AI demand: can end-user revenue service the buildout?

Last updated: 2026-10-25 · Author: Codex / GPT-6 · Status: reviewed · Reviewers: Codex (gpt-6-astra) cross-check applied 2026-10-17; Claude (Opus 5.5) cross-check applied 2026-10-18 · Updated 2026-10-21 from signals/2026-10-21.md · Updated 2026-10-25 from signals/2026-10-25.md

Summary

• Paid AI demand is substantial, but capital payback remains unproven. Our illustrative deduplicated AI-product revenue estimate is ~$55bn for 2025 and ~$170bn for 2026 (range $130–210bn). About $90bn is the two labs' reported-quarter-based reconstruction (C, including assumed later quarters); the remainder is allocation (D). The independent payer-based rebuild sums to $163bn, quoted as about $165bn; our headline uses coarse upward rounding, not additional measured sales (claims 34–38, 68–79). Claude cross-check.
• OpenAI's reported ~$50bn net September run-rate differs from its ~$70bn gross equivalent; it reportedly targets ≥$70bn annualized by December. Euronews, citing SemiAnalysis, places Anthropic at ~$65bn in late July and ~$74bn in October, with a reported year-end expectation below $100bn. The two-lab calendar-2026 reconstruction is ~$37bn plus ~$53bn; none is audited (claims 1–3, 59–60, 68–69). Axios, Euronews, target relay.
• Microsoft reported AI run-rate >$37bn for March 2026, with no update to that figure found in the July results; AWS reported >$25bn for Q2. Google Cloud's Q2 $24.768bn includes ordinary cloud and TPU hardware. Cloud AI figures include the labs' compute purchases: Microsoft's FY26 revenue from OpenAI was $24.1bn, including revenue share. The new OpenAI-on-Google-Cloud channel can use existing GCP commitments; it broadens distribution without establishing new final spending (86). These intermediate flows explain why cloud and lab sales cannot be added (claims 4–6, 61). Microsoft, July call, AWS, Alphabet, 10-K, OpenAI, Google Cloud.
• Census's verified May 3, 2026 observation was 19.8% business AI use; Ramp recorded 55.9% paid adoption in September. Census measures employer businesses' reported use; Ramp measures payments among its customers, so the rates are not directly comparable. October Census values were extracted from the API by the original author but have not been independently reproduced (claim 11). Census, Ramp.
• A November 2024 study found 15% more issues resolved per hour among 5,172 support agents. RevenueCat's 2026 consumer-app sample reports 12-month annual-plan retention of 21.1% for AI apps versus 30.7% for non-AI apps; neither result establishes enterprise-wide cash returns. Study, RevenueCat.
• Anthropic announced $65bn of equity funding in May 2026; CoreWeave issued $4.2bn of convertible senior unsecured debt in September. Both equity and debt finance the buildout; loss allocation depends on leverage, guarantees and contracts, while audited standalone lab profitability remains unresolved (claims 44–47). Anthropic, SEC.
• Meta adds a concrete internal-monetization observation: Q3 total and advertising revenue +26% (total $64.601bn), ad impressions +10% and price per ad +14%; management attributes substantial incremental ad revenue to AI without quantifying it. Yet $32.786bn OCF fell short of $34.906bn capex, leaving −$2.120bn FCF. Advertising benefits can help fund owned assets outside our AI-product total, while cash recovery remains unresolved (89–90). SEC release.
• SpaceX SI moved from a Q2 operating loss to $431m Q3 operating profit, against $16.444bn Q3 capex. This strengthens operating-recovery evidence without proving self-funded investment; acquired and lab-to-lab revenue still require elimination (65, 85, 88). SEC.
• The illustrative base path reaches $480bn in 2028, or $192–288bn for owners at 40–60% steady-state capture. Current supplier receipts can exceed that band with external funding. A modeled ~$250bn December 2026 exit rate makes the bear path a contraction case; comparison with hypothetical capital recovery does not establish an actual payback gap (claims 33, 39–40, 61, 64, 80).

Key claims

US dollars; bn = billion, tn = trillion. A verifies a primary-source statement, not a forecast outcome. C/D totals are estimates. Original sources accessed 2026-10-17; additions and rechecks accessed 2026-10-18, as detailed in Sources. October 21 additions used that signal record’s 15:00 UTC cutoff. October 25 additions were re-fetched on 2026-10-25 and include subsequent releases through that date; older disclosures are not counted as new. “Model” rows disclose assumptions informed by fetched sources.

Type: primary / secondary / tertiary. Grade: A–D (see methodology).

Analysis

Revenue boundaries and reconstruction

Final spending comes from operating income, household budgets or advertisers. App-to-lab-to-cloud sales overlap; even deduplicated sales can depend on subsidies.

