Amodei concedes that a coordinated AI slowdown may need government antitrust cover

The second stage is where the plan stops being something a company can do by itself. Amodei asked frontier developers in democratic countries to coordinate on common safety requirements and on limits for unchecked capability growth ^[10], and in the same essay he acknowledged that parts of that coordination could raise antitrust issues and might require narrow government authorization ^[11]. Competitors agreeing to hold capability back has the structure of the agreements competition law exists to police.

Stage one is Anthropic's alone to grant ^[9]. That leaves two of the three stages with parties outside the frontier labs ^[18]. Stage three needs other governments, and Amodei's essay contemplates agreements with China at several possible levels, running from restrictions on AI-assisted biological weapons to an eventual limit on rapid, automated model improvement ^[12].

David Sacks put his objection at the middle stage, arguing that laboratories could slow voluntarily without requiring a coordinated regulatory framework from competitors ^[6]. Voluntary pacing has no verification step. Verification is the reason Amodei himself ranked a comprehensive international pause as the least likely form of agreement to reach ^[13].

Yakovenko answered the whole debate in four words on Sept. 13: "Profitability at $1 trillion mcap." ^{[7] [8]} He did not say which company the figure referred to or show the revenue, valuation and profitability arithmetic behind it ^[8]. Anthropic and OpenAI are both privately held, so a market-capitalisation figure attached to either can only describe estimates from private transactions or a proposed listing ^[15]. Reuters reported that both were preparing for potential initial public offerings, without establishing that either had reached a $1 trillion market capitalisation or had made profitability contingent on that valuation ^[14], and no named company has confirmed that its support for slower development depends on reaching that number ^[17].

The incentive theory is the weaker of the two accounts. Amodei's stated reasons are model-control failures, cyber risks and the speed at which AI systems are helping create newer systems ^[16]. Yakovenko framed his post as a sarcastic observation, and he offered no evidence that Anthropic, OpenAI or xAI had coordinated their statements for financial reasons ^[19]. The antitrust caveat is the part of this that costs something. If the coordination ask turns into a legislative request with a defined exemption, pacing becomes a compliance cost that every frontier developer pays and that shows up in release schedules. If it stays in essays, the only thing that ships is the evaluator access Anthropic can grant on its own authority. Altman promised further implementation details soon ^[22].