ED ZITRON: The best counterargument, that a smaller AI industry survives, does not fit OpenAI and Anthropic, whose commitments and valuations assume a far larger one

The Gist

Ed Zitron says the best pushback he hears is that AI will shrink but survive. He argues that does not work for OpenAI and Anthropic, because they have signed enormous compute contracts and are being sold to investors on giant market claims. Promises that new chips will make AI much cheaper have not yet shown up as real savings for these companies. 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 strongest counterargument to the AI bubble thesis, that a smaller AI industry survives the boom with OpenAI and Anthropic trimming their costs, does not fit those two firms, because they carry compute commitments estimated at about $1.3 trillion, are being valued on claims of markets measured in tens of trillions of dollars, and run inference businesses that must buy capacity ahead of uncertain demand; the usual rescue, that new chips will make serving customers far cheaper, has not yet been shown as realized savings for these firms.

Premises

1. The most plausible counterargument to the bubble thesis is that a smaller AI industry survives after the boom, with the leading labs cutting training costs and continuing at a reduced scale.
2. OpenAI and Anthropic have taken on compute commitments estimated at about $1.3 trillion combined, much of it payable whether or not the capacity is used, so they cannot shrink to fit a smaller market without breaching or renegotiating those contracts.
3. Both firms are being prepared for public listings on the premise of very large future markets, and Anthropic has pitched IPO investors on a total addressable market above $30 trillion.
4. Running inference resembles running an airline: capacity is rented ahead of demand, so too little demand loses money on idle capacity and too much degrades service, and today the shortfall is covered by investors during a period of intense enthusiasm rather than by customer cash.
5. Claims that new chips will cut costs are typically stated as throughput per megawatt on selected workloads, and the labs have not shown how much cheaper serving their own customers will become in realized dollars.
6. An argument that depends on an unspecified future jump in value, prices, or cost efficiency is a hope rather than an economic case.

Assumptions

• The $1.3 trillion combined figure is one estimate; reported totals vary by horizon, with OpenAI's planned compute spending through 2030 at about $750 billion and Anthropic's commitments at least $518 billion, of which about $413 billion is non-cancelable.
• Anthropic's $30 trillion figure is a marketing estimate of addressable work, and it is somewhat below, not above, annual U.S. nominal GDP of about $32.6 trillion.
• Nvidia states that its Vera Rubin systems produce up to ten times the tokens per megawatt of the prior generation on a DeepSeek R1 workload, which it says implies about one tenth the cost per million tokens, so some cost claims are stated in money terms, though not as the labs' realized serving costs.
• The argument concerns OpenAI and Anthropic specifically; it does not deny that some smaller AI industry could exist after a downturn.
• Announced U.S. data center pipelines exceed 300 gigawatts, while forecasts of what will actually be installed are far lower, so planned capacity figures describe ambition rather than likely construction.

Analysis

Argument type: Inductive.

