ED ZITRON AND WILL GUYATT: Run rates are undefined marketing figures, and the OpenAI net versus gross episode shows reporters should not repeat them without the method

The Gist

Ed Zitron says 'run rate' numbers are marketing, because companies rarely say how they were calculated. OpenAI's reported $70 billion turned out to be investors' gross-up of OpenAI's own $50 billion figure, and stocks moved on both stories. He wants reporters to stop using these numbers unless the method is spelled out. This packet steelmans the arguments Ed Zitron makes in this episode, in their best light, for analysis; it is not an endorsement of his conclusions.

Conclusion

Annualized run-rate figures for AI labs are not standardized measures of financial condition, because the period annualized, the treatment of partner sales, and the base are rarely disclosed; the October 2026 episode in which a roughly $70 billion OpenAI figure turned out to be an investor gross-up of OpenAI's roughly $50 billion net figure shows how such numbers move markets, so reporters should not repeat them without stating how they were calculated.

Premises

1. A run rate annualizes a shorter period of revenue, and the period used, whether a day, a week, a month, or 28 days, is often unstated.
2. Axios reported on Sept 29, 2026 that OpenAI's annual recurring revenue neared $70 billion without a stated calculation basis.
3. On Oct 8, 2026, the Financial Times reported that OpenAI's own figure was about $50 billion, and follow-up reporting explained that the higher figure came from investors restating OpenAI's revenue on a gross basis to match Anthropic's method.
4. Anthropic books the full value of sales made through Amazon and Google cloud marketplaces as revenue and records the partners' share as a sales and marketing expense, while OpenAI books certain partner revenue net; both treatments are permitted under accounting rules.
5. AI-linked stocks fell on Oct 8 after the $50 billion report and rebounded on Oct 9 after Bloomberg reported that OpenAI expects $70 billion or more by year-end, showing that loosely defined run-rate figures move prices in both directions.
6. A reporter who cannot state the period, the base, and the net or gross treatment of a run rate is relaying a company's preferred number rather than a verified financial measure.
7. Accountability for public claims made during a boom is what deters the next one, which requires recording what was said and how it was measured.

Assumptions

• OpenAI's $50 billion net figure is consistent with its own earlier net disclosures, and CNBC reported 77 percent quarter-over-quarter run-rate growth; the episode was a change of basis, not a revenue shortfall.
• In 2025 Amazon and Google kept about $351 million of $2.16 billion of Anthropic marketplace sales, about 16 percent, so restating Anthropic on a net basis would reduce 2025 revenue from about $4.6 billion to roughly $4.25 billion, not to the $2.43 billion figure given in the episode.
• Describing run-rate reporting as market manipulation is the speaker's characterization; no regulator has made such a finding.
• Run rates can be informative for fast-growing firms when the period and basis are disclosed.

Analysis

Argument type: Inductive.

