Greg Brockman says OpenAI's tech found chip optimizations that would've taken humans weeks (www.businessinsider.com)

OpenAI announced it used its own AI models to assist in designing custom chips with Broadcom, claiming the models uncovered optimizations that would have taken human engineers weeks to find. By “pouring compute” into already-optimized components, the models suggested changes that led to “massive area reductions” — smaller, more efficient layouts — and trimmed weeks off the production schedule. Greg Brockman said the optimizations weren’t magic inventions beyond human capability, but rather many practical tweaks (he estimated about 20 items) that designers would have reached only after another month of iteration; human experts still review and validate the AI’s proposals. For the AI/ML community, this signals a meaningful acceleration of hardware co-design: automated search through large combinatorial design spaces can speed EDA workflows, lower chip area and power, and shorten iteration cycles for accelerators tailored to large models. OpenAI and Broadcom plan a 10-gigawatt rollout of these custom chips across OpenAI facilities and partner data centers between 2026 and 2029, underscoring how bespoke silicon plus AI-driven design could materially reduce costs and scale compute capacity — while also highlighting the need for rigorous human-in-the-loop verification as models take on more substantive engineering tasks.