GAVIN BAKER: A dominant OpenAI and Anthropic duopoly would harm nearly every other company, so model-layer fragmentation is better for AI infrastructure and applications

The Gist

Gavin Baker says if OpenAI and Anthropic ran the whole model market, they would squeeze everyone who sells to them or builds on them. More competition is better for chips, clouds and apps, and the big two can still grow. This is a steelman: Gavin Baker's thread is reconstructed in its strongest fair form, not endorsed.

Conclusion

Fragmentation and competition at the model layer are better for AI infrastructure suppliers and application companies than a two-lab oligopoly, and the leading labs can still grow within a competitive market.

Premises

1. If two labs dominated the model layer, they would be the main buyers of cloud capacity and data center chips, giving them monopsony power over suppliers.
2. Dominant buyers with that power tend to press supplier prices and eventually integrate vertically.
3. A duopoly at the model layer would also squeeze the application layer that depends on it.
4. The model layer has become markedly more competitive since mid-2026, with new frontier entrants and open-weight models.
5. The combined business of OpenAI and Anthropic kept growing quickly over the last three months even as share shifted, so competition need not shrink the leaders.

Assumptions

• The claim compares market structures; it does not predict that either leading lab fails.
• Baker calls this view quasi-consensus in Silicon Valley; that describes opinion, not evidence.

Analysis

Argument type: Inductive.

Potential Fallacies

• Non sequitur (missing bridging premise) (Inference from P1–P3 to the conclusion) — P1–P3 describe harms of a duopoly, but the conclusion says fragmentation is better. That requires an unstated premise: fragmentation brings no comparable or larger harms to suppliers and applications (demand volatility, weaker customers, higher integration costs), and a duopoly's anchor-buyer benefits are smaller than its harms. Without this, only half of the comparison is made.
• One-sided evaluation / selective consideration (Conclusion) — Only the downsides of one market structure are weighed. Long-term commitments from large creditworthy buyers, which can help finance chip and data-center build-outs, are not considered, and neither are the capital and counterparty risks of a more fragmented field.
• Hasty generalization / argument from analogy (P2) — The tendency of dominant buyers to squeeze suppliers and integrate vertically comes from industries like retail and autos. It is applied to AI compute without showing that conditions are similar, since suppliers here are concentrated, supply is scarce, and vertical integration is already happening under competition.
• Short-window extrapolation (P4 and P5) — Three months of growth and a recent competitive shift are used to support general claims about market structure and leader economics. In a fast-expanding market, rising demand can mask share loss and says little about margins or longer-run outcomes.
• Counterfactual treated as settled (P1 and P3) — The behavior of a hypothetical duopoly is stated with more confidence than an unobserved scenario supports, and it is hard to refute with current evidence. Offsetting forces such as supplier counter-power, hyperscaler rivalry, open-weight alternatives and regulation are not weighed.
• Possible false dichotomy (Title and conclusion) — Contrasting a two-lab duopoly with broad fragmentation omits intermediate structures, such as three to five strong labs, that could change the effects on suppliers and applications. The harmful endpoint is also treated as a stable outcome rather than one possibility among several.
• Appeal to consensus / authority (mitigated) (A2 and sourcing) — Invoking Silicon Valley 'quasi-consensus' and the speaker's standing lends social proof. The author concedes this is opinion, which limits the problem, but agreement among parties with aligned interests is correlated rather than independent evidence.

Counterarguments

These challenges are not ranked.

• Premise 1 (High impact) — Suppliers hold the leverage. GPUs, advanced packaging, memory and power are scarce and concentrated, so labs compete for allocation and may pay more, not less. Circular financing, in which suppliers invest in labs, further blurs the picture. The monopsony mechanism may never operate.
• Conclusion (High impact) — Concentrated, creditworthy anchor buyers sign multi-year, multi-gigawatt commitments that de-risk fab and data-center build-outs. Fragmented demand from weaker labs raises counterparty risk, stranded capacity and bankruptcies, so fragmentation could be worse for suppliers.
• Conclusion (Medium impact) — The usual platform analogies (Intel/Microsoft, Apple/Google, the big three clouds) saw complement and supplier ecosystems thrive under concentrated, growing platforms. Concentration at the model layer need not harm the rest of the stack.
• Premise 3 (Medium impact) — Two strong labs competing with each other and with open-weight models would likely push token prices down, which benefits applications. Application squeeze by labs moving up-stack happens regardless of model-layer structure.
• Premise 4 (Medium impact) — Frontier training costs may favor scale and reconsolidation. Recent entrants and open-weight models may not be durable or frontier-competitive, and the competitive moment could be transient, in which case the duopoly warning is not moot.
• Premise 5 (Medium impact) — Absolute revenue growth in an expanding market is compatible with margin compression and weaker pricing power. Growth in a boom says little about leader economics in a slowdown or price war.
• Conclusion (Medium impact) — 'Better' is defined only from the standpoint of suppliers and apps. Consumer welfare, safety coordination, and frontier-research funding may point the other way, and a stakeholder-benefit standard could justify protecting incumbents rather than competition.