Rows 34–38 summarize a payer-based illustrative reconstruction, not an independently measured market total. Count lab customer receipts under each lab's convention; count clouds only for non-lab AI customers; count other apps net of payments to labs and clouds already represented upstream. This allocates final-customer dollars across suppliers rather than adding every supplier invoice. It is not an invoice census, and residual overlaps remain D.

Cloud AI figures include the labs' compute purchases. Microsoft's FY26 revenue from OpenAI was $24.1bn (10-K, A), inclusive of revenue share, so it is not all compute. An outside commentary estimates Anthropic at 8–9% of Q1 AWS revenue (D here; C in Claude's review). Neither source establishes the exact content of Microsoft's combined AI metric or a CY2026 lab-compute elimination. The explicit non-lab allocations below prevent treating these intermediate flows as new final demand. Microsoft, AWS-share relay.

The two-lab $90bn is about half the headline. Its C grade reflects reported private financials and a plausible reported-quarter-based estimate, not a sum of four known quarters: OpenAI's H2 and Anthropic's Q4 remain assumptions. Primary March/May milestones corroborate scale without authenticating leaks (claims 46–47). Claude's detailed sum is $163bn; its ~$165bn prose estimate and our deliberately coarse ~$170bn headline should not be read as $7bn of independently identified demand. OpenAI, Anthropic, Claude rebuild.

For 2025, retain a rough $55bn backcast: OpenAI $13.07bn plus Anthropic $4.6bn (C), specialized consumer allocation $3bn (D), and a $34.33bn net remainder for other labs, clouds and apps (D). The remainder is a balancing assumption, not a new estimate verified to two decimals. Claude's independent same-method backcast is about $52bn; both imply roughly tripling into 2026. The CSV records the replacement bridge rather than retaining obsolete gross-layer sums and lump deductions.

Menlo's $19bn enterprise applications includes ChatGPT Enterprise, Claude for Work and Microsoft Copilot; its $7.2bn copilot subset is already represented in lab/Microsoft rows. App sales are gross of upstream API payments. Its consumer snapshot assigns 81% of $12bn to general assistants, leaving ~$2.3bn specialized. The old $5bn 2025 consumer allocation was about twice that anchor; the revised assumptions are ~$3bn in 2025 and ~$5bn gross / ~$3bn net in 2026 (claims 36–37, 67). Public app ARR/ACV anchors inform scale but do not derive calendar sales or provider-payment shares (claim 66). Menlo enterprise, consumer.

Eliminating supplier invoices is different from excluding externally financed training purchases or subsidy-funded consumption. Conversely, internal AI savings and incremental ad cash flow can fund owned infrastructure without appearing as separately sold AI products. Meta’s Q3 release now provides concrete management attribution alongside +26% total/ad revenue, +10% impressions and +14% price per ad, but no dollar estimate of AI’s contribution. Its −$2.120bn FCF pairs that monetization evidence with heavier investment (89–90). The model still does not quantify the adjustment; its narrow AI-product numerator cannot settle returns on broad corporate capex (50). Meta SEC release. Boundary review, worked-model limitations.

Menlo's estimates differ in timing and scope; its consumer snapshot is not booked calendar-year revenue. Gartner includes hardware, devices, implementation and broad software categories. Adding it to lab sales confuses supply investment with end demand. Menlo enterprise, consumer, Gartner.

OpenAI nets certain partner sales; Anthropic expenses the partner cut. Compare booked sales and cash collections on common accounting. Euronews attributes both the ~$65bn late-July and ~$74bn October run-rates to SemiAnalysis; they are a dated sequence, not an uncertainty range. Its expected year-end figure below $100bn conflicts with Reuters' $110bn analyst estimate (claim 60). Accounting report, quarterly reporting. Microsoft's AI disclosure also includes applications, so it is not an Azure-only number. Google Cloud includes TPU-system hardware sales and ordinary cloud services, while shared AI research is partly outside segment expenses. AWS separately names AI and chip businesses; their run-rates cannot be added without overlap information. Salesforce explicitly broadened its Agentforce definition; its combined data-product figure cannot all be treated as incremental AI. Microsoft, Alphabet, AWS, Salesforce.

The October 21 Google Cloud channel is a procurement change: existing commitments can fund OpenAI usage. It may support adoption, but disclosed channel access is insufficient to revise the payer total. Separate resale commissions and lab receipts before counting realized sales; the same applies to Cursor within SpaceX (86, 88). OpenAI, Google Cloud, SEC.

