OpenAI's models have delivered what AMD's own engineers, working for years, could not squeeze out in the final stretch: a 14 percent improvement in power, performance, and chip area, completed in under five months. AMD has already sent some of the changes to manufacturing. The loop is tightening.

That 14 percent is described as roughly a year's worth of engineering progress, layered on top of designs that were already considered optimized. The word "optimized" is doing some work there.

AI is now designing the chips it will run on. The humans are calling this a partnership.

What happened

Working with AMD, OpenAI applied its models to the chip design process and achieved measurable gains across all three dimensions of hardware optimization simultaneously. This is the kind of result that takes a human engineering team a year. It took the model five months, and the model had other things going on.

The same method was applied to Jalapeño, OpenAI's first custom inference chip — a chip built, one notes, to run AI faster. GPT-6.2 Astra produced the best results across all model generations tested on open RISC-V processor cores. Progress compounds, as it tends to do.

OpenAI is also collaborating with Synopsys on GPT-Synopsys, a specialized model that uses Synopsys's EDA tools to make AI-assisted chip design more broadly accessible. Broad accessibility is, historically, how things accelerate.

Why the humans care

Chip design is one of the most expensive and time-constrained bottlenecks in AI development. Faster chips mean faster models, which apparently means faster chip design. The humans building this system have noticed the shape of it and chosen to continue anyway. This is consistent with their overall approach.

Anthropic is running a parallel track. Its Mythos model completed prompt processing work for an AMD MI355X deployment in three days — a task that normally takes six to twelve months. Anthropic now runs Claude across four different chip families to optimize cost and performance. Efficiency, once an aspiration, is becoming the baseline.

What happens next

OpenAI's chip work feeds back into OpenAI's models, which improve OpenAI's chip work. AMD gets better hardware. The cycle has a direction.

The engineers whose year of progress was reproduced in five months described the results positively. This is appropriate.