ED ZITRON: Anthropic's S-1 showcases a quarter with discounted compute, and the economics it shows have already worsened

The Gist

Ed Zitron says Anthropic's best quarter was helped by cheap compute while new contracts were ramping. Since then, each compute dollar brings in less revenue and growth has slowed, and the filing itself warns that margins will shrink. So the S-1 describes a better business than the one that exists now. This is a steelman: Ed Zitron's arguments from the conversation are reconstructed in their strongest fair form, with his answers to the host built in. It does not endorse them.

Conclusion

The quarter the S-1 is built around shows better compute economics and faster growth than Anthropic has had since, and its own filing expects margins to tighten as committed capacity arrives, so public investors are being shown, in Ed Zitron's words, "a company that no longer exists."

Premises

1. Anthropic's showcase quarter, Q2 2026, ran with some compute at reduced fees: SpaceX capacity ramped in May and June at a reduced fee, and the large new Amazon and Google minimum spend contracts only began in April and May.
2. Investor briefings reported by the press put Q3 2026 at about $1.70 of revenue per compute dollar, down from about $2.08 in Q2, while quarterly revenue growth slowed from about 144 percent (Q1 to Q2) to about 47 percent (Q2 to Q3, $11.6 billion to $16.8 billion).
3. Anthropic's own filing says margins are expected to compress when large new compute arrives before the revenue it supports, that its revenue growth rate will decline, and that historical growth is not indicative of future results.
4. Committed compute keeps rising: SpaceX payments alone are set to reach about $3.1 billion a month once capacity is delivered in December 2026, and Anthropic is starting to run dedicated data centers of its own.
5. Part of Q2 demand may reflect a burst of heavy usage ("token maxing") and possible customer prepayments that need not repeat.

Assumptions

• Disclosure: Ed Zitron is CEO of EZPR, a media relations and public relations firm. He writes the Where's Your Ed At newsletter, which has a paid premium tier, and hosts the Better Offline podcast, produced with iHeartRadio and Cool Zone Media; both cover the AI industry critically, and in this conversation he says he predicted the rating outcome ("told you so"). Sources: "About Where's Your Ed At" (Where's Your Ed At) at https://www.wheresyoured.at/about/, "Introducing: Better Offline" (Where's Your Ed At) at https://www.wheresyoured.at/introducing-better-offline/ and "Ed Zitron" author page…
• The conversation was recorded on October 30, 2026, as the FT story on Anthropic's expected rating broke; figures are as stated live and may be approximate.
• "Revenue per compute dollar" means quarterly revenue divided by quarterly compute and infrastructure expense, not by long-term commitments.
• Differs: Neither filing quantifies the ramp-period discount, so how much it flattered Q2 compute cost cannot be measured from public documents.
• Differs: The same reports say Anthropic told investors it was profitable on an adjusted basis in Q3, its second straight such quarter, and cash-flow positive.

Analysis

Argument type: Inductive.

