The FTC chair says his "alarm bells go off" at industry safety coordination, a challenger CEO calls it incumbents dictating terms, and the chips rallied anyway.
The debate over whether frontier AI development should slow down widened this week from a technical and ethical argument into a competition-policy one, and the new participants have changed its shape.
Anthropic chief executive Dario Amodei and OpenAI chief executive Sam Altman have reportedly converged on the position that AI development needs to slow, while acknowledging concrete obstacles to any actual slowdown: competitive dynamics with China, and the profit incentives created by capital raises on an IPO track, including Anthropic's own.
That convergence is what drew in the regulators and the competitors.
The antitrust objection
Federal Trade Commission Chair Andrew Ferguson has been publicly sceptical of granting any antitrust waiver to permit industry-wide safety coordination, saying "all of my alarm bells go off."
The objection is not frivolous. A small number of firms with dominant positions agreeing among themselves to limit the pace of product releases, in a market with high barriers to entry, describes the structure of a cartel regardless of the motivation. Antitrust law assesses effects, not intentions, and the effect of coordinated pacing among incumbents is to freeze the competitive position of the firms doing the coordinating.
The competitor's objection
Cohere chief executive Aidan Gomez has made the commercial version of the same point, framing the safety-pacing push as incumbents attempting to "dictate the terms for everyone else."
This is the argument that gives the story genuine two-sidedness. A challenger firm needs to ship aggressively to close a capability gap. A pacing norm set by the leaders locks in the existing ordering, whatever its stated purpose.
The political and technical responses
Speaking at the All-In Summit in Los Angeles, President Trump called AI safety fears "a hoax" and separately described data centres as "the oil of the next 20 to 25 years."
At the same event, Elon Musk proposed that major AI labs test each other's models using shared security harnesses rather than relying on self-evaluation, arguing it would raise the legal and accountability cost of unsafe releases. He reaffirmed that "Dario is right" and noted any framework would need to be acceptable to China. Peer evaluation is a partial answer to the antitrust objection, since mutual testing is a verification mechanism rather than an agreement to restrict output.
Separately, a cluster of safety researchers has moved to METR, an independent frontier-AI evaluator. Jacob Coxon and Joe Benton departed Anthropic and Josh Engels departed Google DeepMind in early September, with a July 2026 sandbox-escape incident reported as a shared reference point. DeepMind's Bilal Chughtai resigned earlier, stating that AI "has the potential to kill us all."
The market did not agree
For all the escalation, the semiconductor complex rose. Nvidia closed Tuesday at $212.17, up 0.57%, and AMD closed at $504.20, up 2.19%, on a session when the S&P 500 fell 0.45%. The Philadelphia Semiconductor Index gained 0.40% against a broadly lower tape.
One reading is that investors concluded a safety debate conducted at conferences does not slow infrastructure capital expenditure. That interpretation has been offered by market commentary rather than established.
It is also worth separating the complex rather than describing a uniform chip trade. Memory suppliers remain supply-constrained on hyperscaler demand in a way logic and accelerator names are not, and Micron's results on September 30 are the next scheduled test of the memory shortage thesis.
Anthropic's own IPO process remains, per the company's June 2026 announcement, a confidentially submitted draft registration statement with no share count, price, venue or timeline confirmed by the company.