DEAN BAKER AND ED ZITRON: The revenue story behind the AI labs is not holding up, with Anthropic's run rate flattening and price competition forcing cheaper models

The Gist

The AI companies' plans only work if their sales keep soaring. Recent numbers suggest Anthropic's growth stalled, and OpenAI's widely reported revenue figure turned out to be an investor estimate about $20 billion above OpenAI's own net number. Meanwhile, cheap Chinese models are forcing price cuts, and popular free apps don't yet earn much. This packet steelmans the arguments made by Ed Zitron and Dean Baker in their best light for analysis and is not an endorsement of their conclusions.

Conclusion

The labs' commitments assume revenue that compounds rapidly for years, but recent evidence points the other way: Anthropic's annualized revenue appears to have flattened, OpenAI's widely quoted headline figure turned out to be an investor gross-up about $20 billion above OpenAI's own net figure, showing how loosely these metrics are reported, competition from low-cost Chinese models is forcing price cuts, and heavily promoted free consumer apps do not yet show a path to matching revenue.

Premises

1. Commitments of the size the labs have signed are payable only if revenue keeps growing very fast, on the order of several percent or more each month, for years.
2. An unofficial tracker put Anthropic's annualized revenue run rate at about $76.8 billion in late September 2026, up only about 1.8 percent from the prior quarter.
3. The widely reported figure of nearly $70 billion in OpenAI annualized revenue was not OpenAI's own number but an investor estimate that grossed up partner sales to match Anthropic's method; OpenAI's own figure, on a net basis, was about $50 billion, and AI stocks fell when the gap was reported on Oct 8, 2026.
4. Annualized revenue figures are not standardized: they can multiply a single strong month or day, and the labs count sales made through cloud partners differently (Anthropic records the full customer payment, OpenAI records only its share of certain partner sales), so headline figures can overstate durable revenue and are not comparable across labs.
5. In 2026 OpenAI and Anthropic cut prices on lower and mid-tier models to compete with cheaper Chinese open-weight models, which lowers revenue per unit of use.
6. High-profile AI consumer apps, such as Meta's Muse, gain users quickly by being free to download, which shows interest but not yet revenue at the scale needed.

Assumptions

• The Anthropic run-rate figure comes from a third-party tracker that an analyst has questioned; Anthropic itself told investors $65 billion at the end of July.
• On its own net basis, OpenAI's revenue is still rising quickly: it reported 77 percent run-rate growth in its third quarter of 2026 and 107 percent enterprise run-rate growth, and the FT, as summarized by other outlets, reported a July net run rate of about $30 billion in OpenAI's latest investor presentation, so the $50 billion figure does not show a shortfall against OpenAI's own earlier net figures.
• OpenAI told investors it expects annualized revenue of $70 billion or more by the end of 2026.
• Anthropic's run-rate figures count cloud-partner sales gross, so its $65 billion to $77 billion figures are not directly comparable with OpenAI's net $50 billion and would be lower on a net basis.
• Muse offers in-app purchases and paid subscriptions, so it has a monetization path, though its scale is unknown.
• Price cuts can expand total usage enough to raise revenue, so lower unit prices do not by themselves mean lower total revenue.
• The net-versus-gross clarification was reported on Oct 8, 2026, likely after the conversation was recorded.

Analysis

Argument type: Inductive.