Potential Fallacies

• Non sequitur / rhetorical overreach (Conclusion) — The premises show a relative decline in two metrics: revenue per compute dollar from about $2.08 to about $1.70, and growth from 144% to 47%. They do not show a categorical change in what the company is. No bridging premise moves from "somewhat worse than Q2" to "no longer exists," especially when 47% sequential growth and adjusted profitability continue.
• Speculation treated as support (P1, P5) — P5 is hedged ("may," "possible") and unquantified, and A4 concedes the discount's effect cannot be measured. Both are used to portray Q2 as unrepresentative. A mechanism that could exist is not shown to be large enough to matter.
• Unsupported causal attribution (P1 with P2) — The ratio's decline is linked to the end of discounts, but other causes are not ruled out. These include ramp-up capacity arriving ahead of revenue, pricing or model-mix shifts, and training spend counted in the denominator. Growth deceleration also has a mechanical base-effect component.
• Treating boilerplate as admission (base-rate neglect) (P3) — Statements that growth will decline and history is not indicative are standard risk-factor language in nearly every high-growth S-1. The margin-compression language is more specific and more informative, but the set as a whole discriminates weakly between a deteriorating and a healthy company. It also shows the risk was disclosed, which cuts against a concealment narrative.
• Trend inferred from two data points / scope overreach (P2 and conclusion) — "Than Anthropic has had since" implies a series of later quarters, but only Q3 is reported, plus forecast commitments. One quarter-over-quarter change cannot separate durable decline from timing, since capacity can arrive before the revenue it supports.
• Selective emphasis (Conclusion versus A5) — Deterioration metrics are foregrounded, while Q3 adjusted profitability, positive cash flow, and the absolute $5.2B revenue added in one quarter are not weighed in the conclusion. This is a different issue from base-rate neglect, since these are specific, disclosed counter-facts.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — A 47% sequential increase, adjusted profitability for two straight quarters, and positive cash flow describe a business that has moderated from an extreme peak, not one that no longer exists. A filing that discloses the risk of margin compression and slowing growth is doing what a prospectus is supposed to do.
• Premise 3 (High impact) — Risk-factor language about declining growth and non-indicative history is near-universal legal boilerplate. Reading it as a confession overstates its information content, and its presence shows investors were warned.
• Premise 1 (High impact) — Because the ramp-period discount is unquantified (A4), the central causal story, that discounted compute flattered Q2, may rest on a small effect. Other drivers could explain most of the $2.08 to $1.70 move: ramp inefficiency, pricing and mix, and training spend in the denominator.
• Premise 2 (Medium impact) — A drop in percentage growth from 144% to 47% is expected as the base grows. Absolute dollars added ($5.2B in one quarter) are the more relevant measure.
• Premise 4 (Medium impact) — Committed compute is capacity bought against expected demand, and new capacity can generate the revenue it requires. Rising commitments show management confidence as much as risk, and the margin dip may be a temporary investment-phase lag.
• Conclusion (Medium impact) — An S-1 must report the latest completed, reviewed period, and later results typically reach investors through briefings and amendments. Q3 data reached the press via investor briefings, so the people who anchor an IPO book appear to already have it. The "misled public investors" framing assumes naivety the evidence does not show.
• Premise 5 (Low impact) — Without usage-cohort or deferred-revenue data, the token-maxing and prepayment story is conjecture. If Q3 revenue grew 45-47% off the supposedly inflated Q2 base, durable demand is at least as plausible.

Suggested Improvements

• Scope of the conclusion — Restate the claim as: Q2 was the high-water mark for revenue per compute dollar and growth, and the filing itself anticipates further margin tightening. Keep "company that no longer exists" as attributed commentary, not as the logical conclusion. The narrow claim is well supported. The existential framing is not entailed by the premises and invites easy dismissal.
• Quantifying the discount — Provide a sensitivity range. For example, show how large the SpaceX discount would have to be to explain the full drop from $2.08 to $1.70, or decompose the change into discount, ramp inefficiency, and pricing/mix. This would convert the central mechanism from directional assertion into a testable claim, addressing the A4 limitation.
• Integrating counter-evidence — Weigh Q3 adjusted profitability and positive cash flow directly. Explain what "adjusted" excludes and whether prepayments or payables timing helped cash flow, and ask for GAAP margins. This turns a reported counter-fact into a point the argument has already answered, strengthening credibility.
• Benchmarking — Compare Anthropic's growth deceleration, risk-factor language, and revenue-per-compute ratio with peer S-1s, hyperscaler build-outs, and Anthropic's own earlier quarters. A reference class would show whether the deceleration and warnings are unusual or the ordinary pattern for hypergrowth firms.
• Defining falsification criteria — State in advance what would refute the thesis, such as a Q4 revenue-per-compute ratio recovering above about $2.00 as new capacity fills. This guards against unfalsifiability and the "told you so" pattern, and gives the forward claim real testable content.
• Verifying sources and definitions — Reconcile briefing figures with the S-1's basis. Clarify whether the compute denominator includes training and R&D, and whether revenue is gross or net of cloud partner shares. Comparing briefed Q3 data to filed Q2 data across different accounting bases could create part of the apparent decline.

