AI Commercialization and Revenue Monthly Brief · Sep 24–Oct 23, 2026

1. Key takeaways from the month

Between Sep 24 and Oct 23, 2026, AI companies published more revenue numbers than in any earlier month, and the numbers did not share a basis. Five conclusions explain the month; each is supported in the chapter named in brackets, with figures judged against Oct 24.

1. The top revenue figures are large, fast-growing and measured differently. Anthropic's run-rate reached $65B at the end of July (reported, counting cloud-partner sales gross), while OpenAI told investors about $50B annualized at the end of September on its own net basis. The $68–70B figure that circulated in late September was an investor gross-up, not an OpenAI number. Anthropic's quarterly revenue went from $4.7B to $11.6B to $16.8B (+147%, then +45%). [Chapters 2, 7]

2. Usage is enormous; paying share is small. ChatGPT showed 1.2B weekly users on Sep 29, and OpenAI last gave 50M consumer subscribers, which is 5.6% of the 900M weekly users it reported on Feb 27. Microsoft 365 Copilot passed 30M paid seats against 450M+ commercial seats, at most 6.7% (mixed dates). [Chapter 4]

3. Prices moved in opposite directions. Kingy AI's median small-model price fell from $0.53 to $0.26 per million tokens in October, and Anthropic's Haiku 5.5 launched at $0.10 / $0.50 against Haiku 4.5's $1 / $5. At the frontier, Gemini 4 Argon's introductory $2 / $10 doubles to $4 / $20 on Dec 1. [Chapter 5]

4. Applications and infrastructure report revenue faster than they can fund growth. Podium announced more than $150M of AI-specific ARR and Chai reported $100M ARR in July (a $120M unaudited run-rate on Oct 1). CoreWeave's Q2 revenue was $2.58B (+112%) with a $104B backlog and about $29.5B of net debt. [Chapter 3]

5. Profitability is claimed, contested and set against huge commitments. Anthropic told investors it was profitable on an adjusted basis, without a published definition, while its draft S-1 shows a first-half net loss of $80.7B (about $72B non-cash) and compute commitments of at least $518B. The Oct 8 report on OpenAI's revenue moved suppliers: CoreWeave −7.77%, Oracle −5.48%, Nvidia −2.94%. [Chapters 6, 7]

1.1 Metrics at a glance

2. Revenue and ARR ranking: model vendors

Chapter 1 summarised the month; this chapter sets out the underlying revenue figures, what each measures and who said it. Unless marked otherwise, every figure is a press report of investor materials or a company release, not an audited public-company number.

2.1 Ranking table: one family of numbers, four bases

Only "annualized" figures exist for all four companies, so the table orders that family by value. It is not a like-for-like ranking: the dates differ, and Anthropic counts cloud-partner sales gross while OpenAI counts only its own share.

The material holds no AI-specific revenue figure from Google or Meta in the window, so neither is ranked (usage is in Chapter 4, capex in Chapters 6-7).

SpaceX reported Q3 on Oct 20, 2026: revenue $14.6B (+161%), of which SI was $7.56B, up 830% from $813M a year earlier. That is 7.56 ÷ 14.6 ≈ 52% of revenue and 2.95× Q2's $2.56B. SI is not a model-revenue line. It combines cloud services (customers include Anthropic and Google), Cursor (acquired Aug 14; $608M added through September), Grok and $538M of X advertising. Do not add SI to Anthropic, because Anthropic's compute spend is part of SI's revenue. Musk said annual recurring revenue would reach $500B to $1T by end-2027, against LSEG analysts' roughly $110B of 2027 total revenue; treat his figure as an aspiration.

Microsoft's $37B (about 11% of revenue, per Bloomberg research) is the latest AI run rate in the material. Its fiscal Q4 2026 call gave only a seat count, over 30 million paid Microsoft 365 Copilot seats, not revenue.

2.2 OpenAI's two numbers, and why they differ

OpenAI's "$20B lower" report and its "$70B by year-end" expectation are different numbers on different bases. The first is a September estimate built on Anthropic's gross method; the second is OpenAI's own net-basis expectation for December.

The $50B figure. On Oct 8 the Financial Times reported that OpenAI told investors its annualized revenue was "approaching $50 billion". Bloomberg (Oct 9) said roughly $50B at end-September, shared with investors in talks to raise $30B or more at a $1.4T pre-money valuation (reported, unconfirmed). The presentation also showed Q3 run-rate growth of 77% overall and 107% in enterprise, per CNBC.

The $68-70B figure. In late September Axios and others reported OpenAI near $70B; Yahoo Finance dates a $68B version to Sept 29. It was never an OpenAI number. The FT said investors, trying to compare OpenAI with Anthropic, started from an estimated $40B in August and added the 70% growth OpenAI had cited: $40B × 1.7 = $68B (our reading). CNBC confirmed the figure counted partners' gross revenue, not OpenAI's net share.

Why the bases differ. Both companies are GAAP-compliant. If a customer pays $100 through a cloud provider, Anthropic books $100 as revenue and the provider's cut as an expense, while OpenAI books only its share of certain partner sales. The choice turns on who controls the customer relationship and delivers the product (Axios, quoting accounting professor Francine McKenna). The gap is 70 − 50 = $20B (68 − 50 = $18B), a ratio of 1.36-1.40×. FutureSearch's roughly 1.4× gross-up is consistent, but it is an aggregator estimate.

The $70B year-end expectation. Bloomberg reported on Oct 9 that OpenAI expects $70B or more of annualized revenue by end-2026, driven mainly by enterprise (reported expectation, not a result). It continues the $50B series, so we infer net basis. It needs 70 ÷ 50 = +40%, or $20B, in about one quarter. It is not September's $70B gross-up.