Suggested Improvements

• Quantitative evidence — Supply data on model-layer concentration (revenue and token share, HHI), benchmark gaps between leaders and entrants or open-weight models, and lab revenue and margins over 12 months or more, with sources. P4 and P5 are currently bare assertions over a short window. Data would show whether competition is durable and whether leaders' economics are intact.
• Supplier-side structure — Analyze buyer concentration against supplier concentration, differentiating among chip vendors, hyperscalers, neoclouds, memory and power suppliers. Specify which suppliers would be harmed. This addresses the biggest weakness in P1–P2 and replaces the uniform 'nearly every other company' claim with a tiered one.
• Symmetric comparison — Explicitly weigh the costs of fragmentation (demand volatility, counterparty risk, duplicated capex, reduced frontier funding, integration costs) against the costs of a duopoly, and explain why fragmentation nets out better. The comparative conclusion is not earned until both sides of the comparison are evaluated.
• Mechanism and analogues for P3 — Show concrete evidence of application-layer squeeze (API price changes, lab entry into app categories, app gross margins) and compare with historical platform cases. P3 is currently asserted rather than demonstrated and is the least supported premise.
• Scope and definitions — Define 'better' (revenue, margins, survival, growth), specify the time horizon, and address intermediate structures with three to five strong labs. This removes ambiguity and the false-dichotomy risk, and makes the claim testable.
• Probability of the duopoly — State how likely a duopoly is and what conditions (scale economics, capital access, distribution, data flywheels) would produce or prevent it. If fragmentation is already occurring, the argument needs to explain why the duopoly counterfactual matters.

Scenario Tests

• Compute supply stays the binding constraint, and chip and memory suppliers keep strong pricing power against all buyers. (Challenges) — The monopsony mechanism in P1–P2 fails to operate, removing the main supplier-harm channel regardless of how many labs exist.
• Frontier training costs grow so large that only two or three labs can fund them, and open-weight and new entrants fall behind the frontier. (Challenges) — Fragmentation proves unsustainable, so the practical question becomes how suppliers and applications cope with concentration. This undermines P4 and the argument's relevance.
• Model competition intensifies, and API prices fall while application margins rise and multi-model routing becomes standard. (Supports) — This aligns with the claim that a competitive model layer benefits applications, and it shows leaders can grow in a competitive market, though supplier effects remain uncertain.
• A fragmented field of weaker labs signs large compute commitments, then several fail or consolidate during a funding downturn. (Challenges) — Infrastructure suppliers suffer stranded capacity and defaults, showing fragmentation can be worse for suppliers than concentrated anchor demand.
• A two-lab duopoly emerges and the labs launch first-party products that undercut their API customers while raising prices. (Supports) — This would corroborate P3 and the platform-squeeze concern. It would not by itself settle the supplier side.
• Demand growth slows and price competition compresses leaders' margins, even though revenue was growing during the earlier boom. (Challenges) — P5's three-month evidence would not generalize, and competition could shrink leaders' economics, weakening the second half of the conclusion.

Coherence & Relevance

The argument is internally coherent as an inductive case. One strand (P1→P2→P3) argues that a duopoly would harm suppliers and applications. A second (P5) pre-empts the objection that competition hurts the leaders. P4 is largely background context. The main gaps are an implicit bridging premise that fragmentation has no offsetting harms, generic quantification in P2–P3 that is stretched into a near-universal conclusion, and an unexamined assumption that suppliers are the weaker party. The conclusion is conjunctive and so only as strong as its weakest half. Given the stated assumptions (a structural comparison, and consensus offered as opinion only), the argument is a plausible, directionally reasonable case that is not yet proven.

• If two labs dominated the model layer, they would be the main buyers of cloud capacity and data center chips, giving them monopsony power over suppliers. (Strong) — Opens the supplier-harm strand, but the step from buyer concentration to bargaining power ignores supplier concentration and scarcity, as well as other large buyers.
• Dominant buyers with that power tend to press supplier prices and eventually integrate vertically. (Strong) — Links buyer power to harm, but vertical integration occurs under any structure, and whether integration harms or helps suppliers is unaddressed.
• A duopoly at the model layer would also squeeze the application layer that depends on it. (Strong) — Directly supports the application-company half of the conclusion, but it is stated categorically and has no mechanism or evidence. Offsetting factors such as price competition and multi-sourcing are not addressed.
• The model layer has become markedly more competitive since mid-2026, with new frontier entrants and open-weight models. (Weak) — Provides context that fragmentation is occurring, but does not show that fragmentation benefits suppliers or applications. It also makes the duopoly scenario more hypothetical.
• The combined business of OpenAI and Anthropic kept growing quickly over the last three months even as share shifted, so competition need not shrink the leaders. (Moderate) — Relevant only to the second clause of the conclusion, as a rebuttal to the objection that competition hurts the leaders. It is silent on margins, durability, and the supplier and application comparison.

View this argument on LogicFirst.ai