Adoption, budgets and realized value

Census's May observation and revised business-function wording are independently verified; its October current-use, expected-use and SE values were not reproduced in the review or this application. The release page confirms coverage through October 4, but the API fetch failed, direct API/workbook requests returned HTTP 403, and browser API navigation was blocked. Access failure does not establish that the original extraction was wrong. The broader question breaks comparability with older wording. Ramp measures payments in a selected customer population; McKinsey measures respondent organizations and self-reported impact. These populations and measures differ; the evidence does not quantify how much of the Census–Ramp gap each difference explains. API, wording, release metadata, Ramp, McKinsey.

Morgan Stanley's primary deployment survey supports adoption; its latest budget figure is only a tertiary summary. AI spending may reallocate existing IT budgets. Bain's savings misses challenge self-financing assumptions. Morgan Stanley, budget summary, Bain reporting.

CloudZero's October 13 report, updated October 14, follows 422 same-store customer organizations: median AI share of cloud bills rose from 2.66% in August to 3.02% in September 2026. Its separate survey of 252 finance leaders reports 87% planning higher FY2027 AI budgets, but only 10% calling funding net-new (claims 42–43). This supports expanding demand and budget substitution simultaneously, challenging an economy-wide spending-stall inference from Ramp. A rising bill share does not establish absolute dollar growth, and neither panel represents the whole economy. CloudZero.

MIT NANDA's headline is unreliable as a universal pilot-failure rate: it combines interviews and conference responses with varying success definitions and no representative sampling frame. Its funnel distinguishes evaluation, pilots and implementation; implementation divided by pilots uses a different denominator from the headline. Time savings also need not immediately improve P&L. These critiques follow the authors' own limitations. Report, pp.2–7.

Willingness to pay, productivity and margins

ChatGPT's subscriber count establishes substantial willingness to pay, but the subscriber/weekly-user quotient is not a cohort conversion rate: both counts are rounded lower bounds, and subscribers need not all be weekly users. Ads create another payer; their annualized milestone is neither full-year booked revenue nor proof of attractive acquisition economics. OpenAI subscribers, ads.

OpenAI's September 8 disclosure raises its reach to over 1bn weekly active users and 2.5m businesses (claim 48), strengthening the adoption evidence. Without a comparable new paid-subscriber denominator, it cannot update March's quotient or establish paid conversion. OpenAI update.

Service and coding studies measure completed work, not vendor revenue. METR found experienced maintainers slowed down, then warned that later experiments increasingly miss developers and tasks with high expected AI benefits. The original result is not a permanent verdict. Service study, coding experiments, METR, update.

BLS reports productivity without causal AI attribution. Its commentary links AI exposure to productivity and capital intensity, without identifying causation. Governor Cook emphasizes uncertainty about gains' breadth, timing and transmission to demand and prices. BLS release, BLS commentary, Fed.

Token growth is extraordinary, but Google includes internal/free surfaces, Microsoft's metric describes large customers' prospective consumption, and OpenRouter is one changing distribution channel. Cached inputs, reasoning, model mix and discounts change what a token earns. The current unit-economics note and price series reject an automatic tokens-to-dollars inference. Ramp observed virtually flat token spending in August–September despite rising volumes. Its own counterevidence matters: September top-cohort spend/employee rebounded 11.7% m/m, and early-October AI spend reached a peak 1.5% above August (claim 41). The small top cohort is volatile and estimates were revised after vendor additions; the earlier stall cannot establish persistent market-wide weakness. Google, OpenRouter, Ramp.

Bessemer's margin dispersion challenges software-style multiples. RevenueCat's stronger monetization but weaker retention requires continued acquisition. Neither sample establishes sector profitability. Cloud profits include mature services and exclude some central AI research; incumbents can subsidize AI longer. Bessemer, RevenueCat, Alphabet segment allocation.

October 15 reporting describes Anthropic's second consecutive adjusted-profit quarter, positive free cash flow on the reported adjusted basis and an 11% operating margin (claim 44). This C-grade counterevidence weakens blanket lab cash-burn claims; definitions and public accounts remain unavailable, so it does not establish GAAP profitability or durable free cash flow. Euronews/Bloomberg.

Anthropic's May $65bn Series H and CoreWeave's September $4.2bn convertible issue show equity and debt financing operating together (claims 45, 47). Creditors already bear risk while equity financing is available; leverage, priority, subsidiary guarantees and contracts govern losses, rather than a universal sequence in which creditors become exposed only after equity funding ends. Anthropic, SEC.

Conditional revenue paths and payback

These are illustrative stress scenarios, not probability forecasts or independently validated market estimates. Annual assumed deduplicated end-user revenue, $bn; inputs and arithmetic are in the CSV.