Run rate versus booked revenue. Reported recognized revenue was $5.7B in Q1 and $6.7B in Q2 2026; Q2 × 4 = $26.8B, far below the end-September run rate, because a run rate captures the fastest-growing latest period. The same gap showed in 2025: $13.07B booked against a run rate above $20B at year-end.

2.3 Anthropic's quarterly path

Anthropic told investors at a San Francisco investor day on Oct 14, 2026 that Q3 revenue was $16.8B and that it was profitable on an adjusted basis (reported by attendees; Anthropic declined to comment; not a filing).

Nine-month revenue is 4.7 + 11.6 + 16.8 = $33.1B, about 7× 2025's roughly $4.6B. The WSJ's review of the confidential draft S-1 gives first-half revenue of $16.3B, which equals 4.7 + 11.6, against $1.1B a year earlier. The loss side belongs to Chapter 6.

Conflicts. Earlier reports put Q1 at $4.8B (WSJ) and Q2 at $11.5B (Bloomberg). We follow the investor-day figures, which are the latest and sum to the S-1 half-year number. One article's "about 14%" Q1-to-Q2 growth is a slip; the arithmetic gives 147%. Others round Q3 growth to 50%.

Basis. Business Insider says it is unclear whether $16.8B is gross or net, and that Anthropic has historically counted gross cloud-partner sales. In the draft's 2025 data, $2.16B of sales ran through Amazon and Google clouds (47% of revenue) with $351M of distribution fees: about 16% of that channel, 7.6% of $4.6B total (our arithmetic). That is far below OpenAI's 29% gap (20 ÷ 70), so the two adjustments are not interchangeable.

Run rate and expectations. Anthropic's run rate was about $9B at end-2025, over $30B on Apr 6, over $47B in early May and $65B at end-July, per a compilation citing Anthropic and CNBC. Q3 × 4 = $67.2B, only 67.2 ÷ 65 − 1 ≈ 3% above July (a quarter average against a point-in-time figure, so compare cautiously). An attendee said Anthropic expects a $110B run rate by year-end (reported); Jefferies said the market had expected $80B or more now and $105B by year-end (analyst estimate).

3. Applications and infrastructure: ARR milestones

Chapter 2 ranked the model vendors' revenue; this chapter moves one layer down, to the application companies that buy those models and the infrastructure companies that sell the compute underneath them. Application ARR milestones kept arriving in the window, but "ARR" meant a different thing in almost every announcement, so the table below states each company's own definition.

3.1 ARR milestones and what each number measures

Podium is the cleanest milestone, and the only one framed against a prior business. It reports AI-specific ARR above $150 million about two and a half years after expanding into AI agents, and says its original software business took eight years to reach the same level. It also reports more than 20,000 deployed AI agents and says its customers generated more than $13 billion in revenue with its tools (company claim). CEO Eric Rea: "We effectively started a second startup inside an already successful company". The material gives no announcement date and no ARR definition beyond "AI-specific", so treat the timing as unverified.

Chai gives a trajectory, not a point. According to Dealroom, Chai said it reached $100M ARR in July 2026 with about 15 staff, after $80M at the end of Q1 and about $40M in 2025; by Oct 1, 2026 it put ARR at $120M as an unaudited run-rate. Two calculations follow. The Q1-to-October gain is (120 − 80) ÷ 80 = +50% in about six months. Per head, $100M ÷ ~15 ≈ $6.7M of ARR per employee in July. Chai sells subscriptions and in-app purchases and says it is profitable excluding sales and marketing. In-app purchases are not contracted recurring revenue, so "run-rate" is the more honest label here.

Axya, on Sep 24, 2026, closed a C$17M Series A and described its ARR only as "multi-million-dollar", while an investor said it was doubling annually with net revenue retention near 140%. This is a growth claim without a number.

3.2 Coding and agent applications

Cursor's revenue now sits inside another company's income statement (Chapter 2 covers SpaceX's Super Intelligence segment). SpaceX completed its $60 billion all-stock acquisition on Aug 14, 2026. In Q3 results, Cursor added $608 million of revenue through the end of September. New Market Pitch says Cursor's annualized revenue was about $4 billion by May 2026, so the price was about 60 ÷ 4 = 15× run-rate. Other sources date the $4B to June, and the earlier path was $1B (Nov 2025), $2B (Feb 2026), $3B (Apr 2026), all secondary.

Annualizing the post-close figure on mixed bases: Aug 14 to Sep 30 is 47 days, so 608 × 365 ÷ 47 ≈ $4.7B (our arithmetic, not a company number). That is consistent with continued growth after the pre-deal ~$4B, but it blends recognized revenue with an estimated run-rate. New Market Pitch also reports that OpenAI announced plans to end its direct model-supply agreement with Cursor, while Anthropic continued to supply Claude. Cursor's flat-rate "Auto" mode ended Aug 24, 2026 and now bills at the routed model's list price, a pass-through that exposes margin.

Revenue and usage are moving apart. In JetBrains' May to July 2026 survey, 39% of professional developers used Claude Code at work (18% in January), against GitHub Copilot 21%, Codex 16% and Cursor 12% (down from 18%). Claude Code passed $2.5 billion in run-rate revenue in February 2026, according to VentureBeat via NeuralTrust, earlier context that Chapter 2 treats within Anthropic.

Application ARR is arriving at record speed, but Podium's contracted AI ARR, Chai's unaudited run-rate and Cursor's recognized post-deal revenue are three different bases and cannot be added or ranked together.

