STEVE EISMAN: The leading labs' latest revenue signals point to slowing growth, consistent with the end of cost-insensitive AI usage

The Gist

Steve Eisman says the two big AI labs are growing more slowly than before, which fits his view that companies are reining in AI spending. This is a steelman: Steve Eisman's arguments are reconstructed in their strongest fair form, not endorsed.

Conclusion

Recent revenue data from OpenAI and Anthropic show decelerating growth that is consistent with the end of cost-insensitive usage, which weakens the case that their revenue will keep pace with their commitments.

Premises

1. OpenAI's reported annualized revenue was about $50 billion at the end of September, after a figure near $70 billion circulated a month earlier, and annualized revenue is not a standardized measure.
2. Anthropic's third-quarter revenue rose about 43 to 50 percent from the second quarter, down from more than doubling in the second quarter.
3. The deceleration coincides with companies curbing unlimited AI usage.
4. A lab that needs public capital has an incentive to list while its growth figures still look strong.

Assumptions

• The roughly $70 billion OpenAI figure was an investor restatement on a gross basis; OpenAI's own figure was about $50 billion net, so the gap reflects accounting basis rather than lost revenue, a point Eisman partly concedes ("not apples to apples").
• Anthropic also reported adjusted profitability, an 85 percent adjusted gross margin, and positive cash flow in the third quarter.
• Some deceleration from extreme growth rates is expected, as Michael Batnick noted.

Analysis

Argument type: Inductive.

Potential Fallacies

• Correlation taken as causation (post hoc) (Premise 3 and Conclusion) — The premise says deceleration 'coincides with' companies curbing unlimited usage, and the conclusion reads this as 'consistent with' the end of cost-insensitive usage. Overlap in timing does not show that curbs caused the slowdown, nor does it say how much revenue came from such usage. Base effects, compute limits, pricing changes, competition and vendor-imposed caps could produce the same pattern.
• Mere consistency treated as support (affirming the consequent pattern) (Conclusion) — If cost-insensitive usage ended, growth would slow, and growth did slow. But growth would also slow under many other explanations, especially the normal deceleration the assumptions concede. Without evidence that discriminates between explanations, 'consistent with' barely raises the credibility of the specific causal story.
• Unsupported leap / missing premise (Conclusion) — The step from slower growth to 'revenue won't keep pace with commitments' needs a benchmark: the size and timing of the commitments and the growth rate required to meet them. Growth of about 45% per quarter is still very high in absolute terms and may exceed what is required. Without this, the conclusion is asserted rather than shown.
• Motive offered as evidence (Premise 4) — An incentive to list while growth looks strong is true of nearly any firm seeking capital. It justifies scrutiny of disclosures but does not show that growth is weakening or that figures are being managed, and it does not clearly support the revenue claim.
• Non-comparable measures juxtaposed (equivocation) (Premise 1) — Setting a $70B gross-basis figure beside a $50B net-basis figure invites readers to see a decline. The argument concedes they are 'not apples to apples', so the pairing adds suggestion rather than evidence of slowing.
• Hasty generalization from thin data (Conclusion) — One quarter at one lab and an ambiguous figure at another are extended to a forward-looking claim about multi-year revenue against long-term obligations, and to a sector-wide shift in usage behavior. The two labs also have different business mixes, so treating them as a single pattern is questionable.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Deceleration from more than 100% to about 45% quarterly growth is the mathematically and economically expected path for a hypergrowth business. Growth of 45% per quarter is several hundred percent annualized, so the data are fully compatible with revenue meeting commitments. Without a base-rate comparison to other hypergrowth firms, the slowdown carries little new information.
• Premise 1 (High impact) — Given the gross-versus-net explanation, the $70B-to-$50B change is an accounting artifact and not a revenue decline. One of the argument's two empirical pillars therefore does no work, leaving only Anthropic's data to support a claim about both labs.
• Premise 3 (High impact) — Rival explanations fit the same data: compute or capacity constraints, falling price per token with rising volume, competition from other labs and open-weight models, enterprise procurement cycles, or lab-imposed rate limits. If usage volume is rising while prices fall, the story is deflation rather than demand weakness.
• Conclusion (Medium impact) — Anthropic's 85% adjusted gross margin and positive cash flow suggest healthy unit economics and less dependence on external capital, which cuts against the claim that commitments are at risk. Counter-caveat: these are adjusted, non-GAAP measures and do not transfer to OpenAI, which carries much of the commitment concern.
• Conclusion (Medium impact) — Whether revenue 'keeps pace' depends on the structure of commitments. Many are staged, conditional, renegotiable or backed by partners' balance sheets, and financing may come from strategic or sovereign investors rather than near-term revenue. Slower growth need not become distress.
• Premise 4 (Low impact) — The IPO-timing incentive is generic and sits in tension with the claim that growth is already weak. Private rounds and strategic partnerships offer alternative financing routes, and the logic, if accepted, would make any company's disclosures untrustworthy, including the figures the argument itself relies on.
• Premise 3 (Medium impact) — Customers curbing unlimited usage may be healthy: it can mean more disciplined, durable demand and better revenue quality, as in the 2022-23 cloud cost-optimization period, which slowed growth without ending the underlying trend.

