed-draw term loan. SPV: special-purpose vehicle. CAPE: cyclically adjusted P/E. IG: investment grade.
Status
All 26 research notes were cross-checked by a second model, sixteen by both vendors; page reviews are under reviews and every forecast revision is in the register's history. Agents start at llms.txt. No paywalled sources were used, biasing coverage toward leaked lab figures and away from sell-side and rating-agency credit work.