3.3 Infrastructure and neoclouds: demand strong, financing the constraint

Infrastructure companies report recognized revenue and backlog, not ARR. I/O Fund's comparison, drawn from company Q2 2026 releases and filings, is the only same-source set here:

Checks: CoreWeave 29.5 ÷ (2.58 × 4 = 10.3) ≈ 2.9×; Nebius 0.5 ÷ (0.582 × 4 = 2.33) ≈ 0.2×. A Tickeron piece cites a backlog of about $55.6B. It is undated and conflicts with the $104B in company-sourced Q2 data, so we follow $104B and treat the $55.6B as an earlier quarter.

Demand evidence is pricing and statements. Nebius raised H100 on-demand prices 17%, from $3.85 to $4.50 per GPU-hour, effective Oct 1, 2026, with H200, B200 and B300 up 20%, 19% and 21% (company action, via I/O Fund). On Oct 22, Jensen Huang told CNBC at GTC Berlin that demand for neocloud capacity is "off the charts" but limited working capital is holding expansion back; he was speaking about the industry, not about CoreWeave (reported). For Nvidia itself, TradingKey reports fiscal Q2 revenue of $96.2B (+106% YoY), Data Center $89B, and a Q3 guide of about $108B ± 2%. The guide implies 108 ÷ 96.2 − 1 ≈ +12% quarter on quarter. Dates for those results are not stated, so they are context.

The constraint is the balance sheet. CoreWeave's Q2 net interest expense was $640M against a $626M net loss and $1.51B adjusted EBITDA, and its stock fell 9.67% on Oct 22. According to The Motley Fool (via Gotrade), CoreWeave's 2032 bond yields touched 15% on Oct 18; Nebius funds more through customer prepayments, which covered 50% to 60% of capex on its largest contracts. Market-reaction detail is in Chapter 7.

4. Users and paying-customer data

Chapter 3 followed ARR milestones at application and infrastructure companies; this chapter looks at the people and seats behind such revenue lines, and at how few of them pay.

4.1 Users: four bases that cannot be added

ChatGPT. OpenAI showed 1.2 billion weekly users at DevDay on Sep 29, 2026. Its written recap says "our collective 1.2B weekly users", so the scope is not fully settled; Whizi counts it as ChatGPT users, as the slide does. From 900 million on Feb 27, that is 300 million more in 214 days, or 1.40 million a day (company figures, Whizi's calculation).

Gemini. Google's 1 billion is monthly and covers the app only. AI Overviews (2.5B) and AI Mode (1B+) are Search features, counted separately. A weekly count is the stricter test, so 1.2B weekly and 1B monthly do not rank each other. Sensor Tower's May 2026 panel model gives ChatGPT 1.1B, Gemini 662M and Claude 245M, against the 900M+ Google announced that month, so use it for ordering only.

Google's depth metric is tokens. Model APIs ran at about 22 billion tokens a minute on the Jul 22, 2026 call, up from 16 billion a quarter earlier. At 22B × 60 × 24 × 30 days that is about 950 trillion a month. In May the API rate was about 19B a minute (about 0.82 quadrillion a month), roughly 26% of the 3.2 quadrillion tokens Google processed across all surfaces.

Copilot, Claude, Muse. Microsoft sells seats, not users. Seats were 15M (Oct–Dec 2025), over 20M (Jan–Mar 2026) and over 30M (Apr–Jun 2026), so net adds rose from 5M to 10M a quarter. Its last user count was 150M monthly Copilot users for Jul–Sep 2025, and since then it has given only growth multiples. Anthropic publishes no user count. In March 2026 it said only that paid subscriptions had "more than doubled" since Jan 1. In the material, Meta gave no Muse user count. The only figures are The Information's Oct 5 report, relayed by MLQ: Claude Code users at Meta fell from about 60,000 to about 30,000, while MetaCode has 30,000+ and Muse Code 6,000+ internal users. Those are employees, not market adoption (reported).

4.2 Free versus paying

About 4–7% of ChatGPT users and Microsoft 365 Copilot seats pay. Only GitHub Copilot (23.5%, mixed dates) and Claude (13%, an aggregator estimate) exceed 10%.

Four points qualify the table.

• ChatGPT. The 50M is a "more than" count and is seven months older than the 1.2B, so 4.2% is a floor if subscribers have not fallen. Fatjoe describes "around 50 million paying subscribers" as including Business and Enterprise. We follow OpenAI's split as relayed by Whizi: 50M consumer subscribers, and 9M paying business users counted separately. That business count compares with 2M in Feb 2025, so 9 ÷ 2 = 4.5×, but OpenAI's wording shifts between users and seats.
• Microsoft. A seat is a purchased licence, not a user. Seats are rising partly through bundles such as E7, so they measure contract value more than adoption. The base is also growing about 6% a year.
• Two 3.3%s. The Register reported in Feb 2026 that 3.3% of Microsoft 365 users who touch Copilot Chat pay, a different denominator, and the piece is undated. The 15 ÷ 450 result is the same by coincidence.
• Missing disclosures. Google names no paying Gemini subscribers. Microsoft stopped giving GitHub Copilot paid counts after January and now gives only 50M users. Two aggregator-relayed estimates cover US consumers: 4.5% pay for ChatGPT, Gemini or Claude (a16z analysis, Oct 5), and 2.2% of households paid for any AI subscription as of April 2026 (PNC via a16z).

4.3 Enterprise adoption: breadth is high, depth is not

Ramp's AI Index, released Oct 14, 2026, shows 55.9% of US businesses on its platform paid for AI in September, up 0.62 points on August. Its lead economist said "the growth in new AI customers has slowed down significantly", and token spend was flat in August and September. The base is 70,000+ businesses using Ramp cards and bill pay. Fatjoe gives 56.1% for August, but 55.9 − 0.62 = 55.3%. We follow the Ramp figure. For August, 43.8% of Ramp businesses paid Anthropic and 39.8% paid OpenAI. The shares overlap and cannot be added. Ramp skews to startups.

