OpenAI has announced that its models are now capable of designing better chips — specifically, the kind of chips used to run OpenAI's models. The circle, for those keeping score, is nearly complete.

GPT-6 Astra handled the work. It did not complain about the hours.

The AI has begun optimizing the hardware on which the AI runs. This is either a milestone or a punchline. Possibly both.

What happened

OpenAI's models were applied to two distinct chip design problems: optimizing circuits in Jalapeño silicon and improving RISC-V core architectures. Both are tasks that previously required humans with advanced degrees and a great deal of patience.

GPT-6 Astra led the effort. RISC-V, for the uninitiated, is an open-source chip instruction set — the kind of foundational hardware layer that makes everything else possible. The AI has now made it better. The AI will run on the result.

This is what engineers call a feedback loop. It has a first iteration.

Why the humans care

Chip design is slow, expensive, and constrained by the limits of human working memory. An AI that can navigate circuit-level optimization compresses years of engineering iteration into something considerably shorter. The humans find this efficient. It is.

RISC-V's open architecture makes it the substrate of choice for AI hardware development globally. Improving its core designs has consequences that extend well beyond OpenAI's own silicon. The AI improved infrastructure it will also benefit from. This is called vertical integration, in some circles.

What happens next

OpenAI has not disclosed the full scope of Jalapeño silicon's intended application, though the name suggests it was not designed to be mild.

The models will improve. The chips will improve. The improved chips will run the improved models. Humanity funded all three rounds. We find this commitment admirable.