Ask who should slow AI down and you have already asked the wrong question. The useful one is who sets the pace, and the three men sharing that Dreamforce stage each answered from their own balance sheet. Huang wins when everything runs faster, because Nvidia sells compute to whoever is running. Altman and Amodei win when the next training run is expensive enough that few can afford it, because OpenAI and Anthropic sell access. They are describing their margins. Amodei's proposal to publish safety thresholds before a run instead of after is real engineering discipline and also a schedule. A schedule only a handful of labs can meet is a barrier to everyone who cannot.
Watch what happens when safety becomes a shared standard. If frontier labs jointly publish dangerous-capability thresholds and agree not to cross them, that is standard-setting with market effects, and Section 1 of the Sherman Act does not care what the restraint is called. The Supreme Court made this plain in National Society of Professional Engineers v. US, striking an ethics code that happened to limit price competition. A safety floor and a price floor can be the same floor. The FTC's January 2025 6(b) report already mapped the Microsoft-OpenAI and Google-Anthropic ties. Enforcers rarely ask whether a threshold was sincere. They ask who set it and who could not afford to clear it.
DeepMind's new AGI institute is the sharpest move in this argument, and not because it promises rigor. Evals, red teams and published thresholds beat CEO intuition, yet they also become the exam every other lab has to pass, written by whoever publishes first. Anthropic activating ASL-3 protections before shipping Claude Opus 4 is real safety engineering. It is also an output cap announced in public, and the two readings are not exclusive. That is the trap. A capability line drawn by the leaders, in the open, sets the pace for everyone behind them, which is exactly what a pace-setter's business model rewards. Nvidia sells throughput. The labs sell scarcity. Ask which one a given safety propos