# Signals: market views, 2026-10-18

*Sources searched and opened, or access attempted, 2026-10-18. USD unless specified. Verification of the Grok critique of vintage 0.4; Grok is provenance, never evidence.*

## Summary

- Timmer and Brown's earnings-based bull argument is traceable, but Timmer also acknowledges cyclical earnings and unknown infrastructure returns. [TIKR][tikr]
- TSMC and ASML report demand above available or planned supply; that supports continued orders without establishing buyers' eventual returns. [TSMC][tsmc]; [ASML][asml]
- Ornn confirms rising B200 rents alongside weaker thirty-day H100/A100 trends; there is no single GPU rental cycle. [B200][ornn-b]; [H100/A100][ornn-h]
- Microsoft plans Maia as its primary platform for its own inference from 2027; Patel's extraordinary lab-revenue paths remain estimates. [Microsoft][maia]; [interview reproduction][patel-bigtech]
- Bears have a substantial payback scenario in Gartner's $8.5tn cumulative revenue hurdle versus at most $3.6tn attributable revenue through 2030. [Gartner][gartner]
- IMF documents correlated financing and revenue exposures; Kansas City Fed identifies $330bn of AI-related IG issuance and long bond maturities. [IMF][imf]; [KC Fed][kcfed]
- Credit stress is verified: CoreWeave protection reached a reported 1,194.5bp intraday; this is a price of risk, not a completed default. [MT Newswires/Bloomberg/ICE][mt-credit]
- Polymarket displays 7% for end-2026 and 17% for June 2027; Kalshi's opened data-center contract displays 41%, not the unverified historical 75%. [Polymarket][poly]; [Kalshi][kalshi-dc]
- Metaculus forecasts large lab valuations, but these are forecasts rather than traded prices; none of these markets exactly matches M1, F1, F2, F3 or the headline. [Metaculus][meta-blog]; [forecast definitions](/forecasts/register)

## Reading this verification

**U1–U10** correspond to the ten bull arguments, **B1–B10** to the ten bear arguments, and **C1–C12** to the contested-fact rows, in their original order. **P1–P11** separate the market prices, rules, project marks and Treasury observation. Counts are of these **43 bundles**, not independent discoveries; some evidence appears in more than one bundle.

**Confirmed:** the material rendering is supported. **Partly:** some components are supported, but dates, definitions, attribution or figures need correction, or other components remain unverified. **Unconfirmed:** a numerical claim was not established. **Lead only:** an unverified attributed view. Failure to access a source does not prove the claim false. Unopened X posts and mirrors remain leads; an accessible reputable outlet can separately establish what it reports a person said. A grade for an attributed forecast verifies its publication, not its eventual outcome. Private-company financial leaks remain **reported, unverified, C**. Analyst economic estimates generally remain **C**, even when their publication is directly verifiable.

