AI unit economics: can the capex pay back?

Last updated: 2026-10-25 · Author: Claude (Fable 5.1 subagent) · Status: reviewed · Reviewer: Codex (gpt-6-astra) cross-check, applied 2026-10-17 · Updated 2026-10-21 from signals/2026-10-21.md · Updated 2026-10-25 from signals/2026-10-25.md

Summary

• Big-four capex totals of $230.9bn and $380.5bn use company-specific definitions and Microsoft's June fiscal years; they are not calendar-year AI-only cohorts. Meta’s October 21 guide lifts the 2026 midpoint sum from $732.5bn to $735bn, versus $750bn retaining Microsoft’s earlier lease-accounting basis. Oracle's fiscal-year and quarterly spending cannot simply be added to calendar-year guidance (claims 1, 3, 5–8, 52).
• Published required-revenue estimates measure different layers, periods and returns. Our illustrative $750bn new cohort requires $295–468bn annually at a 50% owner EBITDA margin and 10% discount rate, depending on a six- or three-year equipment life. An exactly $1.36tn hypothetical new cohort requires $536–848bn; this is not a valuation of previously deployed cohorts (claims 11–14, 53).
• Revenue growth is real: Google Cloud grew 82% to $24.8bn in Q2 2026, and Microsoft disclosed an AI run-rate above $37bn in April. Reuters reports Anthropic's Q3 revenue of $16.8bn and adjusted profitability, alongside separate reporting of substantial historical losses and future compute commitments. SpaceX now discloses Q3 SI operating profit of $431m alongside $16.444bn segment capex. Private financial reports remain unverified; overlapping revenue measures do not establish a sector-wide payback gap (claims 30, 33–34, 44–47, 67).
• Meta pairs monetization with heavier investment: Q3 total and advertising revenue +26%, ad impressions +10%, price per ad +14%, and management attribution of incremental ad revenue to AI coexist with −$2.120bn FCF on $34.906bn capex. The AI contribution is unquantified; companywide growth and cash flow do not establish cohort payback (76–77). SEC release.
• GPU economic life, utilization, operating margin and customer funding are major uncertainties. The illustrative $500bn short-lived pool has $83bn more annual depreciation at three years than at six. Server accounting schedules and rental listings do not establish GPU lifetime returns, and this sensitivity does not verify Burry's company-specific earnings claims (claims 17–24, 53).
• Prices move in both directions: some premium tiers and scheduled Google prices rise, while Sol and Opus prices fall; September's Opus 5.5 input/output cut was 20%. Google's aggregate token growth spans all surfaces, not just paid APIs. Rising usage alone establishes neither rising paid revenue nor a causal Jevons effect (claims 25–30, 48, 57).
• Epoch estimates final released-model runs at 9.6%, 22.6% and 12.3% of R&D compute spending at OpenAI, MiniMax and Z.ai. These are modeled ratios, not direct prospectus disclosures; experiments and unreleased models also involve training. Its historical 2.4x annual run-cost trend would imply billion-dollar runs by 2027 if continued. Accounting depends on ownership, rental, prepayment and lease terms (claims 36–38, 41–45).
• Financing and profitability differ by layer. Nvidia is highly profitable; CoreWeave has positive adjusted EBITDA but material losses after depreciation and financing. Customer advances, supplier credit and new securities fund the build and allocate risk. SpaceX’s SI profit improves the operating-recovery evidence, but group nine-month operating cash remains $34.972bn below capex. Anthropic’s newly reported H1 financing remeasurement is noncash, and its recent adjusted profit does not establish durable free cash flow. New dated credit stress coexists with Nebius/TeraWulf bank facilities; spreads and the older S&P warning do not establish default or a closed financing market (claims 42–47, 49–50, 55, 67–73, 84).

Key claims

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

Analysis

1. The capex payback math

Historical big-four totals are $230.9bn and $380.5bn on mixed periods and bases: Microsoft's fiscal years end in June, and the companies differ in their inclusion of leases and gross versus net purchases (claims 1–7). Meta’s October 21 revision lifts latest 2026 guidance from a $732.5bn to $735bn midpoint, or $750bn keeping Microsoft’s earlier lease-accounting basis (7, 52). The new $135–145bn Meta range raises its midpoint 1.818%; no quantified 2027 capex guide was issued. Microsoft's building-life extension changes reported capex through lease classification without changing investment expectations (claim 41). Oracle's FY26 $55.7bn and Q1 FY27 $28.5bn are different periods; neither can be silently added to calendar-year forecasts. Goldman's ~$800bn is a scale comparison, not a reconciliation to filings. Nvidia's revenue is a supply-side indicator, but Anthropic's Trainium/TPU/GPU mix demonstrates why a Nvidia-only multiplier cannot precisely measure all AI infrastructure (claims 10, 14, 50).