Potential Fallacies

• Unstated bridging premise / false dichotomy (Premise 2 and conclusion) — The argument moves from 'the firms cannot shrink without breaching or renegotiating' to 'a smaller survivor does not fit'. That treats renegotiation, restructuring, debt-for-equity swaps or acquisition as failure states. Those are the ordinary mechanisms by which a smaller surviving business emerges, so the step needs an added premise explaining why they are fatal.
• Scope shift between industry and firms (Premise 1, title and conclusion) — The counterargument is about a smaller industry surviving, but the test applied is whether two specific firms can shrink in their present form. Assumption 4 concedes the narrower scope, yet the title and framing suggest a broader refutation. Other firms or successor owners of the labs' assets could constitute the smaller industry.
• Argument from absence of evidence (Premise 5 and the conclusion's final clause) — That the labs have not shown realized serving-cost savings is treated as grounds to doubt the savings. Private firms disclose little, and hardware not yet deployed cannot show realized savings. The conclusion is carefully worded ('not yet been shown'), which limits the overreach, but 'unproven' still should not be read as 'absent'.
• Weak analogy (Premise 4) — The airline comparison captures fixed-capacity risk but carries over a reputation for poor returns. GPU capacity is more fungible and resalable than seats, per-token prices are falling, and demand can be shaped by pricing and tiering. Disanalogies are not addressed, so the analogy illustrates more than it proves.
• Marketing figure treated as evidence of operating assumptions (Premise 3) — A TAM slide is a conventionally inflated upper bound, and investors discount it. Assumption 2 concedes it is a marketing estimate below U.S. GDP. Inferring that valuations or cost structures depend on that scale needs a bridge such as implied revenue paths, which is not supplied.
• Asymmetric burden / overbroad standard (Premise 6) — Defining any reliance on an unspecified future improvement as 'hope' would disqualify ordinary forecasting and historically grounded cost curves. It is also applied one way: the bearish scenario depends on equally unquantified forecasts of demand shortfall and failed renegotiation. A probabilistic standard (how likely, how large) would be fairer.
• Headline-number anchoring (Premise 2 and conclusion) — The $1.3 trillion figure is the largest aggregate, mixing horizons and planned versus non-cancelable amounts, and it pools two separate firms. Without a denominator such as revenue or utilization, its size cannot be interpreted. The stated assumptions calibrate this, but the conclusion keeps the full force of the headline.

Counterarguments

These challenges are not ranked.

• Premise 2 (High impact) — Counterparties such as cloud providers, chipmakers and data center developers are heavily exposed to these two customers and have strong incentives to renegotiate, defer, take equity or resell capacity rather than force default. Commitments are also staged and conditional on build-out milestones that may never be met. Renegotiation or Chapter 11-style restructuring can leave a smaller operating business intact, which is the counterargument itself.
• Conclusion (High impact) — A smaller AI industry can survive even if these two firms fail or are restructured or acquired. As with fiber and telecom, the infrastructure and technology persist under successor owners at lower prices. Showing the labs don't fit a smaller industry therefore does not refute the industry-level claim.
• Premise 5 (Medium impact) — Per-token prices at constant capability have fallen steeply, which is indirect evidence of real cost declines. Non-disclosure by private firms is expected whether or not savings exist, and chips not yet deployed cannot have yielded realized savings. Jevons-type demand growth cuts both ways for the thesis.
• Premise 4 (Medium impact) — GPU capacity can be reallocated among inference, training and research, or resold. Rapid revenue growth supplies customer cash, so investor subsidy may be a normal growth-stage pattern rather than a sign of fragility. Airlines also survive at scale despite poor returns, which supports the smaller-survivor view.
• Premise 3 (Medium impact) — TAM slides are boilerplate upper bounds that investors discount heavily. Valuations rest on probability-weighted revenue paths for a plausible slice of the market, so the pitch says little about what the firms need in order to survive.
• Premise 6 (Medium impact) — Applied consistently, the standard would label early Amazon, cloud, solar or semiconductor-curve investing as 'hope'. It proves too much and applies equally to the bearish forecast.

Suggested Improvements

• Close the logical gap on restructuring — State and defend an explicit premise about what happens to equity, creditors and operations under renegotiation or restructuring, or reframe the claim as 'a smaller-survivor outcome would require restructuring and would impose large losses, so it is not the benign scenario the counterargument describes'. The argument concedes renegotiation as an option without explaining why it defeats the counterargument. Reframing would fit the evidence.
• Provide a denominator for the commitments — Compare annual commitments, split into cancelable and non-cancelable parts by firm, with revenue, gross margin and cash burn for each, and show the revenue needed to service them. A large number is only interpretable against revenue and plausible demand, and separating OpenAI from Anthropic reflects their different exposures.
• Derive valuation-implied expectations — Reverse-engineer the revenue and margin paths implied by recent valuations and compare them to the size of a 'smaller industry'. This would replace the weak TAM inference with a direct test of what the valuations assume.
• Engage the strongest version of the counterargument — Address renegotiation, resale of capacity, counterparty incentives, growth into commitments and observed per-token cost declines. The 'trimming training costs' version is thin, and addressing the stronger forms would make the rebuttal more convincing.
• Replace the 'hope' standard with a probabilistic one — Evaluate the likelihood and size of cost improvements using historical cost curves and realized price trends, and apply the same standard to the bearish forecast. It avoids an asymmetric burden and makes the claim testable.
• Make the claim falsifiable — State time-bound, observable indicators such as renegotiations, deferrals, failed or repriced IPOs, or distressed restructurings, and realized cost per token served. Bubble calls often have the right direction but the wrong timing, so specific indicators would clarify what would count as confirmation or refutation.