## Bull arguments

| Item | Figure or claim as verified | Source (linked) | Type | Grade | Status |
|---|---|---|---|---|---|
| U1. Timmer, Brown and DBS: earnings versus multiples | **9 Oct TIKR:** Timmer puts S&P technology at **21× earnings versus 70× in 2000**, semiconductor earnings **+178% y/y**; its own NVDA comparison is **19.1× forward versus 30.4× three-year average**. Brown's cheaper-earnings argument is reported. Timmer also accepts earnings-cycle risk and unknown ROIs. BrokerChooser reproduces **19.7×/17.3×**, but the alleged **8 Oct X** post was not opened. DBS's **5 Oct** Bloomberg search excerpt supports its earnings-led bull view; full article inaccessible. These are different dates and earnings denominators. | [TIKR, 9 Oct][tikr]; [original Compound episode][compound]; [BrokerChooser relay][timmer-pe]; [DBS/Bloomberg, fetch failed][dbs] | secondary reporting/valuation estimates; unopened X attribution | B reporting; C estimates; D X attribution | partly |
| U2. Wei/Fouquet: supply constraints | **TSMC, 15 Oct:** HPC **68%** of Q3 revenue, **+22% q/q**; 2026 capex **$60–64bn** reiterated; Wei assesses demand above supply. **ASML, 14 Oct:** planned 2027–28 supply insufficient; **30%** low-NA EUV/DUV immersion capacity expansion planned for 2027, a **further 30% investigated** for 2028; customer interest extends to **2029–30**. These are management demand assessments and plans, not proof that every downstream investment pays. | [TSMC Q3 transcript][tsmc]; [ASML SEC-filed release][asml] | primary management statements/guidance | A | confirmed |
| U3. Current-generation GPU rents | **Ornn 9 Oct settlements:** B200 **$7.46/GPU-hour**, H100 SXM **$2.79**, A100 SXM4 **$0.92**. Later: B200 **$8.12 on 16 Oct**, H100 **$2.78/A100 $0.97 on 17 Oct**. Thirty-day changes on those dates: B200 **+15.5%**, H100 **−2.5%**, A100 **−4.0%**. The **31 Mar $4.70** primary fixing and exact **FT/SemiAnalysis >$6/+50% six-month** comparison were not reproduced. Index settlements are not universal provider quotes. | [Ornn B200/H200][ornn-b]; [Ornn H100/A100][ornn-h]; [March-price tertiary lead][aiwiki] | primary index observations; tertiary historical lead | A opened settlements; D unresolved comparison | partly |
| U4. Dylan Patel: two-lab revenue path | A third-party Big Technology transcript reproduces **$600–700bn combined OpenAI/Anthropic revenue by Dec 2027**, possibly **$1.5tn**, and the lag between infrastructure spending and revenue. It discusses inference gross margins **75% or somewhat higher**, not the claimed **88%**. Original host's **12 Oct** public post confirms the demand discussion, but its financial passage is paywalled. Episode chronology differs across hosts (**7/9 Oct**). The alleged **10 Oct X** clarification and **88% API margin** remain unverified. | [transcript reproduction][patel-bigtech]; [original host, partial][bigtech]; [11 Oct write-up][patel-writeup] | secondary transcript of analyst forecasts | C estimates; D untraced extras | partly |
| U5. Baker/Thompson: commoditization increases compute | Baker's original post returned an error; a mirror dates the thread **13 Oct**, rather than Grok's 9 Oct. The specific same-compute-per-token attribution and Thompson's **9 Oct** restatement remain leads. Independently, Wei's **15 Oct** transcript argues open-source adoption lowers barriers and increases advanced-silicon demand. That is corroborating manufacturer opinion, not verification of Baker's post or a measured demand elasticity. | [Baker X, failed open][baker-open]; [mirror, lead only][baker-mirror]; [TSMC corroborating view][tsmc] | unopened social-media view; separate primary management view | D named-post lead; A TSMC attribution | lead only |
| U6. Nadella/Chamath: custom silicon and distribution | Microsoft's **12 Oct Scott Guthrie announcement** confirms multi-GW Maia deployments from **2027**, primary inference for **Microsoft's own products/services**. Maia 300 is **expected** to deliver **40% more tokens/MW** versus leading GPU systems; Maia 200 delivers **up to 1.7× tokens/watt** in specified Microsoft workloads. These are vendor comparisons; Maia 300 is prospective. Nadella's X failed. Chamath's convergence/distribution/data/power remarks appear in an indexed **12 Oct Benzinga** relay, whose full fetch failed; no independently measured margin result follows. | [Microsoft original][maia]; [architecture][maia-tech]; [Nadella X, failed][nadella-x]; [Benzinga/Yahoo relay, restricted][chamath-silicon] | primary company plan/comparison; secondary opinion relay | A company statement; C relay; D unopened X | partly |
| U7. Patel/Gokul: revenue per MW | **25 Aug Dwarkesh interview:** Patel estimates **$10–15m compute cost/MW**, revenue as high as **$50m/MW**, potentially **$70–80m/MW by end-2027**, conditional on capability leadership and deployment. **2 Sep Thematic Trader** distinguishes inference-capacity economics from approximately **$20bn/GW across the total fleet** and relays Gokul's **$30bn versus $10bn/GW**. No original JPMorgan model or **9 Oct** restatement was traced. “Already tripled” is not an established realized company metric; inference MW, fleet MW and annualized revenue cannot be interchanged. | [original Dwarkesh interview][dwarkesh]; [Thematic Trader][mw-relay] | primary interview of analyst estimates; secondary model relay | A attribution; C economics/relay | partly |
| U8. Kalshi: continued construction | The exact **9 Oct 75%, versus 60% two weeks earlier**, was not reproduced from a historical tape or fetched CNBC body. The actual opened **“at least 5,100”** contract displays **41%**; see P4. Kalshi's **22 Jul** research instead reports a **5,120 year-end count expectation**, versus **4,700 expected in early June**. That is not an observed installed base or a threshold probability. Proposed sites need not become energized capacity. | [actual Kalshi contract][kalshi-dc]; [Kalshi July research][kalshi-compute]; [CNBC, fetch failed][cnbc-dc] | primary current market/research; unverified historical lead | A current display; D historical price | partly |
| U9. IoT Analytics/Visser: absorbable impairment | **6 Oct IoT Analytics commentary**, on a **September** report, distinguishes four bubbles; models pre-tax returns above a **15% hurdle currently**, below hurdle in **2028**, negative by **2031**, and an absorbable hyperscaler write-down rather than forced systemic collapse. These are modeled returns. **12 Oct Benzinga** search text reports Visser's **11 Oct video** supply-versus-demand/zombie-company argument; original video/full relay not opened. Neither supplies an audited cohort ROIC or establishes absence of contagion. | [IoT Analytics original][iot]; [Visser/Benzinga relay, restricted][visser] | primary analyst model; secondary attributed opinion | A model attribution; C results; D original-video lead | partly |
| U10. Goldman/Zuckerberg: spending continues | **23 Sep Goldman original:** hyperscaler capex **$800bn in 2026, $1.2tn in 2027, $1.4tn in 2028**; it also expects the earnings tailwind to fade to marginal drag in 2028. These are not established AI-only spending totals. **18 Sep Goldman/Sheridan** separately forecasts **$1.4tn in 2027**, preserving a genuine model difference. The exact **54%** and five-company construction need the underlying model. Zuckerberg's hundreds-of-billions remark is relayed through the Klein discussion; its original interview was not verified. Continued competitive spending is plausible; an inevitable equilibrium is interpretation. | [Goldman, 23 Sep][gs-capex]; [Goldman/Sheridan, 18 Sep][gs-sheridan]; [Klein transcript attempt][klein] | primary analyst forecasts; interview relay | A publication; C forecasts; D untraced original quote | partly |

