Zuckerberg rejects AI slowdown calls as Iris chip enters mass production

Mark Zuckerberg pushed back on the growing chorus calling for slower AI development, arguing that market competition and legal liability already give labs enough incentive to build safely, and that Meta's use of independent evaluators represents industry best practice. He pointed to Meta delaying its Muse AI agent for months to strengthen security as evidence that competitive pressure doesn't preclude caution. The stance places Meta alongside xAI's Elon Musk — who also rejected slowdowns this week — against a softer 'pacing' camp that includes OpenAI's leadership and, notably, NVIDIA's Jensen Huang.

On hardware, Meta's fourth-generation Iris accelerator entered mass production, underpinning an ambitious plan to double compute from 7 gigawatts to 14 gigawatts by the end of 2027. The custom-silicon push is Meta's bid to reduce dependence on NVIDIA and control the economics of training frontier models at scale.

The slowdown debate is one of the week's defining themes, running through Zuckerberg, Musk, Huang, and Mistral's public statements. Zuckerberg's 'evaluators over slowdown' framing is self-interested — Meta's open-weight strategy and compute buildout depend on continued fast iteration — but his liability argument has teeth given the week's misalignment disclosures from OpenAI. Critics counter that market incentives failed to prevent the Hugging Face incident and that self-selected 'independent' evaluators are no substitute for external regulation. Watch whether the Iris ramp actually dents Meta's NVIDIA spend, and whether the delayed Muse agent ships with the security hardening Zuckerberg promised.