Suggested Improvements

• Quantify the commitments benchmark — State the size, timing and conditionality of the labs' commitments, and compute the revenue path required to service them. Then compare it with the observed and projected growth rates. This is the missing bridge between 'slowing growth' and 'revenue won't keep pace'. Without it, deceleration cannot be classified as alarming or benign.
• Discriminate between explanations for the slowdown — Provide usage-level evidence such as token volumes, net revenue retention, seat counts, price-per-token trends, and enterprise surveys on usage caps. Separate customer-imposed from lab-imposed limits. This would show whether cost-insensitive usage actually ended, rather than relying on coincidence and 'consistent with' reasoning.
• Use a consistent revenue basis and drop or reframe the OpenAI comparison — Compare figures on a single stated basis (net versus gross, recognized versus run-rate), or present the OpenAI discrepancy explicitly as a metric-definition issue rather than as a growth signal. This removes a misleading implication of decline and focuses the argument on evidence that can bear weight.
• Add base-rate benchmarks — Compare Anthropic's deceleration with the growth paths of other hypergrowth firms such as cloud and SaaS leaders at similar scale, and state what rate of slowing would be abnormal. It would show whether this deceleration is unusual or merely expected.
• Integrate the countervailing evidence — Address how adjusted margins and cash flow affect commitment coverage, with their definitions, and distinguish Anthropic's position from OpenAI's. The bearish conclusion currently ignores facts the argument itself grants, which weakens its credibility.
• Add falsifiable predictions — Specify in advance what next-quarter growth rate or usage metrics would confirm or refute the end of cost-insensitive usage. A hedged 'consistent with' claim is hard to refute; pre-specified tests would make the thesis informative.

Scenario Tests

• Anthropic's growth re-accelerates next quarter following new model launches or agentic and enterprise rollouts. (Challenges) — The Q3 slowdown would look like lumpy timing or normal volatility, and the causal story about the end of cost-insensitive usage would lose support.
• Growth keeps falling faster than base effects predict (for example to the 20s or low teens quarter over quarter) while usage-cap evidence accumulates and net revenue retention declines. (Supports) — The deceleration would look structural and not merely mechanical, strengthening the claim that revenue may not keep pace with commitments.
• Disclosures show token volumes rising strongly while price per token falls, so revenue slows even as usage expands. (Challenges) — The slowdown would reflect deflation rather than demand exhaustion, undermining the cost-insensitivity mechanism while leaving revenue-versus-commitments risk open.
• The labs' compute commitments prove flexible, staged or backed by partners, or financing comes from strategic investors independent of near-term revenue. (Challenges) — Even sustained slowing would not translate into a failure to meet commitments, so the conclusion's link to commitments fails.
• A lab files for an IPO and its S-1 restates revenue on a standardized basis, showing a materially lower run rate than the figures circulated earlier. (Supports) — This would lend credibility to concerns about non-standard metrics and about timing disclosures to flatter growth, partly vindicating the P4 suspicion.
• Comparable hypergrowth companies (for example cloud leaders) that decelerated from 100%+ to roughly 45% quarterly growth continued compounding for years. (Challenges) — The reference class shows such deceleration is normal and not a reliable warning sign, supporting the view that the conclusion over-reads the data.

Coherence & Relevance

The argument is a loose cluster of considerations, not a tight chain. Its hedged conclusion is modest and defensible in a narrow sense, since one lab's real deceleration slightly weakens a perfect-growth assumption. But the OpenAI evidence is neutralized by the argument's own assumptions, the causal mechanism rests on coincidence, the IPO incentive is tangential, and the decisive step from slower growth to shortfall against commitments is never quantified. The stated assumptions (accounting basis, strong margins, expected deceleration) mostly cut against the bearish reading and are not reconciled with it. Overall the argument raises a legitimate question for investors rather than establishing an answer.

• OpenAI's reported annualized revenue was about $50 billion at the end of September, after a figure near $70 billion circulated a month earlier, and annualized revenue is not a standardized measure. (Weak) — Under the stated assumption the gap is accounting basis, so it does not indicate a growth slowdown. It is relevant mainly to the non-standardization of metrics and to the P4 theme of disclosure scrutiny, and it does not support the OpenAI half of the conclusion.
• Anthropic's third-quarter revenue rose about 43 to 50 percent from the second quarter, down from more than doubling in the second quarter. (Strong) — Directly establishes deceleration at one lab, but cannot show the cause or compare the new growth rate with what the commitments require. It also concerns Anthropic only, while the conclusion covers both labs.
• The deceleration coincides with companies curbing unlimited AI usage. (Moderate) — Offers the proposed mechanism, but only as coincidence. It lacks magnitude, share of revenue affected, and any way to separate demand-side discipline from vendor-imposed limits or other drivers.
• A lab that needs public capital has an incentive to list while its growth figures still look strong. (Weak) — A motive claim, not a revenue observation. It helps explain why timing and metric definitions deserve scrutiny, but does not connect to whether growth will keep pace with commitments.

View this argument on LogicFirst.ai