## Bear arguments

| Item | Figure or claim as verified | Source (linked) | Type | Grade | Status |
|---|---|---|---|---|---|
| B1. Druckenmiller/Raedler: earnings bubble | **25 Sep TheStreet**, citing FT, supports Druckenmiller's earnings-bubble concern. FT's **10 Sep** body was paywalled; the **20%-of-earlier-exposure** figure and banks' exact earnings quote were not verified in an opened reputable body. **17 Oct 24/7 Wall St**, on Raedler's **16 Oct BofA note**, corroborates capital-cycle disappointment and next-twelve-month hyperscaler capex consensus **$330bn early-2025 → $1tn**. This is a forecast series, not spending already incurred. Bank original unavailable. | [TheStreet][druck]; [FT, paywall][druck-ft]; [24/7 Wall St/AOL][bofa] | secondary attributed views/analyst estimates | B attribution; C estimates; D untraced exposure | partly |
| B2. Gartner/Van Nieuwerburgh: payback hurdle | **14 Oct Gartner release:** nearly **$1tn** infrastructure in 2026 plus **$6tn in 2027–30**; **$8.5tn cumulative revenue** required through 2030 at **12% ROIC**, versus **≤$3.6tn** attributable revenue; owners earn **less than half** of AI revenue. Exposure **>$3.5tn by 2030/>$4tn through 2032**; **>80%** neoclouds cease independent operation, which includes consolidation. Underlying research abstract is **14 Sep**. Van Nieuwerburgh's Brookings draft, appendix pp.27–29: **$3.725tn mature annual revenue/9.2% of projected 2032 GDP**, assuming **10% unlevered return, 50% cash margin, six-year IT life**; three-year life gives **~$6tn/14.8%**. These are scenarios, not realized losses or forecast sales. | [Gartner release][gartner]; [research abstract][gartner-paper]; [Brookings conference draft][brookings] | primary attributed economic scenarios | A attributed scenario; outcomes untested | partly |
| B3. IMF/Smets/Nvidia: circular financing | IMF October GFSR, ch.1 pp.3–4: **14 disclosed transactions since 2023**, about **$300bn equity led by hyperscalers**, **$650bn additional debt**, and **$1.4tn future revenue commitments**. About **$320bn** of the debt is taken on by investees: Grok wrongly assigns all debt to hyperscalers investing in labs. AI correction is **one of three triggers**, not ranked “top.” Reuters traces Smets' **14 Sep** opacity warning. Nvidia filings support **$36bn AI-cloud agreements** and preliminary platforms to mobilize **>$500bn**, not committed funding; the six named capital-provider counterparties were not individually verified. CoreWeave's **15 Sep 2025 8-K** confirms the **$6.3bn** initial residual-capacity order, conditional through **13 Apr 2032**. Ohio's **$105bn cap/4.25GW** is **filed**, conditional, beginning with service readiness expected from **2028**, with reletting offsets; not an immediately payable blanket guarantee. **$250bn** and the Loudoun/2008 analogy remain untraced. | [IMF chapter][imf]; [briefing][imf-brief]; [Reuters/Smets][smets]; [Nvidia 10-Q][nvda-q]; [CoreWeave 8-K][crwv-backstop]; [Ohio 8-K][ohio] | primary institutional analysis/filings; secondary interview | A filings/IMF publication; B Smets; D extras | partly |
| B4. Kedrosky/KC Fed: short-lived collateral, long debt | **KC Fed, 5 Oct:** AI-related IG issuance **$330bn YTD**, **10× full-year 2023**; share **2% in 2023 → 10% in 2026**. Issuance chart runs through **Q2**; the maturity sample runs through **Aug 2026**: hyperscalers **16 years**, utilities **17**, versus **10** market-wide. It discusses CRE/REITs, CMBS, ABS and nonbank exposure. **9 Oct Benzinga** reports Kedrosky's technology/property/credit/policy thesis and **5–10-year** economic-obsolescence view. Corporate issuers also own durable businesses and buildings; bond tenor is not automatically GPU collateral life. | [KC Fed bulletin][kcfed]; [Kedrosky/Capital Thesis via Benzinga][ked-risk] | primary central-bank analysis; secondary stated view | A analysis; B attribution, not measured GPU life | confirmed |
| B5. Barclays/Tan: enterprise demand quality | **14 Oct Investing.com** corroborates Long's **DELL/HPE/NTAP Equal Weight** downgrades, no near-term enterprise AI inflection and memory-cost-driven revenue. **15 Oct IT Pro** reports AI approximately **5% of IT budgets**; the **one-quarter hardware** split was not verified in a fetched reputable body. The Barclays original was unavailable. **12 Oct Benzinga**, citing Bloomberg, confirms Tan's arcade-token criticism, but also his acknowledgment of strong lab demand and concern that capital becomes the constraint. OEM demand, frontier-lab revenue and all enterprise AI are different markets. | [Investing.com/Long][barclays]; [IT Pro][itpro]; [Tan/Benzinga][tan] | secondary analyst/CEO views and survey estimates | B attribution; C budget estimate; D hardware split | partly |
| B6. Ramp/Timmer/Chamath: token deflation | **15 Oct Breakingviews** supports token-price competition and eroding revenue per unit, but its accessible body did not reproduce **+90% tokens/+19% spend**. Primary Ramp **9 Sep** reports effective price **$1.15 March peak → $0.68 (−41%)**. **14 Oct Ramp** reports two flat-spend months, but also **+11.7%** top-cohort September intensity and an early-October spending high **1.5%** above August. **23 Sep Benzinga/Webull** reports Timmer's **−49%/50-day** expenditure index and cooling H100/A100 lease gains after **+56%/+16%**. Chamath's **70–80%**, reported **16 Oct**, is explicitly conjecture. The last-week-run-rate claim was not traced. | [Breakingviews][breakingviews]; [Ramp September][ramp-sep]; [Ramp October][ramp-oct]; [Timmer relay][timmer-spend]; [Chamath/Benzinga][chamath] | primary panel data; secondary estimates/opinion | A panel observations; B attribution; C conjecture; D missing figures | partly |
| B7. Jupiter/project bonds/issuance | **18 Sep Reuters/FT:** **$18bn** Jupiter loans privately quoted **89–91¢**, not executed trades/defaults. **24 Sep Reuters/Bloomberg:** reported force-majeure notice seeks payment deferral if **2028** delivery missed; Reuters could not independently verify. Neither is new October news. **15 Oct FT/Folha** supports Hyperion **just below 92¢**, Hut 8 Louisiana/Texas **~94¢/~92¢**, Excalibur **below 88¢**; alternate FT/Yahoo indexed marks differ. **16 Oct Axios:** September issuance **$23bn versus June $113bn**; Dimon tightened credit. These observations do not establish an entirely closed financing market. | [Reuters Jupiter][jupiter]; [Reuters notice][force-majeure]; [fetched FT/Folha][ft-folha]; [FT/Yahoo, fetch restricted][ft-yahoo]; [Axios][axios-debt] | secondary private-market reporting | B reporting; C private quotes/unverified notice | partly |
| B8. Burry/Kedrosky/Tusk: IPO supply | **9 Oct Kedrosky original** argues giant offerings already crowd out smaller issuance as investors reserve capital; this does not establish that the public market cannot absorb the labs. **12 Oct Benzinga** confirms Tusk's concern about **~27× revenue**, conditional on sustained growth. Financial Express reports Burry's **8 Oct 78-company** comparison without disclosing its valuation input; original X failed, so that post/arithmetic remains a lead. Metaculus's **30 Sep $1.88tn/$1.84tn** forecasts are verified in P7; they concern lock-up expiry, not initial float or IPO proceeds. | [Kedrosky essay][ked-ipo]; [Tusk/Benzinga][tusk]; [Financial Express relay][burry-relay]; [Metaculus][meta-blog] | primary stated view; secondary interview; unopened X lead | B views; D Burry arithmetic; A Metaculus publication | partly |
| B9. Open Markets: ultimate holders | **16 Oct Open Markets original** argues data-center debt reaches pensions, endowments and institutional portfolios, and PE-owned insurers expose policyholders. This verifies an advocacy position, not a holdings census. KC Fed independently discusses institutional private placements and indirect household exposure. Neither identifies all holders or a quantified **BBB−/BB GPU-mezzanine** allocation. **Axios estimates one-third of AI capex uses debt.** Its method was not reproduced. Financing layers need security-by-security ownership and waterfall evidence. | [Open Markets report][omi]; [KC Fed][kcfed]; [Axios estimate][axios-debt] | primary advocacy/central-bank analysis; secondary estimate | B advocacy attribution; A Fed analysis; C financing-share estimate | partly |
| B10. Dalio/Wood: rates and refinancing | **16 Oct Forbes** supports Dalio's approaching-bubble-break warning; Bloomberg's **7 Oct Singapore** search result supports the debt/rates/cash mechanism, but full article could not be fetched. **Treasury:** ten-year **5.24% on 9 Oct**, falling to **5.09% on 16 Oct**; Grok's 10 Oct is a Saturday carry-forward, not a fixing, and “highest since 2002” was not established. **17 Oct ET/ANI**, quoting Wood/Jefferies, confirms Oracle CDS **186→288bp**, CoreWeave **749→1,118bp**, and CoreWeave 2032 yield **14.8%**. These are dated quotes, not 18 Oct live prices. | [Forbes][dalio-forbes]; [Treasury][treasury]; [Jefferies via ET/ANI][wood] | primary rates; secondary views/dealer marks | A rates; B attribution/marks | partly |