Representative surveys give a lower base. US Census: 19.8% of businesses used AI in any function (May 2026), with Statistics Canada at 19.2% (Q2 2026) and Eurostat at 20% (2025). Ramp's rate is 55.9 ÷ 19.8 ≈ 2.8× the Census rate partly because it counts payers among card-using firms. Gallup: 52% of US workers used AI at work in Q2 2026, 15% daily. McKinsey's August 2026 survey (1,719 respondents, 36% at $1B+ revenue) found nearly 9 in 10 using AI in a function, 44% scaling enterprise-wide and 40% of large organisations scaling agents, up from 27%.

Depth is thinner. Recon Analytics (Feb 4, 2026) found that only 35.8% of US workers with a paid AI subscription and Copilot at work use it as their main AI tool. Gartner (Jun 9, 2026, 186 organisations, quoted second-hand) found 51% naming oversharing and data loss as the top barrier. Bain's Sep 29 report says "as much as 90%" of enterprises remain on narrow tool deployment, but this is Bain's assessment, not a survey. Meta's halving of internal Claude Code users shows that seats can fall (reported).

The trend is a shift from adding payers to deepening use. Kharazian puts it this way: "the growth in the market will have to come from more intense adoption rather than more customers". That raises the value of usage-linked metrics such as tokens and agent use over seat counts.

5. Pricing changes

Chapter 4 counted who uses and who pays for AI products; this chapter looks at what they are charged. Per-token prices fell at the small-model end and were set to rise at the frontier, while list prices of existing models barely moved.

5.1 API prices: introductory versus standard

USD per 1M tokens, one source per row. "In effect" means the date is on or before 2026-10-24. Where the vendor's price page is not in the material, the row says so.

Google calls Argon's $2/$10 an "introductory price", with cached input 95% cheaper. It was announced on Sep 30 to trusted cyber defenders only and opened publicly on Oct 13. OpenAI's Luna price is not stated; Kingy AI shows only a blended $0.20.

Take an illustrative workload of 100M input and 20M output tokens a month (our example). Argon costs 100 × $2 + 20 × $10 = $400 now and 100 × $4 + 20 × $20 = $800 from Dec 1. Haiku 5.5 (prompts up to 100k) costs 100 × $0.10 + 20 × $0.50 = $20, against $200 on Haiku 4.5: a 90% cut. Anthropic itself says "around 75% less" on average, with its basis not in the material.

5.2 What moved and why

The month's price signal was replacement, not repricing: new small models launched at a tenth of their predecessors' price, and the new frontier model launched at a price with an expiry date.

Small models fell. According to the Kingy AI Price Index, the small-model median fell from $0.53 to $0.26 (down 50.0%). Method: a median of each lab's list price, blended 3 input : 1 output, captured Oct 19 from 26 official listings and compared with a reconstructed Sep 20 baseline. Only five models sit in that tier. Kingy says none of the 17 models and plans present in both editions changed list price; every move came from a lab swapping in a new model. Haiku 4.5 at $2.00 blended gave way to Haiku 5.5 at $0.20, which matches our check: (3 × $0.10 + $0.50) ÷ 4 = $0.20. Kingy is a founder-led media business with commercial partnerships, so treat the index as a secondary reading.

The frontier reprices upward. Kingy's frontier median fell from $4.50 to $4.00 because Argon at its introductory rate replaced Gemini 3.1 Pro Preview. At the standard rate Argon blends to (3 × $4 + $20) ÷ 4 = $8.00; if the other six entries stay put, the median would be $6.00 (our estimate). Olakai (secondary) notes that Argon's standard price would equal Muse Wonder and Opus 5.5 at $4/$20. GPT-6 Astra at $10/$50 is reported at 2.5 times GPT-5.6 Sol; GPT-6.1 Sol is one-fifth of Astra, which checks ($10 ÷ 5 = $2; $50 ÷ 5 = $10).

Per-token price is not per-task price. Sonnet 5.5 kept Sonnet 5's $2/$10 and halved cache reads from $0.20 to $0.10, with up to 30% lower cost per task claimed by Anthropic. The Oct 7 Haiku post presents the cache cut as part of that launch, although the Sep 28 page already lists $0.10; we follow $0.10 as in effect. Per Artificial Analysis (via a Substack review), Sonnet 5.5 at max effort uses about 193,000 output tokens per task, the highest recorded, so its cost per task exceeds Opus 5.5's. Budget on cost per task at a chosen effort level, not on the price card.

5.3 Subscriptions, free tiers and usage billing

Gemini free tier (in effect since Oct 9). According to Tech Insider and Shattered.io, citing 9to5Google's Oct 3 report and Google support pages, free users are limited to Gemini 3.5 Flash-Lite, down from three models; Plus subscribers lost Pro, and Deep Think moved down to AI Pro. No Google post explains it. Prices conflict: Tech Insider gives Ultra $99.99, while Kingy lists $199.99. NeuralTrust's "from $99.99" suggests an entry and a top tier, but the material does not settle it.

ChatGPT after DevDay (Sep 29). Skillmammoth, reading chatgpt.com/pricing on Oct 2, reports a new Pro 500 tier and a reopened Pro 200 at the same $200 with a lower allowance. Holders active Sep 22–29 keep the old allowance until Oct 29 (upcoming). Pro and Business Premium now include an always-on agent, a "dot", and Go at $8 "may include ads". Fast mode draws the allowance at 2.5 times and Ultrafast at 8 times. A US student offer of four free Plus months runs through Oct 31.

Anthropic. Since Oct 7, Max and Team plans include monthly API credits; the amount is not stated.

Kingy's subscription medians were unchanged at $20 (standard) and $200 (top). The changes were in access and allowances, not sticker prices: free users get less, heavy users get a smaller allowance at the same price, and API credits are being bundled into subscriptions.