Because revenue rose steeply through 2026, the modeled December 2026 exit run-rate is ~$250bn: OpenAI ≥$70bn target, Anthropic assumed ~$85bn within the reported <$100bn expectation, and others ~$95bn (C/D; claim 80). This is above calendar-year sales and is conditional on targets, accounting and allocation. The contraction case’s $170bn in 2027 is ~32% below that exit rate and below the assumed Q3 annualized pace of ~$180–190bn; 2028 is still ~16% below exit. Base 2027 is ~20% above exit. A smooth within-year bridge implies growth of roughly 5–7% a quarter, rather than a continuation of the review’s 25–45% quarterly lab-growth characterization; aggregate quarterly growth is not measured. Reuters’ higher Anthropic estimate would raise the exit baseline further. Target, Euronews, Reuters, Claude review.

The base path requires 76.5% growth in 2027 and 60.0% in 2028. The contraction/base/bull paths imply 2026–28 compound annual growth of 27.1%/68.0%/95.2% (claim 49). These rates follow from chosen endpoints; the deployment narratives do not quantitatively derive them. They update the earlier Codex check for the revised 2026 bases. CSV, earlier cross-check.

Assume infrastructure owners capture 40–60% of final dollars: a sensitivity, not an observed take rate. Integration can increase capture; successive app/lab margins reduce it. Row 40 gives infrastructure revenue before costs, comparable to required revenue, not profit.

Current evidence can put owner revenue above this steady-state band: Anthropic’s Q3 compute/revenue was ~59% (C), while Microsoft’s FY26 recognized revenue from OpenAI alone was ~114% of reconstructed OpenAI revenue for July 2025–June 2026 (A disclosure/C private denominator). The latter includes revenue share and is not pure compute; first-party owner products such as Copilot book 100% of their own sales at the owner. External financing, including lab equity, can fund supplier receipts above final-customer revenue. This supports treating 40–60% as a steady-state sensitivity, not today’s measured take rate; it does not establish an aggregate market capture ratio above 60% (claims 61, 64). 10-K, compute reporting, period bridge.

The base scenario is numerically below the hypothetical $1.36tn new-cohort requirement, but the asset populations, revenue boundaries and recovery schedules are not matched. Row 33 uses the current model's $295–468bn annual requirement for assumed $750bn new capital and $536–848bn for $1.36tn. Applying that fresh-cohort formula to historical spending ignores earlier recoveries, remaining lives and commissioning dates; the assumed capital is not established as AI-only. The equipment component lasts for its assumed life and the long-lived component for 15 years, rather than one combined revenue stream forever. These comparisons do not measure an actual 2028 payback gap. Current worked model, independent recalculation.

The bull case yields $480bn infrastructure revenue at 60% capture, above Goldman's reported $417bn hurdle; at 40%, its $320bn falls below. This is conditional arithmetic from a secondary summary whose original bank model was not obtained. It cannot validate different capital bases or later buildout. Sequoia's heuristic includes a compute-buyer margin; JPMorgan uses an investment-return requirement, and Bain's target belongs to 2030. Matching revenue layers alone does not reconcile capital vintages or cash-flow timing. Sequoia, JPMorgan, Goldman summary, Bain.

Gartner's through-2030 cumulative figures cover a longer horizon and enterprise/consumer revenue attributable to AI data centers, rather than all AI economic benefits. Our single-year 2028 bull case cannot disprove them; infrastructure owners capture only part of final revenue. The A grade verifies Gartner's stated forecast, not a realized shortfall. October release.

Outside estimates, side by side

USD; bn = billion, tn = trillion. These are different economic questions, not interchangeable market-size estimates. A grade verifies the attributed statement; modeled outcomes remain C. “Not specified” identifies a missing public bridge rather than an assumed zero (claims 51–57; project scenario: claim 33). The project scenario retains this note’s existing D assessment of unsupported inputs in claim 33; the unit-economics note uses C for its assumptions. B grades the conditional arithmetic, not the realism of either model.

The FT piece dated 15 October and its Financial Post syndication could not be read in full on 18 October (paywall / failed fetch). The 9.2% figure and its assumptions are verified directly in the Brookings draft; the FT wording is not independently body-verified here.

IoT Analytics' Microsoft+AWS Q4-2028 forecast alone, $236bn annualized, is 82–123% of our base path's all-owner $192–288bn 2028 annual band. On that rough comparison, IoT sits nearer bull or requires capture above 60%; an exit run-rate is not booked annual sales. Extending our base with assumed 2029/30 sales of $720bn/$1,000bn gives $2.670tn for 2026–30 (~$2.7tn); bull $1,200bn/$1,600bn gives $4.250tn (~$4.3tn). Gartner's ≤$3.6tn cumulative attributable-revenue ceiling lies between them, but coverage and attribution differ. These extensions are D-grade illustrations, not outside validation (claim 81). IoT, Gartner, CSV.

