Nvidia and Meta Reject AI Slowdown Pushed by OpenAI, Anthropic
Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg have broken with the leaders of OpenAI, Anthropic and SpaceX over proposals for a coordinated slowdown of AI development.
Zuckerberg sided with Huang in opposing a coordinated slowdown, according to reporting published September 16, 2026 CNBC. The split is now explicit. It centers on whether frontier capability gains should pause until alignment work catches up. Frontier means the most advanced systems. Alignment means making those systems behave safely.
Zuckerberg said there is debate about slowing progress on AI capabilities until alignment catches up. He favours a different control. He backs using independent evaluators over a slowdown to ensure AI safety The Straits Times. The approach would keep training and deployment moving while third parties assess risk. Scope, access and enforcement authority were not specified in the verified accounts.
Huang stated his position directly to political leadership. On September 14, 2026, he told Trump that 'we're not going to let [an AI slowdown] happen' TechCrunch. Huang and Zuckerberg have distanced themselves from safety proposals advanced by leaders of OpenAI, Anthropic and SpaceX Financial Times. That distancing was reported September 15. No joint proposal from Huang and Zuckerberg was described.
The counter-position was put forward in parallel. Anthropic CEO Dario Amodei published an essay titled "why the AI industry should slow down" outlining a three-part plan for doing so. OpenAI CEO Sam Altman said September 14 that it is 'right to be afraid of AI' but that the world should trust tech firms BBC. The CEOs of Nvidia, Anthropic and OpenAI also weighed in on AI progress and responsibility at Salesforce's annual event SiliconANGLE. The event provided a shared venue. It did not produce a shared stance.
The broader context here is a choice between two ways to manage risk. A coordinated slowdown means labs agree what triggers a pause, who verifies compliance, and what happens on defection, a bit like drivers sharing a speed limit. An evaluator regime means outside testers check systems before release, a bit like a safety inspection, and turns on evaluator independence, pre-deployment access, test design, and whether a negative finding can block release. Both regimes leave hard questions unresolved.
Looking at what matters for practitioners, the near-term variable is not rhetoric but testability. Under a slowdown, the binding constraint would be the threshold definition and the monitoring of training runs and deployments across labs and infrastructure providers. Under independent evaluation, the binding constraints would be access to models, scaffolding and system cards, plus the legal weight of an evaluator finding. Alignment catching up to capabilities is the stated criterion in Zuckerberg's description of the debate. Without an operational definition of alignment sufficiency, either path risks becoming procedural. For risk, audit and policy teams, the work is mapping which commitments are verifiable, which findings are blocking, and which party bears liability when an eval misses or a voluntary pause fails.