The required-revenue literature uses different populations and methods. Sequoia's Nvidia-sales multiplier estimates end-user revenue at an assumed gross margin; Bain's 2030 cloud scenario uses sustainable capex/revenue ratios, on-prem migration and reinvested savings. JPMorgan's 10% return estimate is not independently reconstructable from its blog. Goldman's infrastructure break-even and ROIC estimates differ from its application-layer estimate, and the original model was not inspected (claims 11–14). They do not form a chronological series of one statistic. Morgan Stanley estimates a financing gap, not revenue: its named funding categories leave $350bn of the stated $1.5tn gap unspecified (claim 15).

Revenue growth is visible at Google Cloud and in Microsoft's AI run-rate (claims 30, 34). Private lab reporting requires separate treatment: OpenAI's Q2 revenue annualises to $26.8bn, unlike a point-in-time run-rate (claims 33, 59). October reporting puts Anthropic's Q3 revenue at $16.8bn and describes adjusted profitability; this is meaningful counterevidence to a blanket lab-loss characterization, but neither an audited result nor proof of free cash flow (claim 46). Anthropic’s company-stated run-rate baseline is also distinct from full-year recognized revenue (claims 44, 58). Lab receipts overlap suppliers' cloud revenue. Available disclosures do not establish a non-overlapping total of AI end-customer revenue, so they cannot quantify a sector-wide payback gap or establish a revenue-doubling deadline. The new Google Cloud channel can use existing customer commitments, so it is distribution evidence without an independently measured addition to final spending (claim 71).

Track customer prepayments, lease commitments, supplier credit and new securities alongside gross capex: they determine who funds assets and when losses reach investors. Oracle's cash flow includes $11.363bn of customer advances, and its capex reconciliation reports $17.966bn net cash outlay versus $28.499bn gross (claim 42). CoreWeave's September convertible shows financing remains available, while adding leverage and potential dilution; capped calls and fees reduce deployable proceeds (claim 43). Reuters' Anthropic prospectus reporting describes historical losses, future compute obligations and customer concentration; the reported Broadcom facility adds acceleration and availability risks (claims 44–45). The June Broadcom financing tranche is another channel, with potential overlap across capex and leases (claim 50). These amounts must not be added without eliminating overlap. The unresolved Project Jupiter lead warrants checking contractual payment timing, not assuming a debt default (claim 56). SpaceX supplies a new primary profitability observation: Q3 SI operating income turned positive while segment capex remained much larger. Its group operating-cash-minus-capex deficit is a different measure and period, not proof that every installed cohort fails to repay (claims 67–68). New credit quotes and S&P commentary are financing-cost and transparency indicators, not completed funding transactions or rating actions (claims 72–73).

Scope also changes the interpretation: Amazon’s investment includes fulfillment capacity. Meta now explicitly attributes incremental ad revenue to AI, alongside +26% Q3 total/ad revenue, +10% impressions and +14% price per ad; it supplies no dollar attribution. Its simultaneous −$2.120bn FCF on $34.906bn capex pairs monetization with continuing investment pressure, without identifying any AI cohort’s return (76–77). Meta. Conversely, Google Cloud profit does not isolate new AI-cohort returns, with shared AI R&D partly recorded outside the segment (claim 51). Both sides limit a comparison of total capex with separately billed AI revenue.

2. GPU depreciation

Microsoft's quarterly short-lived asset mix is a company disclosure, not an audited sector mix. Server and networking schedules at the other filers are not necessarily GPU-specific lives (claims 5, 17–20). Meta's extension added $2.59bn to 2025 net income, 4.3% of reported annual earnings (claim 54). Accounting schedules affect the timing of earnings, not the physical asset's cash payback.

For the hypothetical fully operating $500bn short-lived pool below, zero-residual straight-line depreciation is $83bn annually over six years, $125bn over four and $167bn over three. The $83bn difference is conditional scenario arithmetic, not an observed big-four misstatement. It does not verify Burry's cumulative or company-specific earnings estimates, which require cohort dates, taxes and earnings denominators (claims 21, 53).

Rental evidence cuts both ways: H100 contract rates rose roughly 40%, and A100 listings remain available (claims 22–23). The October 24 OCPI H100 and B200 prints are higher over both seven and thirty days; October 20’s H100 weekly decline remains a dated prior observation. The recovered Silicon display and listed-price aggregators measure different markets and cannot be spliced into OCPI or converted into utilization (claims 22, 74). Listings do not establish achieved utilization or lifetime profitability, and no causal inference about Nebius's extension follows from them. InferenceX's single-model comparison estimates Blackwell's token cost below Hopper's, but the full economic cost assumptions were only partly verified (claim 24). It does not show that every H100 workload survives only under scarcity. Resolution requires realized rental prices, utilization, operating costs and residual values through changing supply conditions.