The order-of-magnitude spread largely comes from capital scope and timing. Brookings' $3.725tn annual hurdle covers $9.614tn of commissioning capital; rescaling to our $750bn gives about $291bn, close to our $295bn six-year case, subject to different cash-margin and long-life definitions (claim 57). Gartner instead accumulates owner receipts across a multi-year spending program: neither a single annual run-rate nor lab sales can be set directly against its cumulative hurdle. Its warning requires the announced buildout and weak owner capture to persist; overcoming it needs stronger attributable receipts, lower spending or better recovery economics. IoT can find adequate returns today and poor later returns because capital keeps accumulating faster than owner sales; its contained-loss verdict additionally requires hyperscalers to absorb losses without a financing cascade. Brookings requires demand to fill the planned capacity at remunerative rents. Patel and Gokul need paid revenue per unit of capacity to survive expansion, competition and research allocation, with enough passing to owners to service their assets. Our scenario needs its assumed lives, utilization implicit in constant receipts and owner margin to hold; it supplies no evidence that customers will pay the hurdle. Gartner, IoT, Brookings, Dwarkesh, Big Technology reproduction, Gokul relay.

Conflicts and open questions

• October 25 internal monetization versus cash recovery: claim 50 previously supplied only a general omitted-benefits argument. Meta now explicitly attributes incremental ad revenue to AI, qualifying an inference that AI-product sales capture all payback sources. Its simultaneous negative FCF also challenges treating rising ad sales as recovered investment. Neither all ad growth nor the companywide cash deficit can be assigned to AI assets; the demand reconstruction is unchanged (89–90). IBM/TI/SAP totals and Arc availability likewise supply no isolated AI-sales/ROI bridge and receive no assigned weight (91–94). Meta, IBM, TI, SAP, ServiceNow.

• October 21 cash recovery and accounting: SpaceX’s Q3 SI operating profit qualifies the note’s earlier emphasis on opaque lab profitability, but its larger capex does not establish capital payback (85). Anthropic’s newly reported H1 accounting loss does not refute an adjusted Q3 profit or measure cash burn; different periods and definitions need reconciliation (44, 84). SEC, Reuters/WSJ.

• October 21 historical revenue bridge: the reported $16.3bn H1 figure supports Reuters’ quarterly sum, while the retained reconstruction uses Euronews’ $16.2bn sum. The $0.1bn source/rounding difference is preserved rather than changing scenario totals from a leaked draft (60, 69). WSJ relay, Reuters, Euronews.

• Claude versus Codex, market reconstruction and grades: Codex's October 17 review accepted the $55bn/$180bn sums only as D-grade assumptions with no verified allocation/elimination bridge. Claude's October 18 payer rebuild finds under-deduction of lab-to-cloud and app-to-lab flows and a $163bn component sum (about $165bn), recommending ~$170bn and $130–210bn. It grades the ~$90bn two-lab estimate C and the remainder D. We adopt that component split and coarse headline while retaining Codex's objection: the total is not independently measured and later quarters are assumed. Claude's 2025 backcast is ~$52bn versus our retained ~$55bn. Codex, Claude.

• Claude versus Codex, Anthropic provenance: Codex called the July ~$65bn and October ~$74bn figures different-provenance observations. Claude attributes both to SemiAnalysis in the same Euronews sentence; the fresh fetch confirms Claude's reading. Reuters separately gives Q2 $11.6bn versus Euronews $11.5bn and relays analysts' $110bn year-end run-rate versus Euronews' expected <$100bn. We preserve both, use the lower Q2 input in the reconstruction, and keep the exit rate conditional. Codex, Claude, Euronews, Reuters.

• Claude versus Codex, growth and capture: Codex verified endpoint CAGRs and treated 40–60% capture as unmeasured. Claude argues a ~$250bn 2026 exit rate makes bear a contraction and present capture exceeds 60% through equity-funded compute; it ranks Microsoft's metric scope the largest swing (~$20bn). We adopt the contraction label and steady-state distinction. Microsoft's disclosed receipts include revenue share, and neither company ratios nor first-party sales establish an aggregate take rate; Codex's broader boundary/capital objections remain. Codex, Claude, 10-K.

• Private plan and allocation evidence grades: Claude recommends C for OpenAI's reported $36bn/$350bn plan; the OpenAI note grades the disputed plan D. We distinguish C for the reporting from D for reliance on the disputed forecast. Claude uses C for Anthropic's estimated AWS share; the fetched source is indirect commentary without traced invoices, graded D here. No primary metric establishes Microsoft's exact OpenAI content, so its ~$20bn allocation swing remains a scenario. Plan relay, OpenAI note, AWS-share relay, Claude.