Potential Fallacies

• Hasty generalization from a thin, contested datapoint (P2 and the conclusion ('appears to have flattened')) — A single quarter-on-quarter reading of about 1.8% from an unofficial, questioned tracker is used to suggest Anthropic's revenue has 'flattened'. Run rates are lumpy, and Anthropic's own $65B figure for end-July implies roughly 18% growth to $76.8B in about two months if the numbers are taken at face value. The sources appear to use different baselines or methods, so 'flat' depends on which series is chosen.
• Conflating measurement unreliability with revenue weakness (P3, P4 and the conclusion) — The $70B versus $50B gap reflects gross-versus-net accounting and an investor's harmonization, not a fall in revenue. OpenAI's net figure was never $70B, and under the stated assumptions it is growing quickly. The episode is good evidence for better disclosure but not directly for a revenue shortfall.
• Unsupported inference from unit price to total revenue (P5) — Lower revenue per unit of use only reduces total revenue if volume does not rise proportionally. No volume or elasticity data is given, and the argument's own assumptions grant that usage expansion could offset the cuts. The cuts also apply to lower and mid-tier models, not frontier models.
• Argument from absence of evidence (P6) — That a free consumer app does not 'yet' show revenue at scale is normal for a freemium product and says little about eventual monetization, especially since Muse has in-app purchases and subscriptions. Meta is also not one of the labs whose commitments are at issue, and most lab revenue is enterprise and API-driven.
• Unproven threshold / hidden premise (P1) — The requirement of 'several percent or more each month for years' is asserted without the commitment schedule, margins, or financing mix. Commitments may be staged, renegotiated, or funded by equity and debt, so 'payable only if' is stronger than shown. The argument also never compares observed growth against the threshold.
• Selective use of evidence / asymmetric skepticism (Conclusion; P2 versus P4) — Run-rate metrics are treated as unreliable when they flatter a lab (OpenAI's gross-up) but are accepted as informative when an unofficial tracker suggests flattening for Anthropic. Counter-evidence (Anthropic's own figure, OpenAI's growth and guidance) is acknowledged only in the assumptions, not weighed in the conclusion.
• Equivocation on 'run rate' across labs (P2, P3, P4) — Anthropic's gross figures and OpenAI's net figures are used as if they measured the same thing, though the assumptions state they are not comparable. This lets the evidence appear to converge when the series are not like-for-like.
• Temporal mismatch of sources (P3 and source framing) — The Oct 8 gross-versus-net clarification likely postdates the recorded conversation, so the argument blends later facts with the speakers' earlier remarks. This makes the case look more contemporaneous and coherent than the source supports.

Counterarguments

These challenges are not ranked.

• Premise 2 (High impact) — Anthropic's own investor figure of $65B at end-July, set against the tracker's $76.8B in late September, implies strong growth rather than a plateau. The 1.8% figure likely reflects a different baseline, methodology, or a spike in the prior-quarter base. A flat reading depends on choosing the weakest-sourced series.
• Premise 3 (High impact) — The $20B gap is a definitional artifact of an investor's gross-up, not a revenue shortfall. On OpenAI's own net basis, run-rate growth was 77% in the quarter, enterprise run-rate growth 107%, and the net run rate rose from about $30B in July to about $50B, with guidance of $70B or more by year-end. This is evidence of sloppy reporting alongside strong growth.
• Premise 1 (High impact) — Commitments are typically staged, conditional, renegotiable, or financed through equity, debt, and partner balance sheets. Without the contract terms and funding mix, the claim that they are payable only with multi-year monthly compounding is an assertion. A revenue shortfall may lead to renegotiation rather than default.
• Premise 5 (Medium impact) — Falling unit prices have historically coincided with rising total revenue in cloud, bandwidth, and semiconductors when demand is elastic. Cuts on lower and mid-tier models may reflect tiering and market expansion, while frontier models retain pricing power. Unit costs may also be falling, preserving margins.
• Premise 6 (Medium impact) — Free-first adoption followed by monetization is the standard path for consumer platforms. Muse has in-app purchases and subscriptions, and the labs' revenue is mainly enterprise and API, so a free consumer app is the wrong comparison class.
• Conclusion (Medium impact) — The run-rate series are noisy and capacity-constrained. A flat quarter could reflect compute supply limits rather than weak demand, and a single-quarter pause says little about multi-year compounding. Bubble skeptics have also been wrong on timing before, so the pattern of evidence is consistent with strong but opaquely measured growth.

Suggested Improvements

• Quantify the commitment-versus-growth test — Present the actual commitment schedules, financing mix, and flexibility terms, then compute the monthly growth rate required and compare it with observed growth on a consistent basis. P1 is the logical bridge between the evidence and the conclusion. Without the arithmetic, the argument never shows that observed growth falls short.
• Use consistent, primary-source revenue series — Restate Anthropic and OpenAI on a common net basis, at monthly frequency where possible, and backtest the tracker against Anthropic's disclosed figures such as the $65B end-July number. This would resolve the conflict between the tracker and the company figure and make cross-lab comparisons valid.
• Separate the epistemic claim from the substantive claim — Frame the supported finding as 'headline revenue metrics are unreliable and growth durability is uncertain' and treat 'growth has stalled' as a separate claim requiring its own evidence. The metric-reliability point is well supported, while the stalling claim is not. Keeping them apart avoids overreach and makes the strong part harder to dismiss.
• Add volume and elasticity evidence for the price cuts — Provide token-volume, enterprise-spend, and retention data before and after the 2026 price cuts, ideally compared against unchanged models. Whether price cuts reduce or raise total revenue depends on elasticity, which the argument currently leaves unaddressed.
• Use a relevant comparison for consumer monetization — Replace or supplement the Muse example with paid-conversion and subscription data for the labs' own consumer products, and show the revenue mix between consumer and enterprise. Muse is peripheral to the labs' revenue, and a mix breakdown would show how much the consumer-monetization question matters.
• Preserve source timing and attribution — Distinguish what the speakers said at the time of recording from later reporting such as the Oct 8 clarification, and label the tracker as unofficial wherever it is cited. This avoids attributing later facts to the speakers and keeps the evidential weight of each source transparent.