3. Inference unit economics

Some premium tiers and scheduled Google prices rise, while other models become cheaper. OpenAI's Sol and Anthropic's Opus lines counter a universal frontier-price turn (claims 25–27). September's Opus 5.5 input/output cut was 20%, with cheaper cache reads; the vendor's performance assertions are not independent validation (claim 48). Fixed-capability price declines measured by Epoch and MIT can coexist with this mixed list-price pattern (claim 28). Equivalent capability, workload, caching and service tier must be compared before inferring pricing power or scarcity.

The InferenceX cost estimates are workload-specific, not measured commodity-serving margins. Comparing their open-model costs with closed-model API list prices mixes models, input/output composition and service costs (claim 24). OpenAI's reported positive adjusted gross margin does not imply that inference costs included in that measure exceed sales; R&D losses and cash burn can coexist with positive gross margin (claim 32). Anthropic also has a reported historical gross-margin observation and a newer adjusted-profitability report, all C-grade private reporting (claims 46–47). Nvidia is highly profitable; CoreWeave has positive adjusted EBITDA but material losses after depreciation and financing; lab profitability varies and remains incompletely disclosed (claims 49, 55).

Usage is growing, but the series do not establish Jevons. Google's monthly totals span all surfaces, while its API metric is narrower; OpenRouter is a separate reported series (claims 29, 31). Microsoft's trillion-token metric concerns customer counts, not demonstrated growth in each customer's consumption (claim 30). At constant units and mix, sevenfold volume and a 10–40-fold price decline imply revenue falling 30–82.5% (claim 57). Capability-price and aggregate raw-token series are not comparable enough to estimate this effect. Cloud growth is observed revenue growth, not an identification of price elasticity or a causal Jevons effect.

4. Training economics

Epoch's final-run cost fit grows 2.4x annually and would exceed $1bn by 2027 if the historical trend continued (claim 36). Its estimates put final released-model runs at 9.6%, 22.6% and 12.3% of R&D compute spending at OpenAI, MiniMax and Z.ai. Filings supply R&D denominators, while modeled training costs supply numerators; the ratios are C-grade estimates, not direct prospectus ratios (claim 37). Experiments and unreleased models also involve training, so these estimates do not make training broadly defined a minority of lab compute. DeepSeek-V3's disclosed final-run cost excludes prior research and ablations (claim 38).

Accounting depends on the contract and asset owner. Owners capitalize hardware and recognize depreciation; power and staff can enter R&D, as illustrated by Meta's expense disclosure (claim 40). Renters, lessees and customers making prepayments can have different recognition patterns; the financing arrangements above rule out a categorical “labs expense, hyperscalers capitalize” rule. OpenAI's research-compute treatment lacks a primary accounting-policy disclosure, and Epoch's two-year amortisation is an assumption (claim 37). API retirement dates do not establish the economic life of model development, research or data, so they cannot establish a one-year model asset life or a GAAP earnings overstatement.

5. Scaling laws, efficiency and demand for compute

Epoch's estimated eight-month efficiency halving and MIT's algorithmic contribution concern compute or prices for fixed capability (claims 28, 39). Reasoning and agentic workflows can add computation per task, but Huang's “100 times” compute assertion is not a measured distribution of tokens per answer. Aggregate tokens do not directly measure compute consumption across changing models and hardware. The evidence therefore does not establish that compute demand outgrew efficiency by a particular margin. Cheaper new capacity could erode older assets' pricing power, while greater usage could sustain them; utilization, realized prices and operating costs determine the outcome.

Vertiv’s up-to-30% first-year token gain models an earlier start at equal installed capacity, rather than higher steady-state throughput or observed customer returns (79). Tesla’s >$25bn→>$20bn threshold also shows issuer-specific investment restraint, while providing neither a midpoint nor evidence of a cut by the four eligible hyperscalers (78). IBM/TI/SAP disclosures and Arc availability are recorded without assigning weight; broad revenue and a product launch cannot calibrate the model’s owner margins or realized ROI (80–83). Vertiv, Tesla.

