THE CRUNCH

OpenAI says its models delivered up to 14 percent improvement in PPA, meaning power, performance and chip area, on AMD's flagship GPU IP in under five months. By OpenAI's own account, that is roughly a year of gains stacked onto circuitry AMD's engineers had spent years refining. Some of the changes have already been folded into a design sent to manufacturing, according to OpenAI.

The same approach was applied to Jalapeño, the first inference chip OpenAI has designed in-house, and in trials on freely available RISC-V processor cores, OpenAI's newest Astra model delivered the best results across model generations. The two accounts differ on its name: OpenAI's post calls it GPT-6 Astra, while The Decoder's report refers to GPT-6.2 Astra. OpenAI is also building GPT-Synopsys with Synopsys, a model that taps Synopsys's EDA tools, the software used to design chips, aiming to widen access to AI-assisted chip design.

Anthropic is doing similar work. It says its Mythos model took care of prompt processing for a model running on AMD MI355X GPUs in three days, a task it says normally takes six to twelve months, and the company now runs Claude across four chip families to optimise cost and performance per dollar. Note that the headline figures come from the companies themselves rather than independent verification.