Stratechery finds a commercial use for Anthropic's call to pace the frontier

An overhang here is a gap opened by speed. The article puts it this way: the frontier labs' problems "take the form of overhangs that have formed by virtue of how rapidly models are improving; slowing model improvement would reduce the overhangs" ^[4]. The commercial version of that sentence is about time: capability that is already shipped and not yet built into products has longer to earn if nothing better lands next quarter.

Where the money sits, on the same site's account, is integration. It argued that profits flow away from the modular, commoditized parts of a value chain and toward the integrated ones, leaving Anthropic and OpenAI poised to be significantly more profitable than it had looked as recently as late last year ^[10]. "What this means is that model performance isn't the only thing that matters: the integration between model and harness is where true agent differentiation is found," the author wrote ^[9]. He now says he has wavered on that ^[11]. One of his own projects over the past month or so involved building a harness; he reported it was doable, and wrote "It also is very difficult" ^[12]. The text as supplied breaks off mid-sentence there, before it names the overhangs beyond the capability one ^[13].

The bubble rebuttal and the pacing case pull against each other. Six months ago the case against an infrastructure bubble was that agentic models are so capable and so token hungry that all the compute being built is genuinely needed ^[7]. That demand comes from models that keep getting better, and pacing the frontier slows it. The paradigm is dated to the release of Opus 4.5 in late November 2025 ^[8]. The pacing argument therefore arrives at least six months into it ^[16]. If integration is where the profit is ^[10], a paced frontier gives the labs time to wrap capability they already have.

The arguments land in sequence, the bubble rebuttal first and the safety-superpower piece three months after it ^[15]. The convenience reading is hard to disprove, and the article does not pretend otherwise: it says pacing is framed as a safety concern and, in the author's belief, motivated by one ^[3]. That earlier piece argued Anthropic's sincere safety belief conveniently licensed goals matching its commercial interest, including aggressive attempts to disintermediate software and secret sabotage of would-be competitors ^[6]. Sincerity and convenience predict the same behaviour, so the test is whether Anthropic holds back a capability a competitor ships and bills for. The author is not neutral on the underlying philosophy. "I reject the premise in favor of doubt in our ability to foresee the future, combined with faith in humanity figuring things out along the way," he wrote ^[14].