ED ZITRON AND WILL GUYATT: Adjusted profits and record chip orders do not show durable end demand for AI, because they exclude costs and reflect a few buyers racing each other

The Gist

Will Guyatt asks whether Anthropic's profits, huge chip orders and steady bank lending prove AI demand is real. Ed Zitron says no: the profit figure excludes real costs, chip orders come from a few giants racing each other, and much of the hardware is not even switched on yet. 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

The signs often cited as proof of real AI demand do not establish it: Anthropic's adjusted operating profit excludes material costs; record chip orders reflect spending by a handful of large buyers locked in a race none can quit first, while much of the hardware is not yet installed; and much of the end revenue that would justify that spending comes from OpenAI and Anthropic.

Premises

1. Anthropic told investors it expected a second straight quarter of positive adjusted operating income, a measure that excludes costs including stock-based compensation, and its reported gross margin above 80 percent is before partner revenue shares and model training costs.
2. TSMC reported record third-quarter 2026 revenue of about $46.7 billion, up about 50 percent, which shows strong orders for AI chips.
3. Those orders come from a small number of buyers, principally the largest cloud and AI companies.
4. Morgan Stanley research estimates that more than half of AI servers sold from 2026 to 2028 may lack a powered place to run, so purchases run ahead of working capacity.
5. Each large buyer faces a dilemma: stopping first risks a sharp stock decline and loss of position, so all keep spending even when returns are uncertain.
6. A large share of the end revenue that justifies cloud AI capacity comes from OpenAI and Anthropic.

Assumptions

• Reporting describes training costs as excluded from Anthropic's gross margin figure; whether they are also excluded from its adjusted operating income measure was not confirmed.
• Estimates of the two labs' share vary by measure; one estimate puts them at about 30 percent of new AI compute in 2026 and 40 to 50 percent of 2027 additions, and other large buyers exist, such as Google's agreement to pay SpaceX about $920 million a month for compute.
• The dilemma framing is an interpretation of incentives, not a documented statement by any company.
• Strong chip orders are consistent with both durable demand and overbuilding; the argument claims only that they do not prove durable demand.

Analysis

Argument type: Inductive.

Potential Fallacies

• Missing bridging premise (non sequitur) (Premise 1 to the first strand of the conclusion) — Showing that a metric excludes some costs establishes that it is partial. It does not show the excluded costs are large enough to flip the result, nor that profitability tells us much about end demand. The conclusion's word 'material' asserts a magnitude the premises never supply, and it is unconfirmed whether training costs are excluded from the adjusted operating income measure at all.
• Concentration treated as fragility (Premises 3 and 6) — Having few large buyers, or two lab customers, is normal in early infrastructure markets and does not by itself show demand is non-durable. The inference needs an added premise that the labs' revenue is itself subsidized, circular, or not backed by real end customers. Other large buyers also exist.
• Modal shift (forecast stated as fact) (Premise 4 and the conclusion) — Morgan Stanley says more than half of servers 'may' lack power over 2026 to 2028, which is a forecast. The conclusion says much of the hardware 'is not yet installed', which is a present-tense claim. Power-constrained deployment may also reflect a timing gap, or even demand outrunning supply, rather than weak demand.
• Unfalsifiable incentive narrative (Premise 5) — The buyers' dilemma explains continued spending whether returns are good or bad, so it carries little discriminating weight. It is acknowledged as interpretation, but it is still the step that turns 'record orders' into 'orders that don't signal demand'. Rational, well-founded competitive investment predicts the same behavior.
• Denominator ambiguity (Premise 6 and the second assumption) — Shares of new compute, of 2027 additions, and of end revenue are different quantities. Using them interchangeably makes 'a large share' and 'much of the end revenue' hard to assess.
• Selective evidence and asymmetric scrutiny (Argument as a whole, and the title framing) — Each bullish indicator is reinterpreted as non-probative while demand-side evidence (usage, revenue growth, enterprise contracts, compute scarcity) goes unexamined. The modest 'not proven' claim is defensible, but the framing can leave a stronger impression of weakness than the evidence supports.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Durable demand is evidenced by signals the argument never addresses: compounding lab revenue, token and API usage growth, enterprise adoption, sold-out compute, rising GPU rental prices, and multi-year compute commitments. Competitive racing can coexist with, and be justified by, real returns.
• Premise 4 (High impact) — Servers lacking power may signal that demand exceeds the physical capacity to serve it (a scarcity signal) rather than overbuilding. Grid, turbine and construction lead times are normal frictions, and delayed hardware may simply be deployed later.
• Premise 5 (High impact) — The dilemma is untestable. Executives investing on conviction in a winner-take-most market, with real balance-sheet discipline, would generate identical behavior, so the narrative adds little discrimination between durable demand and overbuilding.
• Premise 1 (Medium impact) — Excluding stock-based compensation is standard non-GAAP practice, and training costs are normally R&D rather than cost of goods sold. Unless the exclusions are shown to be large, or the GAAP trend is worse, the point is a quibble. Improving unit economics may be real.
• Premise 6 (Medium impact) — The labs are intermediaries whose revenue comes from enterprises and consumers. Other large buyers such as Google's SpaceX compute deal exist, and early-stage markets are normally concentrated, so concentration is a risk to monitor, not proof of circularity.
• Conclusion (Medium impact) — The same reasoning (concentrated buyers, racing, adjusted metrics) would discount demand evidence in railroads, fiber, and early cloud, so it needs a limiting criterion that separates overbuild from sound investment. Historical overbuilds also often found eventual use.