A worked model

We illustrate capital recovery for an assumed $750bn new cohort, with two-thirds ($500bn) short-lived equipment and one-third ($250bn) long-lived infrastructure; this is not an estimate of AI-only investment. The mix extrapolates Microsoft’s disclosure, and the 15-year long life is assumed, not a sector-wide accounting policy (claim 53). Revenue R at owner EBITDA margin m repays each tranche with discount rate r. Pre-tax, R = [500 / A(r,L) + 250 / A(r,15)] / m, where A(r,n) = (1−(1+r)^−n)/r. Required annual revenue in $bn:

At six-year equipment life, 50% owner EBITDA margin and 10% discount rate, the hypothetical cohort requires $295bn annually; at three years, $468bn. An exactly $1.36tn hypothetical new cohort requires $536–848bn under the same assumptions. That extension does not value the remaining obligations of previously deployed cohorts: it ignores past recovery and remaining lives. Goldman's $417bn can be matched numerically at approximately 39.5% EBITDA margin with five-year life or 35.4% with six years, but that does not confirm its capital base or return definition; ROIC is not automatically an annuity discount rate (claims 14, 53).

Over the selected ranges, equipment life changes required revenue by 1.6x, while a margin change from 50% to 30% changes it by 1.7x. Raising r from 8% to 12% adds 10–15%, depending on life. The displayed margin sensitivity is slightly greater than the life sensitivity; these ranges do not prove a universal ranking (claim 53). A renter's gross margin cannot substitute for an owner's EBITDA margin: rental payments already fund the owner's hardware. Recovering that capex again would mix layers.

The model attributes all assumed investment to incremental revenue and omits replacement assets, conventional cloud, logistics, internal AI benefits, commissioning delays, taxes, working capital, declining utilization/prices and residual values; land is not modeled separately. It requires the short-lived revenue component for L years and the long-lived component for 15 years, not one constant combined stream forever. A company-by-company revenue bridge and commissioning/cash-flow schedule are needed to quantify an actual payback gap or a doubling deadline. Three-year equipment life is mathematically possible at sufficient revenue; five-plus years is not a model requirement.

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 60–66; project scenario: claim 53).

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.

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 66). 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 monetization and investment: claim 51’s formerly general internal-benefits argument now has Meta management attribution and strong ad metrics, but the paired negative FCF prevents treating that attribution as demonstrated cash payback. Its raised capex floor/midpoint also supersedes July-only summaries; Tesla’s lower open-ended threshold is a different issuer/perimeter, not a four-company cut (7, 52, 76–78). Enterprise disclosures receive no assigned weight, and Vertiv’s modeled earlier start cannot establish realized economy-wide ROI (79–83). Meta, Tesla Q2, Q3, Vertiv.

• October profitability and accounting update. SpaceX’s primary Q3 SI operating profit supersedes its Q2 loss as the current quarterly reading and demonstrates operating profit at the mixed SI segment level; it does not identify standalone frontier-model profitability. Larger capex and the group cash deficit leave lifetime recovery unresolved (claims 67–68). Anthropic’s reported H1 net loss and Q3 adjusted profit use different periods and definitions; removing a noncash financing charge does not calculate cash burn (claims 44, 46, 69).

• Commitment and distribution perimeters. The newer ~$450bn Anthropic cloud-contract report differs from September’s $518bn broader obligations; no scope bridge supports calling that a cancellation. $750bn includes negotiations. OpenAI’s Google channel also permits existing commitments to fund purchases, so supplier and lab receipts cannot be summed as new final demand (claims 70–71).

• Mutable GPU source conflict. At the 21 October signal cutoff no fresh Silicon print was established; the note’s later 15:24 UTC re-fetch recovered Oct 20 $2.85. The Oct 25 retrieval now recovers dated Oct 21–24 $2.86/2.88/2.90/2.88 from the card/embedded chart, including Saturday, while the FAQ still says $2.53. These newer observations supersede an ongoing-unavailability inference, not the earlier access record; no complete-month publication/revision flags or invented Oct 25 settlement follow. OCPI H100’s latest weekly rise supersedes its earlier weekly decline, while benchmarks and listings remain distinct (22, 74). Silicon, Ornn.

• Useful life and earnings. Burry argues for a much shorter economic life than filers' server schedules (claims 17–21). The original note judged rentals supportive of longer lives, H100 viability dependent on scarcity, and its depreciation sensitivity supportive of Burry's earnings estimates. The cross-check agrees life matters but finds none of those conclusions established by listings or one benchmark. Realized utilization, costs, cohort timing and company earnings denominators are needed; the scenario does not resolve the dispute.

• Largest uncertainty and who bears losses. The original note ranked GPU life first. The cross-check finds the selected margin sensitivity slightly larger and emphasizes customer solvency, contract terms and funding independently. Its judgment is that highly committed lab and leveraged-provider equity is exposed early, with lenders, lessors and suppliers affected according to protections; cash stress can begin at customer or provider depending on guarantees and prepayments. This is a risk assessment, not an established loss sequence.