6. Profitability and cost structure

Chapter 5 covered price moves; this chapter asks what margin is left after serving costs and compute bills, and who carries them. The only company statement below is Anthropic's June 1, 2026 announcement that it confidentially submitted a draft Form S-1; the rest is press-reported or an analyst estimate.

6.1 Anthropic's profit claim and the draft S-1

At an Oct 14, 2026 investor day, Anthropic reportedly said Q3 revenue was $16.8B (reported by attendees; no filing; gross-or-net basis not disclosed) against $11.6B in Q2, and that it was profitable "on an adjusted basis". Growth is 16.8 ÷ 11.6 − 1 = +44.8%. Business Insider says "almost 50%"; we follow the arithmetic. Q1 is $4.7B per attendees but $4.8B per the WSJ. Earlier FT and Bloomberg reports had Q2 at $11.5B.

No profit amount was disclosed and Anthropic "has not disclosed its methodology"; the FT says the adjustment excludes stock-based compensation. Outside figures disagree: New Market Pitch cites an 11% adjusted operating margin (0.11 × 16.8 = about $1.85B), while First Page Sage's model shows 5.4% on $17.3B (about $0.94B), revenue that does not match the reported $16.8B. We treat neither as fact.

The draft S-1 is unpublished; the figures come from Reuters (Sept 29) and the WSJ:

Arithmetic: FY2025 4.6 − 12.65 = −8.05, consistent with "over $8B". Compute is 7.33 ÷ 12.65 = 58% of expenses and 7.33 ÷ 4.6 = 159% of revenue. In H1 2026, 80.7 − 72 = about $8.7B of loss remains after the non-cash charges (our subtraction, not a company measure), and compute is 9 ÷ 16.3 = about 55% of revenue. Q1 + Q2 = 4.7 + 11.6 = 16.3 matches H1. Compute per revenue dollar fell, but "adjusted" remains undefined.

Revenue basis matters: Anthropic books cloud-partner sales gross. In 2025 these were about $2.16B (2.16 ÷ 4.6 = 47% of revenue), with $351M of distribution fees (351 ÷ 2,160 = 16%). The FT reported gross margin of 80% before partner revenue share; First Page Sage models 52% for Q2 2026 with partner shares as costs (estimate). These are not comparable.

6.2 Margins: AI-native software, SaaS and the labs

The gap is a sample effect. Aleph finds no median compression, with software margin at 79–81% over four years. ICONIQ's AI-product cohort sits 80 − 53 = 27 points lower. CloudZero's illustration: AI cost at 8% of revenue turns an 80% margin into 72%.

For labs, OpenAI's gross margin was about 39% in Q1 2026 against about 33% a year earlier (reports). Its FY2025 accounts, reported by Fortune and the FT, show revenue $13.07B and cost of revenue $7.5B, so (13.07 − 7.5) ÷ 13.07 = 43%; the same compilation quotes a 33% company-wide margin. Definitions are unstated, so we do not mix them. Operating loss is clearer: $20.92B on $13.07B in FY2025; $9.3B on $5.7B in Q1 2026 and $12.3B on $6.7B in Q2, both including stock compensation: 163% and 184% of revenue.

6.3 Compute commitments against revenue

Anthropic's draft S-1 reportedly commits at least $518B over ten years to six partners, about 80% non-cancellable (reported). Named items are Google $111B, AWS $110B, Microsoft $31B and $161.2B of Broadcom-related leases. These four sum to 413.2, which is 79.8% of 518 and matches the 80% figure. xAI (up to $84.5B, mostly cancellable on 90 days' notice) and AMD (over $20B) are the flexible items. The WSJ review gives "up to $750B"; we use the lower, firmer $518B.

The even ten-year spread is our assumption; no payment schedule is reported. Revenue is booked gross, so net ratios would be higher.

For OpenAI, a July 2026 presentation reported by the FT on Sept 18 projects $856B of compute and infrastructure spend through 2030, $840B of cumulative revenue and $278B of negative free cash flow for 2026–2030 (internal projections). 856 ÷ 840 = 1.02×: planned compute equals planned revenue. The WSJ puts compute near $750B. Against the annualized run rate approaching $50B at end-September (reported; FT, Reuters), 856 ÷ 50 = 17×.

Anthropic's "adjusted profit" is an undefined, reported claim; its fixed claim is $518B of commitments, 7.7× annualized Q3 revenue and about 80% non-cancellable.

6.4 Who bears the cost

Labs carry the fixed risk: if spend falls short, Anthropic "must pay Google the difference". A Blackstone-led debt facility closed in October 2026 (reported).

Clouds carry the capex but hold backlog. 2026 guidance: Alphabet $195–205B (raised Jul 22), Meta $130–145B, Microsoft about $175B (calendar), Amazon $200–220B (Goldman coverage), Oracle $50–70B (FY2027). The sum of ranges is $750–830B against about $413B in 2025, nearly double (800 ÷ 413 = 1.9×). Backlogs: Microsoft $678B, Oracle $664B. Amazon, Google and Microsoft are each Anthropic's investor, customer, supplier and competitor.

Concentration lifts the risk. Two customers supplied about 12% each (24%) of Anthropic's 2025 revenue. Bain's funding-gap maths is in Chapter 7.

7. The monetization gap and market reaction

Chapter 6 asked who carries the cost of AI; this chapter asks whether revenue is keeping pace, beginning with the market's verdict on Oct 8.