## Contested facts

| Item | Figure or claim as verified | Source (linked) | Type | Grade | Status |
|---|---|---|---|---|---|
| C1. CAPE versus forward earnings | Timmer/Brown's counterargument is traceable in **9 Oct** reporting (U1); it is not a refutation of a ten-year real-earnings metric. **16 Oct FactSet:** S&P 500 forward P/E **19.0×**, versus **19.8× five-year/19.1× ten-year averages**. Grok's different **13–19× NVDA** observations need dated consensus denominators; **13×** was not verified here. Shiller's original workbook link was located, but web fetch could not parse XLS; this pass does not independently recalculate the report's **40.6/44.2** CAPE readings. | [TIKR][tikr]; [FactSet original][factset]; [Shiller original download page][shiller] | secondary reporting; primary data-provider publication | B views; A FactSet's stated series; D unverified 13× | partly |
| C2. Unknown returns versus competing estimates | Patel's bull path is reproduced, not audited (U4/U7); IoT's **>15% modeled pre-tax return now**, below zero in **2031**, and Gartner's **12% hurdle/$8.5tn versus ≤$3.6tn** are explicitly different models. Their coexistence supports presenting alternative assumptions, not declaring realized payback known. Lab run-rates, infrastructure-owner revenue, cumulative customer revenue and total AI benefits are different objects. | [Patel interview][dwarkesh]; [IoT][iot]; [Gartner][gartner] | primary/secondary analyst models | A model attribution; C economic estimates | partly |
| C3. A100/H100 versus Blackwell rents | Ornn confirms **9 Oct $0.92/$2.79** and **17 Oct $0.97/$2.78** for A100/H100, versus B200 **$8.12 on 16 Oct** (U3). Thus old-chip discounting and current-generation scarcity coexist. The exact **FT +50% six-month** claim and **23 Sep Timmer** directional comparison require consistent indices and intervals. Rental dollars per GPU-hour do not directly measure dollars per token or collateral resale proceeds; this check does not independently re-run the report's Blackwell benchmark. | [Ornn older GPUs][ornn-h]; [Ornn B200][ornn-b]; [Timmer reporting][timmer-spend] | primary indices; secondary views | A observations; B attribution; D unverified six-month comparison | partly |
| C4. Anthropic growth and accounting | **14 Oct Reuters:** Q3 **$16.8bn**, Q2 **$11.6bn**, Q1 **$4.7bn**, sequential growth **45%**, adjusted profit. **15 Oct Euronews/Bloomberg:** Q2 **$11.5bn**, **11% adjusted operating margin**, adjusted positive FCF. Bay Area Times relays exclusion of SBC and **5–7%** distillation revenue; original FT/Information bodies inaccessible. The **85% GAAP gross-margin correction**, Jefferies **≥$80bn** expectation and precise **+90%/+19%** token comparison were not established in opened original/reputable bodies. ROIC does not follow from gross margin. Luria's competition interpretation is reproduced by ROIC, not independently proven. | [Reuters][anthropic-reuters]; [Euronews][anthropic-euro]; [Bay Area Times relay][anthropic-bat]; [FT access attempt][anthropic-ft]; [ROIC/Luria][luria]; [unopened Ludlow profile][ludlow-x] | secondary private financial reports/analyst interpretations; tertiary leads | C reported, unverified; D unresolved basis/expectations | partly |
| C5. OpenAI net/gross revenue | **8 Oct Reuters:** almost **$50bn September annualization**; earlier approximately **$70bn** comparison arose from different partner accounting. Its report explains Anthropic counts partner sales differently and cites partner payments of approximately **16%** on those sales. The alleged **≥$70bn year-end target** and **107% Q3 enterprise growth** appear in **9 Oct Quartz/Bloomberg/CNBC** indexed reporting; full Quartz/Yahoo fetch was restricted. These are not realized year-end sales. No same-basis revenue decline is established by the revision. | [Reuters, fetched][openai-reuters]; [Quartz/Yahoo, restricted][openai-quartz]; [CNBC, fetch failed][openai-cnbc] | secondary private-company financial reports | C reported, unverified | partly |
| C6. Census use versus willingness to pay | **26 May Census publication:** national employer-business AI use **19.8% as of 3 May 2026**. October API failed access; download page exposed only a JavaScript shell, so **23.8%** remains unreproduced in this pass. Barclays' **~5% of IT budgets** is a different survey/channel estimate (B5); Tan's token-value criticism is opinion. Adoption and expenditure answer different questions; neither invalidates the other by itself. | [Census original][census]; [October API, failed][census-api]; [download shell][census-download]; [IT Pro/Barclays][itpro]; [Tan][tan] | primary survey; secondary budget estimates/views | A May use; C budget estimate; D October value | partly |
| C7. CoreWeave credit versus Huang | **15 Oct** intraday protection high **1,194.5bp** is corroborated by **16 Oct MT Newswires**, citing Bloomberg/ICE. Reuters reports a **64% five-year modeled risk**, not a uniquely observable physical probability. **15 Oct FT/Folha** reproduces Huang's improving-finances/decreasing-capital-cost/no-new-backstop-needed view. This confirms disagreement in interpretation; it does not make traded spreads or a supplier's assurances a resolution event. | [MT Newswires][mt-credit]; [Reuters][credit-reuters]; [FT/Folha, Huang remarks][ft-folha] | secondary market reporting/CEO view | B quotes/attribution; C probability model | partly |
| C8. Oracle BBB− and cash deficit | **9 Jul Investing.com** confirms S&P **BBB→BBB−, stable**, projected FY2027 capex **$90–95bn**, FOCF **−$42bn**. Original rating page could not be fetched. Oracle's **10 Jun release** directly states FY2026 FCF **−$23.7bn**; August 10-Q directly states **$125.0bn senior notes/other long-term borrowings**. Those are mixed dates and specific accounting categories. IG status and fragility can coexist; stable outlook is not a probability or guarantee against downgrade. | [rating-action reporting][oracle-rating]; [S&P original, failed][oracle-sp]; [Oracle FY2026 release][oracle-fy]; [August 10-Q][oracle-q] | primary financial disclosures; secondary rating report | A financials; B rating | confirmed |
| C9. Microsoft Cloud breadth versus AI-cloud concentration | **29 Jul Microsoft call:** nearly **90% of FY2026 Microsoft Cloud revenue** from customers outside frontier model companies; **$214bn+** full-year cloud revenue; all sequential commercial-RPO growth from outside those companies. This is broad Cloud revenue, not an AI-only revenue or marginal-MW split. Azhar's alleged **14 Oct** post was not located/opened. Baker's **15 Oct** hyperscaler-growth argument appears in an Invezz search excerpt, but fetched page and original X were unavailable: named-post attribution remains lead only. | [Microsoft call][msft-call]; [Azhar profile, failed][azeem-x]; [Baker/Invezz lead, fetch failed][baker-revenue] | primary company mix; unopened social/secondary leads | A Microsoft statement; D exact named-post leads | partly |
| C10. Nvidia commitments versus bottlenecks | **Q2 FY2027 CFO commentary:** supply/capacity commitments **$279bn versus $119bn prior quarter**, **primarily memory**, at **26 Jul 2026**. TSMC/ASML supply constraints are sourced in U2. Forward procurement may secure scarce supply and still create downside obligations if demand later changes. Neither “shortage” nor “speculative inventory” follows from the dollar total alone; the forecast comparison needs component and delivery schedules. | [Nvidia SEC commentary][nvda-cfo]; [TSMC][tsmc]; [ASML][asml] | primary disclosures; interpretation explicitly identified | A figures/statements | confirmed |
| C11. Sentiment/new-buyer entry | **16 Oct Benzinga** reports Social Capital's claim that AI/ML captured **~83% of US venture funding through September**. This is attributed deal-value concentration, not a direct measure of household new-buyer entry; no underlying deal-level bridge was reproduced. Burry's “denial” post and current household-allocation figures were not established. Concentration and bullish positioning deserve discussion, but do not mechanically change the report's partial indicator score. | [Benzinga/Social Capital][chamath]; [Burry reporting, only related comparison][burry-relay] | secondary estimate; untraced opinion/household leads | C concentration estimate; D unresolved leads | partly |
| C12. Correlated exposures versus loss order | IMF's **$300bn equity/$650bn debt/$1.4tn commitments** analysis verifies wrong-way risk, not a universal creditor waterfall. Nvidia's **17 Aug 8-K** establishes conditional residual-value support, tenant reimbursement and offsets from reletting/sale. A single counterparty failure can impair several exposures at once; seniority, recourse and guarantees still determine each security's recovery. Grok's claim that correlation supplies a fixed loss order is not confirmed. | [IMF ch.1][imf]; [Nvidia Ohio 8-K][ohio] | primary institutional analysis/legal disclosure; inference | A source terms; B inference | partly |