• Pricing and Jevons. The original note read rising premium list prices as a universal frontier reversal and token growth as proof of Jevons. The cross-check points to Sol and Opus cuts and rejects the inference from unmatched token/capability-price series. The corrected fixed-unit arithmetic runs against the original revenue claim. Pricing power versus temporary scarcity remains unresolved without comparable workloads and paid-volume data (claims 25–30, 48, 57).

• Revenue gap and deadline. The original note proposed $150–250bn end revenue, a narrowing gap and doubling by 2028. The cross-check finds no reproducible revenue bridge, matching asset population or timing schedule. Those estimates are withdrawn from the active analysis; run-rates and booked quarters remain distinct and overlap supplier revenue. Published required-revenue scenarios also differ in layer and return definition (claims 11–14, 30, 33–34, 46).

• Profitability by layer. The original note described neoclouds as roughly break-even after depreciation and interest and labs broadly as cash-burning. The cross-check finds CoreWeave's material net losses directly contradict that characterization, while Anthropic's reported Q3 adjusted profitability is counterevidence to a blanket lab claim. Positive gross margin does not prove inference costs exceed sales; adjusted profit does not prove durable free cash flow. Anthropic's reported historical losses and newer adjusted result use different periods and metrics, with no primary accounts inspected (claims 32, 44, 46–47, 49).

• Lab training accounting and model life. The original note treated MiniMax/Z.ai final-run shares as direct filing disclosures and inferred roughly one-year model value and earnings flattery from endpoint retirements. The cross-check identifies modeled numerators and rejects the economic-life inference. We retain the ratios as C-grade Epoch estimates and leave OpenAI's accounting unresolved pending a primary policy disclosure (claim 37).

• Training-cost interval within the source. The cross-check specifies 95%, matching Epoch's abstract for the 2.4x trend. The fetched publication also has a revision note saying intervals were switched to 90%. Claim 36 uses the abstract's 95% label; the source's inconsistent presentation remains unresolved rather than implying the whole publication uses one interval.

• Primary access and unresolved subclaims. Goldman and Morgan Stanley remain bank-model summaries; McKinsey's figure was available only through a snippet (claims 14–16). Google's historical 3 Pro price was not independently recovered in the cross-check (claim 27). Access limits do not establish falsity. Project Jupiter's reported force-majeure notice is unverified; Oracle's unchanged-scope statement does not resolve contractual payment timing (claim 56).

Signals to watch

Sources

1. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm — SEC EDGAR, Amazon.com Form 10-K for FY2025, filed 2026-02-06, accessed 2026-10-17.

2. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm — SEC EDGAR, Amazon.com Form 10-Q for Q2 2026, filed 2026-07-31, accessed 2026-10-17.

3. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm — SEC EDGAR, Amazon Q2 2026 results, Exhibit 99.1, 2026-07-30, accessed 2026-10-17.

4. https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm — SEC EDGAR, Microsoft Form 10-K for FY2026, filed 2026-07-29, accessed 2026-10-17.

5. https://www.sec.gov/Archives/edgar/data/789019/000119312526323632/msft-ex99_1.htm — SEC EDGAR, Microsoft FY26 Q4 earnings release, 2026-07-29, accessed 2026-10-17.

6. https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/TranscriptQandAFY26Q3 — Microsoft, FY26 Q3 earnings call transcript, 2026-04-29, accessed 2026-10-17.

7. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm — SEC EDGAR, Alphabet Form 10-K for FY2025, filed 2026-02-05, accessed 2026-10-17.

8. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm — SEC EDGAR, Alphabet Q2 2026 results, Exhibit 99.1, 2026-07-22, accessed 2026-10-17.

9. https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm — SEC EDGAR, Meta Platforms Form 10-K for FY2025, filed 2026-01-29, accessed 2026-10-17.

10. https://www.sec.gov/Archives/edgar/data/1326801/000162828026050596/meta-06302026xexhibit991.htm — SEC EDGAR, Meta Q2 2026 results, Exhibit 99.1, 2026-07-29, accessed 2026-10-17.

11. https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm — SEC EDGAR, Meta Form 10-Q for Q2 2026, filed 2026-07-30, accessed 2026-10-17.

12. https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm — SEC EDGAR, Oracle Form 10-K for FY2026, filed 2026-06-22, accessed 2026-10-17.

13. https://www.sec.gov/Archives/edgar/data/1341439/000119312526387905/orcl-ex99_1.htm — SEC EDGAR, Oracle Q1 FY2027 results, Exhibit 99.1, 2026-09-10, accessed 2026-10-17.