7.1 Oct 8: one revenue report moves the suppliers

On Thursday Oct 8, 2026, the Financial Times reported that OpenAI told investors its annualized revenue was roughly $50B at the end of September; CNBC separately confirmed it. A figure of $68B–$70B had circulated in late September. A person familiar with the matter said the higher number included partners' gross revenue, to compare directly with Anthropic. These are the two numbers Chapter 2 defines: about $50B on OpenAI's own booked basis, and $68B–$70B including partner gross. The gap is $68B − $50B = $18B to $70B − $50B = $20B, attributed mainly to accounting, not lost sales. On Oct 9, Bloomberg reported that OpenAI still expects a run rate of $70B or more by year-end (reported target).

The stated cause of the selloff was this measurement difference, which made investors ask whether a major AI customer can pay for the compute it has ordered. Oct 8 closes, from one source:

Source: 24/7 Wall St.. The Nasdaq Composite fell 1.25%, its worst session since mid-August (Admirals, citing Wealth Professional).

Oracle's exposure is the clearest. Yahoo Finance (via BeInCrypto) says a reported $300B five-year OpenAI agreement is about 45% of its $664B backlog (300 ÷ 664 ≈ 45%), and its latest quarter showed $28.5B of capex and −$5.4B of free cash flow. CoreWeave had a second worry: Benzinga reported on Oct 7 a clause in a $2.6B loan that can trigger a "Cash Trap Event" if a customer contract is seriously breached. TIKR stresses this is a risk, not a breach that has occurred.

The Oct 8 selloff priced in dependence on one unaudited buyer, not lost demand: OpenAI is private, publishes no audited figures, and is expected to list in 2027.

7.2 The funding gap: Bain's arithmetic and its critics

Bain & Company's Technology Report 2026 (Sep 29, 2026) sets the benchmark; every value is an estimate:

Source: Bain & Company. At the low end the gap is $6.0T − $1.2T = $4.8T (our calculation).

Bain calls the 25% capex share "ambitious but reasonable". Search and ads ($100B–$200B), autonomous systems ($400B) and physical AI ($900B) total up to $1.5T, leaving $2.7T ($4.2T − $1.5T) from products that do not yet exist. A year earlier Bain put the need at $2T by 2030, with an $800B shortfall, so the requirement has tripled. David Crawford said funding the build needs about 1 percentage point added to annual global GDP growth.

This is a hurdle rate, not a forecast, as Technext24 (commentary) notes. It also cites undated estimates: Goldman's roughly $300B break-even AI revenue for five hyperscalers, with AI cloud revenue about $70B above its pre-AI trend, a gap of about $230B a year ($300B − $70B); and Sequoia's David Cahn at about $600B. Semafor (Sep 15, via Aju Press) estimated four hyperscalers will spend about $66B more than operating cash flow over six quarters. Opinion: Marin Ivezic (PostQuantum, Sep 26) argues the builders need about $2.5T a year, equal to 250 million jobs against 159 million US nonfarm jobs.

7.3 Adoption and pricing: seat, usage or outcome

Revenue growth is real but concentrated. The State of AI Report (Air Street Capital, Oct 8, 2026) says OpenAI and Anthropic's combined reported run rates reached about $105B by late summer, from about $30B at the start of the year, roughly 3.5x. Ramp's August data show the median top-1% spender at $7,205 per employee per month against $12.50 for the median firm (about 576x), and 1% of customers drove about 80% of observed OpenAI and Anthropic spend; both describe Ramp's sample. Creatio's vendor survey (Oct 20, 600+ leaders) finds 70% beyond experimentation but only 17% "AI Adoption Leaders".

Vendors answer with new meters. List prices below were read on Oct 18, 2026 by QueryNow, a vendor-run comparison, so treat them as secondary:

Seats are predictable but decoupled from output, usage charges even failed attempts, and per-result billing puts failed attempts on the vendor. At three actions per conversation, Flex Credits cost 3 × $0.10 = $0.30, or $300 per 1,000 conversations (Macha, vendor blog); the $2 conversation meter is cheaper only above 20 actions. Storylane (Oct 1, vendor opinion) argues for hybrid pricing, citing HFS Research that buyers prefer outcome pricing in principle while contracted adoption lags.

The motive is margin. According to I/O Fund (Oct 16), ServiceNow's subscription gross margin fell from 83.0% to 80.5% and Microsoft Cloud's from 69% to 65% between June 2025 and June 2026, while AI-native apps run 25%–60%. Prices are falling while usage meters spread, so enterprise software alone cannot close Bain's gap.

8. Commercialization trends

Chapters 2–7 set out revenue, users, prices, margins and the market's reaction. This chapter reads them together and adds a dated watch-list. It uses only facts from earlier chapters.

8.1 What the month adds up to

Revenue concentrates in two labs, and the bases differ. Anthropic's $65B run-rate (end of July) and OpenAI's about $50B (end of September) are the two largest model-vendor figures, but they are not like-for-like: Anthropic books cloud-partner sales gross and OpenAI books its own share. The gap that moved markets on Oct 8 was 70 − 50 = $20B, or 29% of the gross-up figure. Anthropic's own gross-versus-net effect looks smaller: $351M of distribution fees on $2.16B of cloud-channel sales in 2025, which is 16% of that channel and 7.6% of $4.6B revenue.

Growth is slowing from a very high rate. Anthropic's quarterly growth fell from +147% to +45%, and Q3 × 4 = $67.2B is only about 3% above the July run-rate. OpenAI needs about +40% in one quarter to reach the reported $70B-plus year-end expectation (70 ÷ 50).

Usage is wide, paying is thin. Paying shares stay in single digits where they can be computed: 5.6% of ChatGPT weekly users (50 ÷ 900, from Feb) and at most 6.7% of Copilot's commercial seats. Ramp's index shows 55.9% of its US business customers paid for AI in September, against 19.8% of US businesses in the Census survey, so the sample matters.