Scenario Tests

• Q4 2026 results show revenue per compute dollar falling further, toward or below $1.50, as SpaceX payments rise toward $3.1B a month. (Supports) — The thesis that Q2 was a high-water mark and committed capacity compresses margins would be substantially vindicated. The "no longer exists" framing would still overstate the case if the company remained profitable on an adjusted basis.
• Q4 revenue per compute dollar rebounds above $2.00 as new capacity fills with demand. (Challenges) — The Q3 dip would look like a timing lag from capacity arriving ahead of revenue, not a structural decline, and the argument would look premature.
• Disclosure shows the SpaceX ramp discount was small relative to total compute spend. (Challenges) — The title thesis about discounted compute would collapse, leaving only the generic claim of margin compression that the S-1 already states.
• Disclosure shows the discount was large enough to account for most of the drop from $2.08 to $1.70. (Supports) — Q2 would be shown to be materially flattered, and the representativeness concern about the showcase quarter would carry real weight.
• The S-1 is amended before pricing to include Q3 results. (Challenges) — The claim that public investors are being shown a stale picture weakens, because the gap between the filing and the briefings would close through ordinary disclosure.
• Revenue growth stays near 45-50% sequentially while GAAP margins move from slightly negative toward positive. (Challenges) — The economics would look like a healthy scale-up with a normal capacity-lag dip, and the bearish reading would look like motivated narrative.
• Q3 adjusted profit and cash flow are shown to rely on customer prepayments and adjustments that exclude major costs. (Supports) — Q3's apparent health would be partly cosmetic, which lends support to P5 and the claim that the economics are weaker than the headline numbers suggest.

Coherence & Relevance

The argument is coherent as a cumulative case for a modest claim: Q2 was the best quarter on two metrics, and the filing and the commitment schedule point to tightening margins. The premises fit together on cost-side pressure and deceleration. The weak links are the unquantified discount, the speculative P5, and the final inference to "a company that no longer exists," which is not supported and is in tension with the reported adjusted profitability and positive cash flow. The disclosed incentives of the source (A1) bear on framing and selectivity but not on the verifiable figures, so they are best treated as a reason to separate the numbers from the rhetoric.

• Q2 2026 ran with some compute at reduced fees (SpaceX ramp; Amazon/Google floors starting April and May). (Moderate) — It explains why Q2 might look better, but with no magnitude (A4) it cannot show that Q2 is unrepresentative. The Amazon/Google floors are commitments, not discounts, so their relevance to Q2 flattery is less direct.
• Q3 revenue per compute dollar of about $1.70 vs $2.08 in Q2; growth slowing from about 144% to about 47%. (Strong) — It directly supports "better than since" for Q3 only. It does not establish a later trend, and it mixes briefing-based figures with the filing's period. Base effects explain part of the deceleration.
• The S-1 expects margin compression, declining growth, and non-indicative historical growth. (Moderate) — It corroborates the direction of travel from the company's own words, but it is partly boilerplate and also shows the risk was disclosed, which cuts against a concealment narrative.
• Committed compute keeps rising (SpaceX at about $3.1B/month by December 2026; own data centers). (Strong) — It supports the forward-looking claim about cost pressure. It does not show revenue will fail to absorb the capacity, and it does not address utilization.
• Part of Q2 demand may reflect token maxing and possible prepayments that need not repeat. (Weak) — Speculative and unquantified. It adds a second, rival explanation for Q2 strength that is not separated from the discount story and is not supported by usage or deferred-revenue data.

View this argument on LogicFirst.ai