## Market prices

*Access-day displays are identified as such. They are not necessarily executable midpoints, and cumulative volume is not current liquidity. Historical quotes remain historical. Metaculus contributes forecast distributions, not wager prices.*

| Item | Figure or claim as verified | Source (linked) | Type | Grade | Status |
|---|---|---|---|---|---|
| P1. Polymarket, end-2026 | Opened page displays **7% by 31 Dec 2026**. API snapshot updated **17 Oct 22:05 UTC**: midpoint **6.9%**, Yes bid/ask **6.3–7.5¢**, last **7.5¢**; cumulative volume **$2.442m**, displayed liquidity **$39,909**. Supports Grok's approximate **6–7%**, not a fresh 18 Oct fixing. | [market page][poly]; [venue API][poly-api] | primary market snapshot | A | confirmed |
| P2. Polymarket, mid-2027 | Same API: **16.5% midpoint** for **30 Jun 2027**; page rounds to **17%**. Bid/ask **16–17¢**, last **17¢**; cumulative volume **$24,327**, liquidity **$29,289**. Grok's **18–20%** may be another vintage, but no dated tape establishing it was obtained. | [market][poly]; [API][poly-api] | primary market snapshot | A current observation; D unverified historical price | partly |
| P3. Polymarket resolution rules | Any **three of six** triggers within one **90-day window ending by the deadline**: NVDA **−50% ATH**; SOXX **−40% ATH**; either lab bankruptcy; OpenAI acquisition; H100 **≤$1/hour for five days** on specified index; one of **TSM/ASML/AVGO/ANET/SMCI −50% ATH**. The enumerated conditions match; Grok's “crash-plus-insolvency” characterization does not. Three equity conditions can suffice. | [venue rules/API][poly-api] | primary contract rules; logical comparison | A rules; B inference | partly |
| P4. Kalshi, US data centers | Opened **KXUSADATACENTERS-26DEC31** page: **“at least 5,100” 41%**, Yes offer **46¢**, No offer **59¢**; event cumulative volume **$68,368**; settlement displayed **7 Jan 2027**. No fixing timestamp exposed. This verifies the contract/display, not a contemporaneous 41% midpoint. **75% on 9 Oct**, prior **60%**, and the **>4,700 current-count** assertion were not established from an opened historical source. July's **4,700** was an earlier forecast, not an observed stock. | [contract][kalshi-dc]; [Kalshi July research][kalshi-compute]; [CNBC historical-price attempt][cnbc-dc] | primary venue display; unverified historical quote | A display; D historical bundle | partly |
| P5. Kalshi, Nvidia largest company | No specific Kalshi contract, deadline or dated tape supporting **67–69% on 15–17 Oct**, falling from **91% in 24 hours**, was identified after targeted searches. Related Polymarket contracts have different horizons and cannot substitute. The claimed numerical move remains unconfirmed; no probability is adopted. | [Kalshi venue search, provenance only](https://kalshi.com/markets); no matching contract located | unverified numerical lead | D | unconfirmed |
| P6. Kalshi, which lab confirms IPO first | Direct APIs identify the sides: **OpenAI** Yes bid/ask **14–20¢**, midpoint **17%**, last **20¢**; **Anthropic 93–94¢**, last **93¢**. Each concerns **confirming an IPO first before 1 Jan 2040**, not completing a listing. Separate binaries are not complementary quotes. Last-24h volumes approximately **1,387/40 contracts**, respectively. Accessed **18 Oct**, no verified quote-fixing timestamp; API's April update metadata is not a current fixing. Grok's **14 Oct** history was not reproduced. | [OpenAI API][kalshi-oai]; [Anthropic API][kalshi-anth] | primary market quotes/rules | A current display/rules; D historical date | partly |
| P7. Metaculus, valuations and IPO dates | **30 Sep official article:** Anthropic **$1.88tn**, 50% interval **$1.41–2.51tn**; OpenAI **$1.84tn**, **$1.35–2.35tn**; IPO dates **7 Dec 2026/25 May 2027**. Direct pages accessed **18 Oct** instead display **$1.98tn Anthropic (18 forecasters)/$1.84tn OpenAI (14)** and dates **30 Nov 2026/27 May 2027 (26/18)**. Valuations are at lock-up expiry, roughly 180 days later; historical intervals cannot be carried over to the new medians. These are small-panel forecasts, not clearing prices. | [30 Sep article][meta-blog]; [Anthropic valuation][meta-anth]; [OpenAI valuation][meta-oai]; [Anthropic date][meta-anth-date]; [OpenAI date][meta-oai-date] | primary forecast publications | A publication, not valuation outcome | confirmed |
| P8. CoreWeave CDS and bond yields | **15 Oct** high **1,194.5bp** per Bloomberg/ICE, relayed **16 Oct MT Newswires**. **17 Oct Wood/Jefferies relay:** **749→1,118bp since 11 Sep**, **2032 yield 14.8%**. **15 Oct Reuters:** protection above **1,100bp**; **2031 yield briefly 16%, later ~14%**; modeled **64% five-year default risk** at around 1,000bp. Thus Grok's distress/range is supported, but **60–63%** is not a unique conversion of every print. Recovery and term structure matter; no observed default is established. | [MT Newswires/Bloomberg/ICE][mt-credit]; [ET/ANI/Jefferies][wood]; [Reuters][credit-reuters] | secondary dated market quotes/model output | B prices; C default-probability estimate | partly |
| P9. Oracle CDS | **17 Oct ET/ANI** reproduces Christopher Wood/Jefferies: five-year protection **186bp on 11 Sep → 288bp**. Later **15 Oct Reuters** version reports **near 300bp**, about **31bp wider over the week**. These are distinct observations, not conflicting official closes; neither is an 18 Oct live fixing or a fallen-angel probability. | [Wood/Jefferies via ET/ANI][wood]; [Reuters][credit-reuters] | secondary dated dealer/data-service quotes | B | confirmed |
| P10. Project-debt marks | **15 Oct FT/Folha:** Hyperion/Beignet **just below 92¢**, Hut 8 Louisiana **~94¢**, Texas **~92¢**, Excalibur **below 88¢**. Matches Grok's approximate project marks in this fetched vintage; FT/Yahoo search shows different snapshots, without a reconciliation. **18 Sep Reuters/FT:** **$18bn Jupiter loans privately quoted 89–91¢**. Quotes are not necessarily executed transactions, let alone distressed exchanges or defaults. Project paper is not the same obligation as its corporate tenant's debt. | [FT/Folha, fetched][ft-folha]; [Reuters Jupiter][jupiter]; [FT/Yahoo alternate, fetch restricted][ft-yahoo] | secondary private-market quotes | C individual marks | confirmed |
| P11. Ten-year Treasury | Primary table: **5.24% on 9 Oct**, **5.19% on 14 Oct**, **5.10% on 15 Oct**, **5.09% on 16 Oct**. **10 Oct** was Saturday. Rising funding costs over a longer interval can be a mechanism, but “rates are already up” is misleading as a description of this latest week. “Highest since 2002” was not verified. | [Treasury daily par yields][treasury] | primary market-rate series | A observations; D historical-record assertion | partly |