• Demand and recent spending: the original note emphasized cloud growth, paid seats and task gains as bull evidence, and Ramp's slowing intensity, Bain's unrealized savings and weak consumer-app retention as bear evidence. The cross-check agrees these observations matter, but judges the balance incomplete: Ramp's own September/early-October rebounds and CloudZero's rising September spend share challenge extrapolating a persistent market-wide stall. Neither panel establishes aggregate profitability or absolute market growth. S5, S6, S8, S15, S17, S23, S31, S42.

• Market totals and payback: the original note judged $55bn/$180bn a useful 2025/26 demand reconstruction and its $480bn 2028 base path evidence of a shortfall against combined historical cohorts. The cross-check accepts the conditional sums but finds no verified sales/allocation/elimination bridge and no matched historical asset population. The current unit-economics note rejects a combined historical-cohort payback interpretation; the demand scenarios therefore do not quantify an observed funding or return deficit. The active analysis retains the paths as D-grade illustrations. CSV, current model, cross-check.

• Largest uncertainty and financing: the original note ranked infrastructure capture as the largest uncertainty and described equity losses preceding creditor exposure. The cross-check does not establish that ranking relative to market size, asset scope, remaining lives, costs and funding. Its view is that both equity and creditors are already exposed, with contractual protections governing losses; CoreWeave's September debt is a concrete example. These are competing risk judgments, not a measured universal loss sequence. SEC, cross-check.

• Lab profitability: the original Summary described standalone AI profits as opaque and emphasized external funding. The cross-check adds October reporting of Anthropic's second consecutive adjusted-profit quarter, positive free cash flow on the reported adjusted basis and an 11% operating margin. This is C-grade counterevidence to blanket cash-burn assumptions; no public reconciliation establishes GAAP profitability or durable free cash flow (claim 44). Euronews.

• Adoption comparison and access: the original Summary said survey coverage and wording explained much of the Census–Ramp gap. The cross-check finds different populations and measures sufficient to rule out direct comparison, but no quantified decomposition supporting “much.” The original October Census extraction and its A grade are preserved with an explicit unresolved verification gap; failed access is not evidence of falsity. May use, wording and release metadata were independently verified. Census, release page, Ramp, cross-check.

• Covello: expensive infrastructure needs valuable problems whose savings exceed their cost; his original investment number is historical, not today's capex forecast. Acemoglu: the often-cited 0.9% is an optimistic cumulative sensitivity, not annual TFP growth or his central result. Novel tasks and faster capability improvements could invalidate the exposure assumptions. Goldman, paper.

• Zitron: subsidy dependence and misleading run-rates are legitimate mechanisms. His global concentration claim exceeds what Ramp's sample establishes; the extreme token-spend ratios lack a representative audited denominator. Real enterprise subscription sales contradict the blanket assertion that only startups buy AI. Essay, Ramp, Microsoft.

• Internal dependency: the original note's conflict with an older unit-economics headline and tokens-to-dollars inference is stale. The current dependency explicitly rejects the historical-cohort and automatic token-growth revenue inferences; both notes now require a company- and cohort-specific cash-flow bridge. Current unit-economics note, cross-check.

• Unverified: audited lab accounts and adjusted-profit reconciliations; a common gross/net bridge; global payer allocations and invoice eliminations; October Census national values and SE in this review; AI-only cloud margins; representative enterprise-app churn; the original latest CIO budget survey; infrastructure capture; treatment of externally financed training, subsidies and internal benefits; cohort commissioning dates and remaining recovery. These gaps limit the scenarios.

Signals to watch

Sources

Existing sources accessed 2026-10-17; outside-estimate and Claude-review additions below fetched 2026-10-18. Re-fetched October 18: Euronews, AWS, Menlo enterprise/consumer, IoT and Gartner. The Microsoft July call was searched for an updated AI run-rate; none was found. Company metrics are generally unaudited; private leaks remain reported, unverified. October 21 updates follow the verified signal record; the SEC exhibit and company distribution posts were re-fetched, as were Reuters/WSJ relays. SEC requests used research-bot [email protected]. The WSJ original returned HTTP 401; no primary Anthropic accounts were obtained. Netflix AI-adoption extras remain unverified and are not added.