Potential Fallacies

• Post hoc / unsupported causal attribution (P5 and the 'move markets' clause of the conclusion) — P5 treats a price fall and a rebound that followed two reports as showing that loosely defined run rates move markets. The rebound followed a forward-looking year-end projection, a different kind of claim from a run rate. Sector sentiment, Oracle-specific news and financing concerns could also explain the moves. The sequence is suggestive, but it does not isolate definitional looseness as the cause.
• Hasty generalization (Conclusion; P1) — A single well-documented OpenAI/Anthropic episode is used to support a claim about AI-lab run rates in general ('rarely disclosed', 'not standardized'). The episode shows the problem can occur and matters when it does. It does not show how often it occurs. A sample of coverage would be needed to establish frequency.
• False dichotomy / persuasive definition (P6; title) — P6 sets a fully specified, verified measure against a 'company's preferred number', leaving out the middle ground of an attributed, caveated, unaudited figure. It also calls the number a company's preference, though P3 says the $70B came from investors restating the figure. Much of the conclusion's force is built into the wording of the premise.
• Is-ought gap (missing normative bridge) (P6, P7 to the conclusion) — The premises are mostly descriptive (what run rates are, what happened), while the conclusion is prescriptive (reporters should not repeat them). The unstated bridge is that reporters have a duty not to relay measures whose basis they cannot state. P7 addresses record-keeping and deterrence, not that duty.
• Unsupported causal claim / appeal to consequences (P7) — P7 says accountability deters future booms but offers no mechanism or evidence, and the historical record on deterrence is mixed. The point works better as a claim about good record-keeping than as a claim about deterrence, and the conclusion does not need it.
• Terminological drift (P2, P3, P5) — P2 describes Axios's figure as 'annual recurring revenue', while the rest of the argument treats it as a run rate. ARR, run rate, trailing revenue and year-end guidance are distinct measures. The argument's point that the labels are used loosely is partly a demonstration of the problem, but the shift should be made explicit.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — The episode shows the information system working. Axios published a figure, the FT clarified it within about nine days, and follow-up reporting explained the gross-up. Both figures were valid under different permitted conventions, and sophisticated investors already discount run rates. Attribution plus prompt correction is a better remedy than a rule against repeating figures.
• Premise 6 (High impact) — Reporters covering private companies often cannot obtain the method. A rule of 'state the method or don't print' would thin out coverage or push it toward leaks and less careful outlets. The accepted practice of attributing a figure and flagging that its basis is unknown already guards against relaying a preferred number as verified.
• Premise 5 (High impact) — The Oct 9 rebound followed a forward-looking year-end projection, not a clarified run rate. Sector sentiment, Oracle's credit and financing news and other events also move these stocks. Without an event study, the price moves do not show that run-rate ambiguity drove prices.
• Premises 3-4 (Medium impact) — A2's own numbers put Anthropic's partner share at about 16%, which cannot alone account for a roughly 40% gap between $70B and $50B. Period choice, the revenue streams included, or other adjustments may contribute, so the episode may illustrate a broader definitional problem more than a clean gross/net mismatch.
• Conclusion (Medium impact) — If permitted accounting options (principal versus agent) make a figure 'undefined', then ordinary revenue, SaaS ARR and adjusted EBITDA are all undefined marketing figures. The argument has to distinguish metrics with accepted conventions from those with none. Treating gross and net as 'undefined' conflates a missing standard with a standard that allows options.
• Premise 7 (Low impact) — Accountability records existed after the dot-com and crypto booms and did not prevent later ones. Incentives and credit conditions drive booms more than documentation does, and the link between stating a run-rate method and deterring the next boom is unspecified.
• Conclusion (Medium impact) — The reporting norm addresses a symptom downstream of the source. Investors and labs generated the gross-up, and private companies face no disclosure standard. Lender or investor term-sheet requirements, or auditor-attested metrics, would act closer to the root cause.

Suggested Improvements

• Scope of the reporting norm — Reframe the conclusion from 'should not repeat without the method' to 'should state the period, base and gross/net treatment where known, and explicitly flag when the company or source has not provided them', with attribution to the party supplying the figure. This meets the argument's goal and is workable when private firms withhold methods. It also answers the objection that the rule would reduce public information.
• Evidence for prevalence — Add a content analysis of a sample of AI-lab run-rate reports, showing how many state the period, base and net/gross treatment. 'Rarely disclosed' is currently inferred from one episode. A sample would turn an anecdote into a measured frequency.
• Market-impact claim — Separate run-rate reports from forward projections, and support the claim with an event-study of AI-linked equities around report timestamps with sector and market controls. Otherwise, soften 'move markets' to 'coincided with market moves'. The urgency of the norm rests on P5, which is the weakest causal link. The norm still stands without it.
• Decomposing the gap — Break the ~$20B difference into gross/net treatment, period choice and other adjustments, and reconcile it with the ~16% partner share in A2. This would show how much of the discrepancy the headline mechanism explains and would prevent oversimplifying a multi-cause gap.
• Terminology — Define and keep apart run rate, ARR, trailing revenue and year-end guidance, and state which one each cited report used. Conflating these measures is part of the problem being criticized, and the argument's own premises should avoid it.
• Framing and sources — Reconcile the 'undefined marketing figures' title with A4 and P4, which concede that run rates are informative when disclosed and that gross and net are both permitted. Cite primary documents or company statements for the accounting treatments, and note the source's numerical error. This removes a gap between rhetoric and evidence and anchors the key facts in more than one speaker's account.
• Premise 7 — Either drop the deterrence claim or recast it as a claim about preserving provenance, such as archiving what was claimed and how it was measured, without asserting that this prevents future booms. The modest version is defensible, while the deterrence version adds an unsupported empirical burden.