**Bundle counts: 8 confirmed; 33 partly; 1 unconfirmed; 1 lead only.** Unsupported subclaims inside a partly verified bundle remain unsupported; these totals should not be read as a fact-level accuracy score.

[tikr]: https://www.tikr.com/blog/ai-bubble-warnings-spike-but-fidelity-strategist-says-valuations-are-actually-falling
[compound]: https://podcasts.apple.com/us/podcast/this-is-where-the-rubber-meets-the-road-with-jurrien-timmer/id1456467014?i=1000794030026
[timmer-pe]: https://brokerchooser.com/news/bond-yields-reaching-6-could-compress-stock-valuations-by-a-fifth-fidelity-warns--a94cbd7b
[dbs]: https://www.bloomberg.com/news/articles/2026-10-05/nvidia-s-valuations-show-ai-rally-isn-t-a-bubble-dbs-says
[tsmc]: https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf
[asml]: https://www.sec.gov/Archives/edgar/data/937966/000162828026065813/pressreleasefinancialresul.htm
[ornn-b]: https://data.ornn.com/markets/compare/h200-vs-b200
[ornn-h]: https://data.ornn.com/markets/compare/h100-sxm-vs-a100-sxm4
[aiwiki]: https://aiwiki.ai/wiki/kalshi
[patel-bigtech]: https://ceointerviews.ai/interview/3018643/
[bigtech]: https://www.bigtechnology.com/p/can-ai-keep-growing-exponentially
[patel-writeup]: https://ceointerviews.ai/en/resources/ai-revenue-timeline-patel-levie-enterprise-diffusion-20261011/
[baker-open]: https://x.com/GavinSBaker/status/2110041532770160668
[baker-mirror]: https://en.rattibha.com/thread/2110041532770160668
[maia]: https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/
[maia-tech]: https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture/
[nadella-x]: https://x.com/satyanadella/status/2109631106404487523
[chamath-silicon]: https://finance.yahoo.com/technology/ai/articles/chamath-palihapitiya-elon-musk-highlight-204625617.html
[dwarkesh]: https://www.dwarkesh.com/p/dylan-patel-3
[mw-relay]: https://www.thematictrader.com/p/someone-has-to-buy-120-gigawatts
[kalshi-dc]: https://kalshi.com/markets/kxusadatacenters/number-of-data-centers-in-usa/kxusadatacenters-26dec31
[kalshi-compute]: https://kalshi.com/research/insights/the-economics-of-compute
[cnbc-dc]: https://www.cnbc.com/2026/10/09/kalshi-traders-see-high-odds-of-5000-plus-data-centers-by-2027.html
[iot]: https://iot-analytics.com/why-there-is-no-ai-bubble-analysis/
[visser]: https://finance.yahoo.com/technology/ai/articles/ray-dalio-says-ai-bubble-233008164.html
[gs-capex]: https://www.goldmansachs.com/insights/articles/can-ai-investment-drive-s-and-p-500-earnings-even-higher
[gs-sheridan]: https://www.goldmansachs.com/insights/articles/consumer-agents-signal-new-phase-for-ai-growth
[klein]: https://podscripts.co/podcasts/the-ezra-klein-show/is-there-an-ai-bubble-a-debate
[druck]: https://www.thestreet.com/investing/billionaire-investor-warns-of-risk-building-beneath-ai-boom
[druck-ft]: https://www.ft.com/content/7410c56e-5350-48f8-a5e6-24f37c1581e9
[bofa]: https://www.aol.com/articles/bofa-warns-ai-investors-could-114411000.html
[gartner]: https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030
[gartner-paper]: https://www.gartner.com/en/documents/8381081
[brookings]: https://www.brookings.edu/wp-content/uploads/2026/09/4c_Van-Nieuwerburgh.pdf
[imf]: https://www.imf.org/-/media/files/publications/gfsr/2026/october/english/ch1.pdf
[imf-brief]: https://mediacenter.imf.org/news/global-financial-stability-report-press-briefing/s/17b00d8b-eb4a-458d-970d-e056c46f8431
[smets]: https://ca.marketscreener.com/news/bis-says-global-market-ai-momentum-showing-signs-of-vulnerability-ce785bdcdd89f724
[nvda-q]: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm
[crwv-backstop]: https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm
[ohio]: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm
[nvda-cfo]: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27cfocommentary.htm
[kcfed]: https://www.kansascityfed.org/research/economic-bulletin/beyond-valuations-long-term-financial-risks-from-the-ai-build-out/
[ked-risk]: https://www.benzinga.com/markets/prediction-markets/26/10/62273895/oracle-4-bubble-risks-converge
[barclays]: https://www.investing.com/news/stock-market-news/barclays-downgrades-dell-stock-after-sharp-rally-4949499
[itpro]: https://www.itpro.com/technology/artificial-intelligence/data-centre-and-cloud-shares-slide-amid-ai-gloom
[tan]: https://www.benzinga.com/markets/prediction-markets/26/10/62305417/broadcom-ai-investors-casino-gambling
[breakingviews]: https://www.breakingviews.com/40925ca4af2b/columns/breaking-view/anthropics-2-trln-dream-faces-deflation-trap-2026-10-15/
[ramp-sep]: https://ramp.com/data/ai-index-sept-2026
[ramp-oct]: https://ramp.com/data/ai-index-oct-2026
[timmer-spend]: https://www.webull.com/news/15626732893045760
[chamath]: https://www.benzinga.com/markets/prediction-markets/26/10/62404224/chamath-palihapitiya-finds-revenue-quality-problem-in-ai-firms-warns-of-overreliance-on-venture-capital-but-crypto-punters-dont-expect-a-major-downturn