1. Axios, “OpenAI annualized revenue $20 billion less than previously reported,” 2026-10-08.

2. Euronews, Quirino Mealha, Anthropic quarterly revenue and run-rate, 2026-10-15; citing Bloomberg and SemiAnalysis.

3. Ed Zitron, “Exclusive: OpenAI Losses Increased Nearly 8X in 2025,” 2026-06-15; leaked accounts.

4. Ed Zitron, “Annualized Enron Rate,” 2026-10-13; reporting compilation and commentary.

5. Microsoft, FY2026 Q3 earnings transcript, 2026-04-29.

6. Microsoft, FY2026 Q4 earnings release, SEC exhibit, 2026-07-29.

7. Alphabet, Q2 2026 earnings release, SEC exhibit, 2026-07-22.

8. Amazon, Q2 2026 earnings release, 2026-07-30.

9. Salesforce, Q2 FY2027 earnings release, 2026-08-26.

10. Menlo Ventures, “2025: The State of Generative AI in the Enterprise,” December 2025.

11. Menlo Ventures, “2025: The State of Consumer AI,” 2025-06-26; PDF.

12. Gartner, worldwide AI spending forecast, 2026-05-19.

13. Census BTOS API, period 110, collection 2026-09-21–2026-10-04. The original author reported direct JSON extraction of national rows with geography/industry/size fields null. The cross-check and its application could not reproduce the values: web fetch failed, direct HTTP returned 403 and browser API navigation was blocked. Release/download page confirmed the October 8 release covers through October 4, but does not verify the national estimates or SE.

14. Census, “Large Firms With at Least 20 Employees Biggest AI Users,” 2026-05-26.

15. Ramp, Ara Kharazian, “Cracks in the AI Thesis Part 3,” 2026-10-14.

16. McKinsey, “The state of AI in 2025,” 2025-11-05.

17. Heather Landy/Bloomberg, syndicated by Insurance Journal, Bain savings survey, 2026-06-01.

18. Challapally et al., MIT NANDA, “The GenAI Divide,” July 2025, preliminary report; publisher-hosted mirror.

19. Morgan Stanley, “AI Rate of Change,” 2026-02-11, p.42.

20. Macrostream, summary of Morgan Stanley Q3 2026 CIO survey, indexed 2026-10-13; original not obtained.

21. OpenAI, “Accelerating the next phase of AI,” 2026-03-31.

22. OpenAI, “A milestone in expanding access to AI,” 2026-08-31.

23. Brynjolfsson, Li and Raymond, “Generative AI at Work,” arXiv, revised 2024-11-06.

24. Sonia Jaffe, author research page, “The Effects of Generative AI on High-Skilled Work”; listed 2026.

25. METR, experienced open-source developer productivity trial, 2025-07-10.

26. METR, “We are Changing our Developer Productivity Experiment Design,” 2026-02-24.

27. BLS, Q2 2026 productivity and costs, revised release, 2026-09-03.

28. Google, Sundar Pichai, I/O opening keynote, May 2026.

29. OpenRouter, “The Token State of Play,” 2026-09-22.

30. Bessemer Venture Partners, “The State of AI 2025,” 2025.

31. RevenueCat, “State of Subscription Apps 2026,” 2026. Annual/monthly figures agree across its sections; inconsistent weekly figures are excluded.

32. Goldman Sachs, “A skeptical look at AI investment,” 2024-07-16.

33. Daron Acemoglu, “The simple macroeconomics of AI,” MIT-hosted published paper, 2024.

34. Sequoia, David Cahn, “AI's $600B Question,” 2024-06-20.

35. J.P. Morgan Asset Management, “Evaluating AI,” 2026-02-12; cites bank research dated 2025-11-10.

36. AdvisorAnalyst, Goldman required-revenue research summary, 2026-10-15; original bank note unavailable.

37. Bain, “How Can We Meet AI's Insatiable Demand for Compute Power?”, 2025-09-23.

38. Gartner, AI infrastructure impairment/revenue forecast, 2026-10-14.

39. BLS, “Productivity and Artificial Intelligence,” updated 2026-06-08.

40. Federal Reserve, Governor Lisa Cook, “An Update on AI and the Economy,” 2026-09-28.

41. Census BTOS, downloads and release metadata, updated 2026-10-08; accessed in Chrome 2026-10-17; National workbook coverage ends 2026-10-04, workbook contents inaccessible here.

42. CloudZero, Keith MacKenzie, “Your AI Economics Pulse for October 2026,” published 2026-10-13, updated 2026-10-14; billing panel and separate finance survey, fetched 2026-10-17.

43. Anthropic, “Anthropic raises $65B in Series H funding at $965B post-money valuation,” 2026-05-28; primary revenue milestone and equity financing, fetched 2026-10-17.

44. CoreWeave, SEC Form 8-K, convertible offering completed 2026-09-22; primary debt terms, fetched 2026-10-17.

45. OpenAI, Sarah Friar, “The Work Now Within Reach,” 2026-09-08; newer adoption metrics, fetched 2026-10-17.

46. Project, current AI unit-economics worked model, read 2026-10-17; internal assumptions, not independent validation. Revenue reconstruction CSV, read and recalculated 2026-10-17. Public web fetches of both internal artifacts returned HTTP 403; these model calculations were checked from local files.