Scenario Tests

• Anthropic's own reported net run rate keeps rising at double-digit rates quarter after quarter, and the tracker is shown to have a mismatched baseline. (Challenges) — The central 'flattening' pillar collapses, leaving only the metric-reliability point and general pricing pressure.
• OpenAI reaches its $70B+ year-end annualized target on a net basis. (Challenges) — The company's own consistent net numbers would show compounding revenue, undercutting the claim that the revenue story is failing, though the headline-metric criticism would still stand.
• Token volume and enterprise spend rise sharply after the 2026 price cuts, with total revenue per model tier increasing. (Challenges) — P5 becomes neutral or positive evidence for revenue, and price competition looks like market expansion rather than erosion.
• Run rates stay flat for several consecutive quarters on a consistent net basis while compute payments step up on schedule, with limited contract flexibility. (Supports) — The commitments would become hard to service from revenue alone, forcing renegotiation, deferral, or new capital, and exposing counterparties such as Oracle and neoclouds.
• Commitments are revealed to be staged, cancellable, or largely partner-financed. (Challenges) — P1's 'only if' logic weakens: a revenue shortfall would produce renegotiation rather than default, reducing the significance of any slowdown.
• Further disclosures show other lab or supplier revenue figures were inflated by inconsistent gross-versus-net or single-month annualization. (Supports) — This would reinforce the claim that headline metrics are loosely reported and that market valuations rest on unreliable numbers, even if it does not by itself show a revenue decline.

Coherence & Relevance

The argument is a convergent, cumulative-case inductive structure with a clear organizing question: can revenue compound fast enough to service the commitments? It is coherent in form, but its strands are unequal. The metric-unreliability strand (P3, P4) is strong yet addresses information quality, not revenue trajectory. The strands that bear on trajectory (P2, P5, P6) are thin, contested, or offset by the stated assumptions. The conclusion's confident phrasing ('points the other way') outruns what the premises support, and what survives is a more modest claim: reported revenue figures are unreliable, price pressure exists, and the durability of rapid growth is uncertain.

• Commitments of the size the labs have signed are payable only if revenue keeps growing very fast, on the order of several percent or more each month, for years. (Strong) — Sets the benchmark for the whole argument, but no premise computes observed growth against it, and contract terms and financing flexibility are absent.
• An unofficial tracker put Anthropic's annualized revenue run rate at about $76.8 billion in late September 2026, up only about 1.8 percent from the prior quarter. (Strong) — Directly relevant to growth, but a single quarter from a questioned source that conflicts with Anthropic's own figure; the baseline mismatch is not reconciled.
• The widely reported figure of nearly $70 billion in OpenAI annualized revenue was not OpenAI's own number but an investor estimate that grossed up partner sales to match Anthropic's method; OpenAI's own figure, on a net basis, was about $50 billion, and AI stocks fell when the gap was reported on Oct 8, 2026. (Moderate) — Relevant to metric reliability but not to revenue trajectory. It also postdates the recording and shows a reporting-basis correction rather than a decline.
• Annualized revenue figures are not standardized: they can multiply a single strong month or day, and the labs count sales made through cloud partners differently (Anthropic records the full customer payment, OpenAI records only its share of certain partner sales), so headline figures can overstate durable revenue and are not comparable across labs. (Moderate) — Supports caution about all run-rate figures, including the tracker figure used in P2, so it weakens the argument's own evidence as much as the labs' claims.
• In 2026 OpenAI and Anthropic cut prices on lower and mid-tier models to compete with cheaper Chinese open-weight models, which lowers revenue per unit of use. (Moderate) — Missing volume and elasticity data, unit cost trends, and the share of revenue from the affected tiers.
• High-profile AI consumer apps, such as Meta's Muse, gain users quickly by being free to download, which shows interest but not yet revenue at the scale needed. (Weak) — Meta is not a lab whose commitments are at issue, 'scale needed' is undefined, and lab revenue is mainly enterprise and API.

View this argument on LogicFirst.ai