Suggested Improvements

• Quantify the excluded costs — Report GAAP versus adjusted operating income, size the stock-based compensation, revenue-share and training costs relative to revenue, and benchmark against peers' conventions. This would turn 'excludes material costs' from an assertion into a measured claim and settle the unconfirmed training-cost question.
• Supply the bridge on revenue quality — Show that the labs' revenue is funded by the same suppliers or investor capital, or depends on continued external fundraising, and quantify shares by revenue and not only by compute additions. Concentration implies fragility only with this missing premise, and the denominators in the estimates currently shift.
• Engage positive demand evidence — Address usage growth, enterprise retention, backlog composition, utilization, and GPU rental prices, and explain why they do or do not show durable demand. This answers the strongest opposing case and shows how the author's view differs from simple reflexive skepticism.
• Separate delay from overbuilding — Estimate what share of the Morgan Stanley shortfall is a timing lag versus stranded capacity, and track installed and energized capacity over time. The premise supports 'purchases outrun capacity' but not 'demand is weak' unless permanent idleness is shown.
• Make the claim falsifiable — Pre-specify indicators that would count as evidence of durable demand (for example, a first buyer cutting capex without a stock penalty, or utilization and pricing trends) and cite behavioral signatures of the dilemma. This prevents the thesis from absorbing any outcome and gives the modest 'not proven' claim predictive guidance.
• Align framing with scope — Keep the headline consistent with the 'does not prove' scope, and keep accounting concerns separate from the demand claim. This avoids a motte-and-bailey reading in which a modest claim carries the rhetoric of a bubble thesis.

Scenario Tests

• Anthropic and OpenAI revenue keeps compounding, GAAP margins improve, and enterprise usage broadens. (Challenges) — The adjusted-profit and circularity strands would look overly cautious. The core 'not proven' claim would still technically hold, but its practical relevance would shrink.
• Power and data-center constraints ease and the uninstalled hardware is deployed and utilized. (Challenges) — The Morgan Stanley shortfall becomes a timing delay and not evidence of overbuilding, removing the strongest evidence for the second strand.
• One large buyer cuts capex and is punished by markets, while others reassess. (Supports) — This would show the dilemma is real and that spending was partly sustained by competitive pressure, though not that end demand is absent.
• The labs' growth slows and the hyperscalers' backlog proves dependent on their commitments. (Supports) — The concentration and circularity concerns would be vindicated, and capacity justified by the two labs' revenue would look overextended.
• Training costs turn out to be included in Anthropic's adjusted operating income, or the exclusions are small. (Challenges) — The first strand would collapse while the chip-order and concentration strands remain, showing the strands are independent.
• Strong orders and weak orders both prove compatible with durable demand over a multi-year period. (Neutral) — This is the underdetermination the fourth assumption concedes. Orders alone cannot separate the hypotheses, so distinguishing evidence has to come from utilization, pricing, and end-user revenue.

Coherence & Relevance

The argument is a coherent conjunction of three largely independent strands (incomplete profit measures, concentrated and race-driven orders ahead of installable capacity, and lab-dependent end revenue), each aimed at the modest thesis that the signals are insufficient. At that scope it holds up as a cumulative case. The coherence weakens where the premises are read as implying overbuilding or fragile demand, because several links (the materiality of excluded costs, the quality of lab revenue, and the meaning of unpowered servers) are asserted rather than shown. The assumptions candidly disclose the main gaps, which improves honesty, though the title leans harder than the premises can carry.

• Anthropic's adjusted operating income excludes costs including stock-based compensation, and its 80%+ gross margin is before revenue shares and training costs. (Moderate) — This bears on profitability, not directly on end demand. It needs a premise that the excluded costs are material, and the training-cost exclusion is confirmed only for gross margin.
• TSMC reported record Q3 2026 revenue of about $46.7B, up about 50%. (Moderate) — It is the datum being explained and not support. It is relevant as the target of the 'does not prove' claim, and the argument gives no direct positive evidence about demand.
• Orders come from a small number of buyers. (Moderate) — It shows correlated risk, not weak demand. It is unquantified and relevant mainly in combination with the dilemma premise.
• Morgan Stanley estimates more than half of AI servers sold 2026-2028 may lack a powered place to run. (Strong) — It directly supports that purchases run ahead of working capacity. The leap from unpowered hardware to weak end demand needs the added claim that the idle capital will not earn revenue.
• Each large buyer faces a dilemma in which stopping first is punished. (Strong) — It supplies the mechanism by which orders could stay strong without strong end demand, but it is interpretive and untested, and it cannot discriminate against rational conviction-driven investment.
• A large share of end revenue justifying cloud AI capacity comes from OpenAI and Anthropic. (Moderate) — It needs the unstated premise that the labs' revenue is subsidized, circular, or fragile. The metric is compute share, not revenue share, and other buyers exist.

View this argument on LogicFirst.ai