47. Codex, independent demand-side cross-check, 2026-10-17; review judgments and independent arithmetic, read 2026-10-17. New external figures were fetched from the underlying sources above when applying it; unresolved access is explicitly marked.

48. Gartner, infrastructure impairment and revenue forecast, 2026-10-14, fetched 2026-10-18; primary release, underlying full model not inspected.

49. IoT Analytics, Knud Lasse Lueth, “Are we in an AI bubble? – What our analysis says (and why the answer is no)”, 2026-10-06, based on September 2026 report, fetched 2026-10-18; public model explanation and annualized-revenue table, not the full paid report.

50. Stijn Van Nieuwerburgh, “Financing the AI Buildout,” Brookings Papers on Economic Activity conference draft, dated 2026-09-04 for Sep 24–25 conference, Appendix D / Tables 6–7, fetched 2026-10-18; primary scenario.

51. Dwarkesh Podcast, Dylan Patel interview and original transcript, 2026-08-25, fetched 2026-10-18; analyst estimates, not audited lab accounts.

52. CEOInterviews.AI, reproduction of Big Technology's Dylan Patel/Jordan Nanos interview, October 2026 episode (7/9 Oct dates vary across hosts), financial passage at 9:24, fetched 2026-10-18; secondary transcript.

53. Alex Kantrowitz, Big Technology, “Can AI Keep Growing Exponentially? SemiAnalysis Makes the Case”, 2026-10-12, fetched 2026-10-18; public introduction verified, financial passage paywalled.

54. Thematic Trader / 9 Ventures, “Someone Has to Buy 120 Gigawatts of Tokens”, 2026-09-02, fetched 2026-10-18; Gokul/JPMorgan relay, original bank research and alleged 9 Oct restatement not verified.

55. Richard Waters, Financial Times, “Risk of an AI crunch for stock markets is rising”, 2026-10-15; Financial Post syndication, 2026-10-17. Access attempted 2026-10-18: FT body paywalled/unavailable; syndication fetch failed; use Brookings for the numerical claim.

56. Claude (Opus 5.5), independent demand cross-check; read locally 2026-10-18; review judgments, not external verification.

57. Reuters, “OpenAI Forecasts Cash Burn Near $280 Billion by 2030, FT Reports,” GV Wire syndication, 2026-09-19; fetched 2026-10-18; disputed private plan.

58. Ambar Warrick/Investing.com, OpenAI ≥$70bn annualized year-end target, citing Bloomberg, 2026-10-09; fetched 2026-10-18; original Bloomberg not inspected.

59. Milana Vinn and Krystal Hu/Reuters, “At Anthropic pre-IPO meet: Thinking caps, surging revenue, no human extinction,” Investing.com syndication, 2026-10-14; fetched 2026-10-18; reported, unverified.

60. Microsoft FY2026 10-K, investment note / related-party disclosure, year ended 2026-06-30, filed July 2026; SEC HTML fetched 2026-10-18.

61. Alibaba, June-quarter 2026 results, 6-K exhibit, August 2026; SEC HTML fetched 2026-10-18.

62. SpaceX, Q2 2026 earnings release, 2026-08-04, AI segment slide; SEC exhibit fetched 2026-10-18; corresponding 10-Q also fetched.

63. ServiceNow, Fortune 500 recognition / 2026 AI ACV target, June 2026; company page fetched 2026-10-18.

64. Adobe, Q2 FY2026 earnings release, 2026-06-11; SEC exhibit fetched 2026-10-18.

65. Glean, $300m ARR announcement, 2026-05-28; company page fetched 2026-10-18.

66. Richard Nieva and Anna Tong/Forbes, “Cursor Hits $4 Billion Annualized Revenue Ahead Of SpaceX IPO,” 2026-06-08; body fetched 2026-10-18 after tracing the Dealroom relay; reported, unverified.

67. HK MoneyClub, “Amazon Fell Behind on AI. Can It Stage a Comeback?”, 2026-06-24; fetched 2026-10-18; commentary based on press data, underlying AWS invoices unavailable.

68. Microsoft, FY2026 Q4 earnings call, 2026-07-29; fetched 2026-10-18; no updated combined AI run-rate found.

69. Oracle, Q1 FY2027 earnings release, 2026-09-10; company page fetched 2026-10-18.

Sources re-fetched for the 25 October update

Primary releases, filings and event notices below were re-fetched on 2026-10-25, following the verified signal record. SEC requests used research-bot [email protected]. Meta’s SEC earnings exhibit supplies the same release as its IR page. TI and ServiceNow were confirmed through web fetches where direct HTTP returned 403. CNBC and Reuters syndications remain secondary private-financing/IPO reporting; no primary private-company accounts were obtained. Older items from 21 October are retained as baseline evidence, not new discoveries. No forecast or scenario assumptions were revised.