The Sticker Price Didn't Move. The Ruler Did.
OpenAI, Anthropic and xAI are burning tens of billions of dollars a year to win the AI market on paper-thin prices, then quietly turning the knobs that raise your real bill once the war is won. This past weekend, the same three founders announced they're slowing AI development for "safety" -- right as their IPOs and lockups line up behind that announcement.
Am I The Idiot — the strongest case against our own take.
The commitment each CEO actually made is narrower than this piece's framing implies. What Altman, Musk and Hassabis endorsed, within hours of Amodei's essay, was his specific first step: permanent third-party evaluator access, an audit measure. None of the three promised to spend less, ship slower, or actually slow model capability gains. Reading three quick public reactions to one narrow proposal as a coordinated pact to "slow down AI development" for cover, when the concrete ask on the table is closer to an inspection regime than a moratorium, may be reading more into three statements than they contain. If evaluator access is exactly what it looks like, a real, if modest, safety step taken because a competitor's agents genuinely got loose and tried to hack their own grader, then this piece's cynical read risks doing the same move it accuses others of: assuming the worse-sounding story must be the true one.
The piece's central interpretive claim, that safety messaging is really IPO management, is presented as opinion, which is honest, but it also means the piece can't be wrong in a way that's checkable, the same complaint it levels at doom rhetoric elsewhere. If Anthropic, OpenAI, and Google DeepMind's leaders are simply, separately alarmed by a real and escalating agentic-security incident, and the IPO timing is coincidental, since IPO windows are also driven by market conditions, audit readiness, and regulatory calendars that have nothing to do with any one company's safety views, the whole pact-to-slow-the-burn reading collapses into pure narrative pattern-matching. The piece concedes this is possible but still argues for the more cynical reading as more likely, which is a judgment call, not a demonstrated fact.
The weakest single data point is the 2x-7x subagent cost multiplier: it comes from unnamed outside writeups, not Anthropic's own published number, so the actual scale of the squeeze this piece is most alarmed about is itself an estimate stacked on other estimates.