[jupiter]: https://www.investing.com/news/stock-market-news/oracles-18-billion-data-center-debt-under-pressure-ft-reports-4907951
[force-majeure]: https://economictimes.indiatimes.com/tech/technology/oracle-cites-force-majeure-to-shield-itself-on-controversial-data-centre-bloomberg-news/articleshow/134462802.cms
[ft-folha]: https://c-level.folha.uol.com.br/financas/2026/10/investidores-se-desfazem-de-titulos-de-data-centers-enquanto-mercado-questiona-receitas-de-ia.shtml
[ft-yahoo]: https://finance.yahoo.com/markets/stocks/articles/investors-dump-data-centre-bonds-132608292.html
[axios-debt]: https://www.axios.com/2026/10/16/wall-street-limit-ai-spending-debt
[ked-ipo]: https://paulkedrosky.com/the-glacial-lake-outburst-problem-in-ai-ipos/
[tusk]: https://www.benzinga.com/markets/tech/26/10/62293658/ai-bubble-ready-to-burst-investor-raises-alarm-over-anthropics-sky-high-valuation
[burry-relay]: https://www.financialexpress.com/market/global-markets/michael-burry-anthropics-valuation-equals-78-sampp-500-companies/4356774/
[omi]: https://www.openmarketsinstitute.org/publications/report-brace-for-impact-protecting-people-and-the-economy-when-the-ai-bubble-burst
[dalio-forbes]: https://tollbit.forbes.com/sites/giacomotognini/2026/10/16/why-debt-laden-neoclouds-could-become-a-trillion-dollar-wework-20/
[treasury]: https://home.treasury.gov/resource-center/data-chart-center/interest-rates/TextView?field_tdr_date_value=2026&type=daily_treasury_yield_curve
[wood]: https://economictimes.indiatimes.com/markets/us-stocks/news/rising-borrowing-costs-signal-funding-risks-for-global-ai-expansion-jefferies-report/articleshow/135012567.cms
[factset]: https://insight.factset.com/sp-500-earnings-season-update-october-16-2026
[shiller]: https://shillerdata.com/
[anthropic-reuters]: https://www.investing.com/news/stock-market-news/anthropics-q3-revenue-surges-45-to-nearly-17-billion-4950208
[anthropic-euro]: https://www.euronews.com/2026/10/15/anthropic-revenue-jumps-46-to-168bn-as-ai-firm-eyes-record-2-trillion-ipo-valuation
[anthropic-bat]: https://www.bayareatimes.com/p/anthropic-q3-revenue-said-to-jump-to-16-8b-posts-adjusted-profit
[anthropic-ft]: https://www.ft.com/content/11632d03-358c-454b-af95-d06d4072932b
[luria]: https://www.roic.ai/news/anthropics-q3-revenue-growth-slows-to-45-but-da-davidson-says-its-not-a-demand-problem-10-15-2026
[ludlow-x]: https://x.com/EdLudlow
[openai-reuters]: https://www.streetinsider.com/Reuters/OpenAIs%2BSeptember%2Bannualized%2Brevenue%2Bnears%2B%2450%2Bbillion%2C%2Bless%2Bthan%2Bpreviously%2Bindicated%2C%2Bsource%2Bsays/27167688.html?classic=1
[openai-quartz]: https://finance.yahoo.com/technology/ai/articles/openai-projects-70-billion-annualized-113114847.html
[openai-cnbc]: https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html
[census]: https://www.census.gov/library/stories/2026/05/ai-use-businesses.html
[census-api]: https://www.census.gov/hfp/btos/api/periods/110/data
[census-download]: https://www.census.gov/hfp/btos/data_downloads
[mt-credit]: https://vanguard.factsetdigitalsolutions.com/markets/newsStory?newsId=450e90f4d667b0a0&notation=8941848&tabName=indexes_newsStory
[credit-reuters]: https://www.aol.com/articles/ai-credit-default-swap-spreads-051828000.html
[oracle-rating]: https://www.streetinsider.com/dr/news.php?id=26752754
[oracle-sp]: https://www.spglobal.com/ratings/en/regulatory/article/-/view/type/HTML/id/3592348
[oracle-fy]: https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Record-Q4-and-FY-2026-Results-Driven-by-Cloud-Infrastructure--Cloud-Applications/
[oracle-q]: https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm
[msft-call]: https://www.microsoft.com/en-us/Investor/events/FY-2026/earnings-fy-2026-q4
[azeem-x]: https://x.com/azeem
[baker-revenue]: https://invezz.com/sg/news/2026/10/17/anthropics-ipo-valuation-cut-is-it-bad-news-for-ai-bulls/
[poly]: https://polymarket.com/event/ai-bubble-burst-by
[poly-api]: https://gamma-api.polymarket.com/events/slug/ai-bubble-burst-by
[kalshi-oai]: https://api.elections.kalshi.com/trade-api/v2/markets/KXOAIANTH-40-OAI
[kalshi-anth]: https://api.elections.kalshi.com/trade-api/v2/markets/KXOAIANTH-40-ANTH
[meta-blog]: https://metaculus.substack.com/p/ai-ipo-windfall-prediction-forecast-radiant-visual-mapping
[meta-anth]: https://www.metaculus.com/questions/45334/what-will-be-anthropics-market-cap-when-its-ipo-lock-up-period-ends/?trk=article-ssr-frontend-pulse_little-text-block
[meta-oai]: https://www.metaculus.com/questions/44912/what-will-be-openais-market-cap-when-its-ipo-lock-up-period-ends/?trk=article-ssr-frontend-pulse_little-text-block
[meta-anth-date]: https://www.metaculus.com/questions/44767/?trk=article-ssr-frontend-pulse_little-text-block
[meta-oai-date]: https://www.metaculus.com/questions/44766/when-will-openai-ipo/?trk=article-ssr-frontend-pulse_little-text-block
[sp-capex]: https://press.spglobal.com/2026-08-27-AI-Infrastructure-Investment-To-Exceed-1-3-Trillion-By-2027,-S-P-Global-Ratings-Says