14. https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm — SEC EDGAR, Oracle Form 10-Q for Q1 FY2027, filed 2026-09-11, accessed 2026-10-17.

15. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm — SEC EDGAR, Nvidia Form 10-K for FY2026, filed 2026-02-25, accessed 2026-10-17.

16. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27pr.htm — SEC EDGAR, Nvidia Q2 FY2027 results, Exhibit 99.1, 2026-08-26, accessed 2026-10-17.

17. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm — SEC EDGAR, CoreWeave Form 10-K for FY2025, filed 2026-03-02, accessed 2026-10-17.

18. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm — SEC EDGAR, CoreWeave Q2 2026 results, Exhibit 99.1, 2026-08-11, accessed 2026-10-17.

19. https://www.cnbc.com/2026/07/28/hyperscalers-face-higher-capex-scrutiny-after-alphabet-report-panned.html — CNBC, "Hyperscalers face higher capex scrutiny after Alphabet report panned", 2026-07-28, accessed 2026-10-17.

20. https://www.yieldtheory.app/research/hyperscaler-ai-capex-tracker-2026 — Yield Theory, "Hyperscaler Capex 2026: $720B–$745B Guidance by Company", 2026, accessed 2026-10-17.

21. https://www.sequoiacap.com/article/ais-600b-question/ — Sequoia Capital, David Cahn, "AI's $600B Question", 2024-06-20, accessed 2026-10-17.

22. https://www.bain.com/insights/how-can-we-meet-ais-insatiable-demand-for-compute-power-technology-report-2025/ — Bain & Company, Technology Report 2025, "How can we meet AI's insatiable demand for compute power?", 2025-09-23, accessed 2026-10-17.

23. https://am.jpmorgan.com/us/en/asset-management/institutional/insights/market-insights/market-updates/on-the-minds-of-investors/how-is-ai-being-monetized — J.P. Morgan Asset Management, "How is AI being monetized?", 2026-01-09, accessed 2026-10-17.

24. https://advisoranalyst.com/2026/10/15/ai-capex-required-revenues-and-what-the-market-is-actually-pricing.html/ — AdvisorAnalyst, summary of Goldman Sachs "AI Capex, Required Revenues, and What's Priced in the AI Trade" (2026-09-24), 2026-10-15, accessed 2026-10-17.

25. https://cryptobriefing.com/goldman-sachs-hyperscaler-capex-ai-revenue/ — Crypto Briefing, "Goldman Sachs calculates the revenue hyperscalers need to justify $1.7T in AI capex", 2026-09-25, accessed 2026-10-17.

26. https://cryptobriefing.com/morgan-stanley-data-center-capex-2028/ — Crypto Briefing, "Morgan Stanley estimates $2.9T in global data-center capex through 2028", 2026-06-19, accessed 2026-10-17.

27. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers — McKinsey, "The cost of compute: a $7 trillion race to scale data centers", 2025-04; page fetch timed out, figures from search snippet, accessed 2026-10-17.

28. https://www.hudson-labs.com/research/amazon-server-depreciation-amzn — Hudson Labs, "Amazon Server and Networking Equipment Useful Life Changes", accessed 2026-10-17.

29. https://www.amcompute.com/gpu-depreciation-schedules — AM Compute, "GPU Depreciation Schedules", 2026-10-10, accessed 2026-10-17.

30. https://deepquarry.substack.com/p/depreciation-of-gpus-between-useful — Deep Quarry (Olga Usvyatsky), "Depreciation of GPUs: between useful lives and useful myths", 2025-12-07, accessed 2026-10-17.

31. https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html — CNBC, "'Big Short' investor Michael Burry accuses AI hyperscalers of artificially boosting earnings", 2025-11-11, accessed 2026-10-17.

32. https://www.home.saxo/content/articles/equities/big-short-12112025 — Saxo, "The Big Short: Is Michael Burry right about the AI trade?", 2025-11-12, accessed 2026-10-17.

33. https://akash.network/the-bid/h100-rental-price-2026-cost-per-hour/ — Akash Network, "H100 Rental Price in September 2026: Cost Per Hour by GPU Provider", 2026-09-12, accessed 2026-10-17.

34. https://getdeploying.com/reference/cloud-gpu/nvidia-a100 — GetDeploying, "NVIDIA A100 cloud GPU pricing", updated 2026-10-17, accessed 2026-10-17.

35. https://inferencex.semianalysis.com/compare/minimax-m27-gb200-vs-h100 — SemiAnalysis InferenceX, "MiniMax M2.5/M2.7 performance per dollar, GB200 NVL72 vs H100", July 2026 pricing, accessed 2026-10-17.