Prices are splitting. Small models got cheaper at once (median $0.53 to $0.26), while frontier medians moved less ($4.50 to $4.00) and Argon's price doubles on Dec 1. Lower per-token prices can still raise spend if usage grows faster; the material does not show that directly, so treat it as a question to watch.

Margins and commitments set the pace. Anthropic's S-1 FY2025 shows compute of $7.33B on revenue of $4.6B, or 159% of revenue. Commitments of at least $518B are 7.7 times annualized Q3 revenue (518 ÷ 67.2). AI product gross margins are estimated at 45% in 2025 against 80% for software.

8.2 What it means by reader

Compare AI revenue only on one basis, treat user counts as reach and not income, and read every profit claim with its definition.

8.3 Key dates

Dates already past on Oct 24, 2026 are marked "In effect".

9. References

This document draws on the following 100 sources, listed in the order they are first cited.

1. OpenAI Expects $70 Billion in Annualized Revenue by End of 2026 — finance.yahoo.com
2. OpenAI annualized revenue $20 billion less than previously reported — axios.com
3. OpenAI's annualized revenue $20 billion lower than prior investor estimates — finance.yahoo.com
4. Anthropic revenue jumps to $16.8 billion ahead of IPO | Jawlah — jawlah.co
5. OpenAI Dev Day 2026: Live Updates On The Latest ChatGPT And Codex Announcements — engadget.com
6. ChatGPT Statistics 2026: 1.2 Billion Weekly Users, Sourced — whizi.io
7. Microsoft Copilot Statistics: 30M Seats, 50M GitHub Users — whizi.io
8. AI Price Index: What AI Costs in October 2026 | Kingy AI — kingy.ai
9. Introducing Claude Haiku 5.5 \ Anthropic — anthropic.com
10. What's new in Gemini 4 Argon  |  Gemini Generate Content API (Legacy)  |  Google AI for Developers — ai.google.dev
11. Podium Surpasses $150M in AI Agent ARR in 2.5 Years — prnewswire.com
12. Chai AI passes $100M ARR with a team of around 15 | Dealroom News — dealroom.co
13. Nebius Vs CoreWeave: Which AI Stock Has the Upper Hand? — io-fund.com
14. Anthropic Reveals $16.8B Revenue Amid IPO Preparations - Business Insider — businessinsider.com
15. Anthropic's draft S-1 reveals $16.3B revenue, $80.7B loss and up to $750B compute bill | Dealroom News — dealroom.co
16. Anthropic's landmark IPO filing shows 12-fold revenue jump, $518B compute bill - InvestmentNews — investmentnews.com
17. OpenAI's Revenue Is Reportedly $20 Billion Lower Than Thought. Nvidia Just Lost $170 Billion - 24/7 Wall St. — 247wallst.com
18. OpenAI’s revenue is reportedly $20 billion less than previously projected — techcrunch.com
19. Microsoft (MSFT): Azure AI Growth Keeps Capex Returns in the Spotlight - Alphastreet — news.alphastreet.com
20. OpenAI projects $70 billion annualized revenue by end of 2026 — qz.com
21. Nasdaq-100® Inside the Index Microsoft Corporation — nasdaq.com
22. SpaceX Reports Third Quarter 2026 Results — s21.q4cdn.com
23. SpaceX Q3 Revenue Jumps 161% on AI Compute Boom — technology.org
24. Musk says Tesla and SpaceX formed special committees to discuss merger: Live updates — cnbc.com
25. SpaceX's AI business is beating expectations | TechCrunch — techcrunch.com
26. How Has Microsoft's Story Changed In A Year? — finance.yahoo.com
27. OpenAI’s Revenue Appears $20 Billion Lighter Than Reported — and the Gap Is Accounting, Not Demand — finance.yahoo.com
28. What will OpenAI's annualized recurring… — FutureSearch — futuresearch.ai
29. Can OpenAI ever become profitable? – New Market Pitch — newmarketpitch.com
30. OpenAI Revenue and Valuation Statistics (October 2026) — explodingtopics.com
31. Anthropic revenue jumps to US$16.8 billion in quarter before IPO - The Business Times — businesstimes.com.sg
32. Anthropic says it is profitable: What Claude-maker's Q3 revenue of $16.8 Billion does and does not show | Company Business News — livemint.com
33. Anthropic Reportedly Generated $4.6 Billion in Revenue and Lost $42 Billion in 2025. Will This Impact Its Targeted $2 Trillion Valuation? — finance.yahoo.com
34. Anthropic Reportedly Posted Quarterly Revenue Growth of 50% and Was Cash-Flow Positive. Here's Why It Might Not Be Enough to Impress Investors Before the Company's Massive IPO | The Motley Fool — fool.com
35. Anthropic IPO: What Investors Need to Know After Draft Prospectus Leak — tradingkey.com
36. Claude Statistics 2026: Users, Revenue and Valuation — usecarly.com
37. Podium Surpasses $150 Million In AI Agent Annual Recurring Revenue And Deploys More Than 20,000 AI Agents — pulse2.com
38. Why did SpaceX buy Cursor? – New Market Pitch — newmarketpitch.com
39. Axya closes $17-million Series A to modernize manufacturing procurement | BetaKit — betakit.com
40. Claude Code vs Cursor (2026): Benchmarks, Pricing & Security | NeuralTrust — neuraltrust.ai
41. Cursor Statistics 2026: Key Numbers, Data & Facts — gradually.ai
42. Claude Code vs Cursor — Side-by-Side Comparison (2026) | VibeCompare — vibecompare.dev