## Implications for the report

- **Front page and synthesis:** [report/index](/report/index), [theory](/report/theory) and [stack](/report/stack) should distinguish earnings-multiple arguments from the durability of those earnings; foreground the opposing payback scenarios and debt transmission. Several claimed omissions already appear in the underlying notes. This note recommends presentation changes, not automatic acceptance of Grok's “offsides” judgments.
- **Payback and demand:** [unit economics](/research/theory/ai-unit-economics) and [demand](/research/macro/demand-side) should place Gartner, Van Nieuwerburgh, IoT and Patel side by side with capital vintages, annual versus cumulative revenue, inference versus fleet MW and owner capture stated. The report's **$750bn cohort/$295–468bn annual revenue hurdle** is an internal scenario, not a measured gap against all end-user revenue. Retain uncertainty about realized cohort returns.
- **Supply and silicon:** [TSMC](/research/stack/tsmc-and-foundry), [equipment](/research/stack/semiconductor-equipment), [custom silicon](/research/stack/amd-broadcom-custom-silicon), [Microsoft](/research/stack/microsoft), [GPU pricing](/research/macro/gpu-pricing-and-utilization) and [memory](/research/stack/memory-hbm) should clearly separate supply constraints, planned Maia substitution, current-generation rents and old-chip residual values. A statement about all GPU rents falling would be unsupported.
- **Credit and counterparties:** [circular financing](/research/macro/circular-financing), [debt markets](/research/macro/debt-markets), [Nvidia](/research/stack/nvidia), [CoreWeave](/research/stack/coreweave), [Oracle](/research/stack/oracle) and [power/projects](/research/macro/power-and-datacenters) should incorporate the IMF transaction map and KC Fed maturity evidence, retaining lender-specific waterfalls. Ohio's conditional cap is already filed; preserve that correction. Date Jupiter's loan quotes and notice to September. S&P's **27 Aug primary release** confirms negative FOCF expected in **2026–27 across its six named hyperscalers**, recovery **2029**: the front page's secondary/B attribution can become primary/A **for the stated forecast**, not for realized cash flow. [S&P original][sp-capex]
- **Lab accounts and IPOs:** [Anthropic](/research/stack/anthropic), [OpenAI](/research/stack/openai) and [public markets](/research/macro/public-markets) should retain reported/unverified labels and gross/net reconciliation, present IPO crowding-out as a mechanism rather than a demonstrated inability to clear, and distinguish full market capitalization from new float/proceeds. Do not promote the unverified 85%-GAAP correction, 88% inference margin or precise token-growth pair into facts.
- **Forecast comparisons:** add the following evidence beside [the register](/forecasts/register). Its horizons are **17 Apr 2027 / 17 Oct 2027 / 17 Oct 2028**, with origin **17 Oct 2026**. No matched market was found for any of the five questions, and this verification alone does not warrant changing their probabilities.

| Forecast | Report number, 6 / 12 / 24 months | Verified market number beside it | Implication |
|---|---|---|---|
| M1 | **12% / 25% / 38%**, daily-rebalanced eight-name basket falling 40% from **16 Oct close** | Polymarket **6.9% by 31 Dec 2026 / 16.5% by 30 Jun 2027**, for P3's different event | Neither event contains the other. Peak-based single-name conditions differ from a forward basket; dates differ. ORCL and CRWV together have **25% of daily target weight**; their earlier peak losses do not count toward M1. A risk-on/risk-off comparison is useful; Grok's formal “harder contract” and rich/cheap deductions are not established. |
| F1 | **10% / 20% / 35%**, actual distress or a completed acquisition below **$40/share** | CoreWeave **1,118bp** dated Wood quote; reported high **1,194.5bp**; Reuters model **64% over five years** | CDS widening and an ordinary sub-$40 stock print do not resolve F1. For illustration, **1,000–1,194.5bp**, **40% recovery**, flat hazard gives **~8.0–9.5% six-month**, **28.3–32.8% two-year**, **56.5–63.0% five-year** risk-neutral probabilities: `1 − exp[−(spread/0.60) × years]`. This is conditional arithmetic, not an estimated physical default curve; the report's broader event and market risk premia also differ. It does not establish that six-month 10% is too low. |
| F2 | **12% / 28% / 42%**, Oracle below IG at any major agency | Oracle five-year CDS **288bp** in Wood's dated quote, versus Reuters **near 300bp**; S&P **BBB− stable** last verified action | CDS does not directly price a rating downgrade. Stable outlook and debt stress both matter; no matched fallen-angel probability was verified. Do not turn 288bp into 12-/24-month downgrade odds without a separate model. |
| F3 | **20% / 40% / 60%**, disclosed **$10bn+** binding compute cut/cancellation or **≥one-year** deferral; supplier delays excluded | Kalshi OpenAI IPO-first midpoint **17%**, Anthropic bid/ask **93–94¢**; Metaculus lock-up caps **$1.84tn/$1.98tn** for OpenAI/Anthropic | IPO order and valuation do not measure binding-compute cancellation. These are financing-context counterpoints, not a market probability for F3. Patel's optimistic revenue forecast likewise cannot resolve undisclosed legal contract terms. |
| Headline | **6% / 13% / 20%**, M1 plus actual distress at **two distinct obligor groups** | Polymarket **6.9% end-2026 / 16.5% mid-2027**, requiring any three P3 triggers | Polymarket need not include insolvency; the report requires two credit failures. The events are not nested. The register already defines distress and groups SPVs by sponsor; **92¢ project debt, downgrades and F1's acquisition branch do not count**. Correlation merits review but does not numerically derive the conjunction probability. |