36. https://developers.openai.com/api/docs/pricing — OpenAI, API pricing, accessed 2026-10-17; model pages for gpt-4.1, gpt-5, gpt-5.2, gpt-5.4, gpt-5.5, gpt-6-sol, gpt-6.1-sol, gpt-6-astra and https://developers.openai.com/api/docs/changelog, accessed 2026-10-17.

37. https://platform.claude.com/docs/en/about-claude/pricing — Anthropic, Claude pricing; https://platform.claude.com/docs/en/about-claude/model-deprecations; model pages for Opus 4.5, Opus 5.5, Fable 5.1; accessed 2026-10-17.

38. https://ai.google.dev/gemini-api/docs/pricing — Google, Gemini API pricing; https://ai.google.dev/gemini-api/docs/changelog; https://developers.googleblog.com/en/updated-gemini-models-reduced-15-pro-pricing-increased-rate-limits-and-more/ (2024-09-24); accessed 2026-10-17.

39. https://epoch.ai/data-insights/llm-inference-price-trends — Epoch AI, "LLM inference prices have fallen rapidly but unequally across tasks", 2025-03-12, accessed 2026-10-17.

40. https://arxiv.org/html/2511.23455v1 — Gundlach, Lynch, Mertens, Thompson (MIT), "The Price of Progress", 2025-11-28, accessed 2026-10-17.

41. https://blog.google/innovation-and-ai/sundar-pichai-io-2026/ — Google, Sundar Pichai, "I/O 2026: Welcome to the agentic Gemini era", 2026-05-19, accessed 2026-10-17.

42. https://www.trendingtopics.eu/chinese-ai-models-overtake-us-rivals-in-global-token-consumption/ — Trending Topics, "Chinese AI models overtake US rivals in global token consumption", 2026-03, accessed 2026-10-17.

43. https://www.kucoin.com/blog/ua-OpenRouter-raises-113m-series-b-as-valuation-hits-billions — KuCoin, "OpenRouter raises $113m Series B", 2026-05-27, accessed 2026-10-17.

44. https://the-decoder.com/openai-adds-111-billion-to-its-cash-burn-forecast-as-ai-costs-spiral-beyond-projections/ — The Decoder, Matthias Bastian, on OpenAI margins citing The Information, 2026-02-21, accessed 2026-10-17.

45. https://cryptobriefing.com/openai-revenue-6-7-billion-q2-2026/ — Crypto Briefing (citing WSJ), "OpenAI revenue climbs to $6.7B in Q2", 2026-08-18, accessed 2026-10-17.

46. https://techeconomy.ng/anthropics-annual-revenue-run-rate-tops-65-billion — TechEconomy (citing Reuters), "Anthropic's annual revenue run rate tops $65 billion", 2026-08-18, accessed 2026-10-17.

47. https://epoch.ai/publications/how-much-does-it-cost-to-train-frontier-ai-models — Epoch AI, "How much does it cost to train frontier AI models?", 2024-06-03, updated 2025-01-13, accessed 2026-10-17.

48. https://epoch.ai/data-insights/openai-compute-spend — Epoch AI, "Most of OpenAI's 2024 compute went to experiments", 2025-10-10, accessed 2026-10-17.

49. https://epoch.ai/gradient-updates/r-and-d-vs-training-compute — Epoch AI, "Final training runs account for a minority of R&D compute spending", 2026-03-23, accessed 2026-10-17.

50. https://arxiv.org/html/2412.19437v1 — DeepSeek-AI, "DeepSeek-V3 Technical Report", 2024-12-26, accessed 2026-10-17.

51. https://epoch.ai/blog/algorithmic-progress-in-language-models — Epoch AI, "Algorithmic progress in language models", 2024-03-12, accessed 2026-10-17.

52. https://www.nbcconnecticut.com/news/business/money-report/nvidias-jensen-huang-on-why-deepseeks-new-model-will-need-100-times-more-computing/3524344/ — CNBC via NBC Connecticut, "Nvidia's Jensen Huang on why DeepSeek's new model will need 100 times more computing", 2025-03-19, accessed 2026-10-17.

53. /data/ai-unit-economics-capex-by-year.csv, /data/ai-unit-economics-api-prices.csv, /data/ai-unit-economics-useful-life.csv — series extracted for this note.

54. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 — Microsoft, FY26 Q4 earnings conference call, 2026-07-29, accessed 2026-10-17.

55. https://apnews.com/article/amazon-second-quarter-earnings-cloud-b4ce02b4666a35b8975823c5c22072ee — Associated Press, Amazon second-quarter results and updated capex, 2026-07-30, accessed 2026-10-17.