Scenario Tests

• A reporter covering a private AI lab receives a run-rate figure from investors on background terms and cannot learn the period or the gross/net basis. (Challenges) — The strict form of the norm ('do not repeat') would mean silence or reliance on worse sources. The workable version is to print with attribution and an explicit note that the basis is unknown, which is a narrower rule than the one the conclusion states.
• A newsroom adopts a style rule requiring period, base and gross/net treatment, or an explicit 'not disclosed' flag, on every AI-lab run-rate figure, and the Sept 29 Axios story is rerun under it. (Supports) — The story would have flagged that no basis was given, so readers would have known the $70B was not comparable to net figures. This sharply reduces the chance that a basis change reads as a shortfall.
• An event study shows that AI-linked stocks moved on Oct 8-9 mainly because of the Bloomberg year-end projection and sector-wide credit news, not the run-rate basis. (Challenges) — The market-moving rationale in P5 would fail. The disclosure norm would still be justified on comparability and accuracy grounds, but with less urgency.
• A company publishes a complete method note that picks a favorable peak 28-day window annualized on a gross basis. (Neutral) — Disclosure is necessary but not sufficient. A stated method can still flatter, so the norm should be paired with comparison to trailing actuals or recognized revenue where available.
• The same gross-up pattern recurs when Anthropic's and OpenAI's figures are next compared, with different outlets giving different bases for the same company. (Supports) — Repetition would turn a single episode into evidence of a systematic comparability problem. That would strengthen the 'rarely disclosed' claim and the case for a convention such as reporting net figures.

Coherence & Relevance

The argument is coherent as a practical, inductive case for a disclosure norm. A vivid, documented episode is paired with a concrete accounting mechanism, and the assumptions rule out a revenue shortfall, a regulatory finding of manipulation, and the claim that run rates are useless. The weak points are the causal reading of the price moves, the leap from one episode to a general pattern, and the stricter-than-necessary phrasing of the norm, which does not account for sourcing limits or the middle path of attributed, flagged reporting. The title's tone also overshoots what the premises and assumptions support. A narrower conclusion, that reporters should state what is known about the basis and flag what is not, would follow well from the evidence.

• A run rate annualizes a shorter period of revenue, and the period used is often unstated. (Strong) — Establishes the definitional looseness. It needs evidence of how often the period is omitted to support 'rarely disclosed'.
• Axios reported on Sept 29, 2026 that OpenAI's ARR neared $70 billion without a stated calculation basis. (Strong) — Supplies the unstated-basis instance. The ARR versus run-rate wording needs reconciling, and it is unclear whether Axios could have obtained the basis.
• The FT reported on Oct 8 that OpenAI's own figure was about $50 billion, with the higher figure an investor gross-up. (Strong) — Directly shows how much basis matters. One case cannot establish typicality.
• Anthropic books marketplace sales gross while OpenAI books certain partner revenue net; both are permitted. (Strong) — Explains the mechanism and shows non-comparability. A2's 16% partner share means it explains part, not obviously all, of the ~40% gap.
• AI-linked stocks fell on Oct 8 and rebounded on Oct 9, showing that loosely defined run rates move prices in both directions. (Moderate) — Relevant to the 'move markets' clause only if causation is shown. The rebound followed a different type of claim, and confounders are not addressed.
• A reporter who cannot state the period, base and net/gross treatment is relaying a company's preferred number rather than a verified measure. (Moderate) — Links the facts to the reporting norm but is partly definitional. It lacks the normative bridge from 'unverified' to 'should not repeat', and it omits attributed and caveated reporting as an intermediate option.
• Accountability for boom-era claims deters the next boom, which requires recording what was said and how it was measured. (Weak) — Only loosely connected to the conclusion. It gives a rationale for archiving methods but supplies no mechanism linking that record to deterrence.

View this argument on LogicFirst.ai