Scenario Tests

• Counterparties renegotiate, defer or take equity in exchange for smaller commitments, and the labs continue at reduced scale. (Challenges) — This is the smaller-survivor scenario achieved through ordinary contract mechanics. It weakens the claim that the firms cannot shrink, though equity holders and suppliers would still bear large losses.
• Demand from coding, agents and enterprise grows fast enough to use the contracted capacity at acceptable prices. (Challenges) — Neither the bubble nor the smaller-industry outcome occurs, and the larger-industry premise embedded in the valuations is borne out.
• Demand grows only moderately, falling short of the commitments, while investor funding tightens. (Supports) — Fixed non-cancelable obligations then bind. Both firms face distress, restructuring or absorption by larger partners, so graceful trimming does not occur.
• The labs publish realized cost per token showing large gains from new chip generations and rising gross margins. (Challenges) — The chip-savings rescue would be demonstrated rather than hoped for, removing one leg of the argument, though commitment rigidity would remain.
• One or both labs are restructured or acquired by a hyperscaler while models and usage persist at a smaller overall scale. (Neutral) — This supports the firm-level claim that they cannot shrink in their present form and the industry-level counterargument that a smaller industry survives. It shows why the scope distinction matters.
• Planned capacity is far above what is actually built (the 300+ GW pipeline versus lower installation forecasts). (Neutral) — It supports skepticism about headline figures but also suggests part of the $1.3T commitment is aspirational and may fall away without breach, softening Premise 2.

Coherence & Relevance

The argument is a cumulative case in which fixed commitments carry most of the weight and the other premises corroborate. It is coherent as a claim that these two firms are poorly positioned for a graceful downsizing and are sold on far larger expectations. It is less coherent as a refutation of the industry-level counterargument, because renegotiation, restructuring and successor ownership are compatible with a smaller surviving industry. The stated assumptions improve calibration and scope. The categorical 'does not fit' should be read as 'would require costly restructuring' rather than 'cannot happen'.

• The most plausible counterargument is a smaller surviving industry with labs trimming costs. (Strong) — It sets up the target but defines the counterargument narrowly, omitting restructuring and growth-into-commitments variants, and shifts between industry-level and firm-level survival.
• ~$1.3T in compute commitments, much payable regardless of use. (Strong) — Directly supports the claim that trimming is hard. It is missing contract terms, a revenue denominator, and an explanation of why renegotiation equals failure.
• IPO preparation and a $30T+ TAM pitch. (Moderate) — Shows the firms' public narrative is far larger than a smaller-industry story, but not that valuations or operating plans depend on that scale. The link from pitch to valuation requires implied revenue analysis.
• Inference resembles an airline, with investors covering the shortfall. (Moderate) — Relevant to demand risk on purchased capacity, but the analogy is unquantified and there is no margin or revenue data showing who is funding the shortfall.
• Chip cost claims are throughput-based and the labs have not shown realized savings. (Moderate) — Addresses the main rescue argument but rests on non-disclosure, and ignores observed price declines and the money-term vendor claims noted in the assumptions.
• Reliance on an unspecified future jump is hope, not an economic case. (Weak) — Works as a framing norm rather than evidence, and requires a limiting principle (specificity or probability) to avoid dismissing all forecasting.

View this argument on LogicFirst.ai