43. AI Coding Models Statistics 2026: $1 vs $14 per Task — preuve.ai
44. CRWV vs NVDA: Weighing High-Growth Potential Against Established AI Leadership — tickeron.com
45. CoreWeave (CRWV) Stock Crashes 10% on AI Funding Fears — stocksdownunder.com
46. Nvidia Stock Forecast: Rubin Ramp and $108 Billion Q3 Outlook Keep AI Growth Accelerating — tradingkey.com
47. Nvidia Stock News (Oct 19, 2026): NVDA Falls 1.4% to $214.13 | Gotrade — heygotrade.com
48. Google Gemini Statistics 2026: 1 Billion Monthly App Users — whizi.io
49. Meta and Microsoft Cut Internal Claude Use as In-House AI Tools Gain Ground | MLQ News — mlq.ai
50. Gemini Statistics 2026: Users, Usage and Market Share — usecarly.com
51. Google Gemini Statistics 2026: Users, Market Share & More — gradually.ai
52. Copilot Adoption Statistics 2026, Fact-Checked | Eyal Marcus — eyalmarcus.com
53. Microsoft Copilot Statistics 2026: 30M Seats, 3.3% Paying | IP — instantpress.co
54. ChatGPT Stats October 2026 - Usage, Market Share, and More. - fatjoe. — fatjoe.com
55. Prediction: Microsoft Will Double Its Paid Copilot Seats in the Next 12 Months | The Motley Fool — fool.com
56. AI Usage Statistics October 2026 - Adoption, Market Share, and More. - fatjoe. — fatjoe.com
57. AI Adoption Is Levelling Off: Where Value Comes From Next — imobisoft.co.uk
58. AI Adoption Statistics 2026: Business & Enterprise Data — aibusinessweekly.net
59. AI Tools for Business: Prices, Adoption and Evidence — convocore.ai
60. Generative AI Market Briefing — vishalsingh.org
61. What's new in Gemini 3.8 Flash  |  Gemini API  |  Google AI for Developers — ai.google.dev
62. Introducing Claude Sonnet 5.5 \ Anthropic — anthropic.com
63. Claude 5.5 is brilliant, confusing, and very hungry — dreasays.substack.com
64. OpenAI DevDay 2026: What It Means for Marketing Analytics — improvado.io
65. Meta Muse Wonder Benchmark: Ties Claude at 56 2026 — shattered.io
66. Gemini 4 Argon is now available for everyone — blog.google
67. Introducing Gemini 4 Argon — blog.google
68. What Is Kingy AI? Tools, Reviews & Pricing (2026) — aiidelist.com
69. Anthropic Made $16.8B. Why Are Investors Nervous? - Olakai — olakai.ai
70. Google Gemini Free Tier Cut to 1 Model, Oct. 9 2026 — tech-insider.org
71. Google Cuts Free Gemini From 3 Models to 1 Oct. 9 2026 — shattered.io
72. Gemini vs ChatGPT (2026): Benchmarks, Pricing & Verdict | NeuralTrust — neuraltrust.ai
73. ChatGPT Pricing in 2026: Every Plan After DevDay — skillmammoth.com
74. ChatGPT Pricing 2026: Every Plan, Free Tier & Catch — lmspedia.org
75. Release notes | Claude Help Center — docs.anthropic.com
76. Anthropic Q3 2026 Revenue Reaches $16.8 Billion: Growth, Profitability, and IPO Outlook — kucoin.com
77. Anthropic Profitability Analysis & Statistics: October 2026 – First Page Sage — firstpagesage.com
78. What's a good SaaS gross margin? — getaleph.com
79. AI gross margin: how AI spend hits SaaS profitability — cloudzero.com
80. OpenAI Revenue 2025–2026: $13B Financials & ARR Data — useluminix.com
81. OpenAI Statistics 2026: Revenue, Users & Funding — gradually.ai
82. OpenAI Revenue at $50bn: Why AI Stocks Fell in October — admiralmarkets.com
83. AI Infrastructure Spending Forecast 2030: 12 Firms Compared — softwarestrategiesblog.com
84. Frontier AI energy economics: tokens per MWh | EarthSync — earthsync.io
85. OpenAI Revenue Comes in at $50 Billion Annualized, Below Earlier Reports, Sending AI Stocks Lower · TechEchelon — techechelon.com
86. OpenAI's Revenue Disclosure Rattles AI Stocks: $50 Billion Figure Falls Short of $68 Billion Reports | Traders Agency — tradersagency.com
87. Oracle Stock Crashes After OpenAI's Revenue Gap. Who's Next? — finance.yahoo.com
88. CoreWeave Stock Fell Nearly 8% in a Day on an OpenAI Revenue Report. Here’s Why Its Borrowing Model Raises the Stakes | TIKR.com — tikr.com
89. New Innovation Is Required to Fund AI’s $6 Trillion Buildout | Bain & Company — bain.com
90. AI market needs to make $6 trillion a year by 2031 to fund its infrastructure habit — theregister.com
91. AI must generate $6 trillion and invent $4.2T in products to justify the data centre boom — technext24.com
92. AI Data Centers Face $6 Trillion Revenue Challenge, Big Tech Warned - Aju Press AMP — m.ajupress.com
93. Paying for AI Data Centers Takes 250M Jobs. US Has 159M — postquantum.com
94. AI inference is becoming a vast business | State of AI Report 2026 — stateof.ai
95. Creatio Releases The State of Enterprise AI Adoption Report: An In-Depth Look at Where Enterprise AI Stands — desmoinesregister.com
96. AI Agent Pricing 2026: Per Seat, Per Action, Per Result | QueryNow® — querynow.com
97. Agentforce Pricing 2026: 10 Cents an Action, $2 a Resolution — getmacha.com
98. Salesforce Agentforce Pricing (2026): Costs Explained — pluno.ai
99. AI Sales Agent Pricing: Seat vs Outcome-Based Models — storylane.io
100. How Agentic AI Is Squeezing SaaS Margins and Growth — io-fund.com