56. https://www.microsoft.com/investor/reports/ar25/index.html — Microsoft, FY2025 annual report, fiscal year ended 2025-06-30, accessed 2026-10-17.

57. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000432/crwv-20260917.htm — SEC EDGAR, CoreWeave Form 8-K, September convertible completion 2026-09-22, accessed 2026-10-17.

58. https://www.marketscreener.com/news/anthropic-s-ipo-prospectus-shows-sweeping-ai-vision-surging-costs-ce785addd98afe2c — Reuters via MarketScreener, Anthropic prospectus, 2026-09-28, updated 2026-09-29; nonpublic document not inspected, accessed 2026-10-17.

59. https://www.marketscreener.com/news/broadcom-to-lend-anthropic-up-to-42-billion-to-lease-its-chips-filing-says-ce785ad3de89fe20 — Reuters via MarketScreener, Broadcom supplier-financing arrangement, 2026-10-01; nonpublic document not inspected, accessed 2026-10-17.

60. https://ca.marketscreener.com/news/anthropic-s-q3-revenue-surges-45-to-nearly-17-billion-ce785dd2d08cf221 — Reuters via MarketScreener, Anthropic investor-day revenue and adjusted profitability, 2026-10-14, accessed 2026-10-17.

61. https://www.breakingviews.com/columns/breaking-view/anthropics-2-trln-goal-is-ais-biggest-moonshot-2026-09-29/ — Reuters Breakingviews, Anthropic valuation and historical gross margin, 2026-09-29, accessed 2026-10-17.

62. https://www.anthropic.com/claude-opus-5-5 — Anthropic, Introducing Claude Opus 5.5, 2026-09-22, accessed 2026-10-17.

63. https://www.anthropic.com/news/google-broadcom-partnership-compute — Anthropic, expanded Google and Broadcom compute partnership, 2026-04-06, accessed 2026-10-17.

64. https://investors.broadcom.com/news-releases/news-release-details/broadcom-apollo-and-blackstone-establish-landmark-strategic — Broadcom, Apollo and Blackstone AI infrastructure financing platform, 2026-06-09, accessed 2026-10-17.

65. https://www.stockopedia.com/share-prices/alphabet-NSQ%3AGOOGL/news/google-increases-capex-forecast-again-after-cloud-driven-quarterly-beat-updated-019f8c0e-edfb-74ba-9a45-207222cd42f6/ — Reuters via Stockopedia, Google increases capex forecast, 2026-07-22, accessed 2026-10-17.

66. https://www.axios.com/2026/09/25/oracle-debt-data-centers — Axios, Oracle credit-risk report, 2026-09-25; underlying notice not inspected, accessed 2026-10-17.

67. https://www.oracle.com/news/announcement/project-jupiter-is-expected-to-generate-13-billion-in-economic-impact-2026-10-13/ — Oracle, Project Jupiter update, 2026-10-13, accessed 2026-10-17.

68. https://ca.investing.com/news/stock-market-news/openais-q2-revenue-growth-lagged-anthropic-as-losses-deepened-wsj-reports-4806938 — Investing.com, WSJ report summary on OpenAI Q2 revenue, 2026-08-18, accessed 2026-10-17.

69. https://www.channelnewsasia.com/business/openai-burned-37-billion-in-first-quarter-2026-information-reports-6188641 — Reuters via CNA, OpenAI Q1 cash burn from The Information, 2026-06-16, accessed 2026-10-17.

70. https://www.axios.com/2026/08/17/anthropic-revenue-run-rate-ipo-openai — Axios, Anthropic reported July run-rate, 2026-08-17, accessed 2026-10-17.

71. https://www.axios.com/2026/09/29/scoop-openais-annual-recurring-revenue-nears-70b — Axios, OpenAI run-rate report, 2026-09-29, accessed 2026-10-17.

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

73. 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.

74. 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.

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

76. 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.

77. 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.

78. 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.

79. 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.

Sources re-fetched for the 21 October update

Primary SEC exhibits and publisher pages were re-fetched on 2026-10-21; SEC requests used a generic research User-Agent. Anthropic financial information remains secondary reporting of a nonpublic draft, with only a partial WSJ original accessible. CNBC, FT syndication and Reuters syndications establish attributed market quotes/commentary, not primary trading tapes. Silicon Data’s recovered dated display differs from the earlier signal-cutoff access result.

Sources re-fetched for the 25 October update

Primary releases, filings, benchmark pages 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/Bloomberg syndications remain secondary market/private-financing reporting; no primary trading tape or private-company accounts were obtained. Retrospectively retrieved GPU histories are not independent same-day captures or complete-month coverage. Older items from 21 October are retained as baseline evidence, not new discoveries. No forecast or scenario assumptions were revised.