OpenAI is using its own models to design the chips that power AI, and it says the approach is starting to pay off in real hardware. In a report published on November 18, the company described work with AMD, open-source processor cores and electronic design tools, with measurable gains in several areas.

The best-known example is Jalapeño, OpenAI's first custom inference chip, which the company announced in August and says was designed with the help of GPT-5 and Codex. Since then, the team has been working with AMD on a benchmark that measures power, performance and area, known as PPA, in actual chip blocks. The collaboration reached up to 14% PPA improvement in select blocks, and AMD has independently verified single-digit percentage gains across key blocks of its upcoming next-generation chips.

OpenAI also tested whether models can design a processor core from a written specification alone. Using the open-source RISC-V cores PicoRV32 and Ibex, models produced working designs in simulation, with GPT-6.2 Astra scoring highest at about 0.7 on PicoRV32 and 0.75 on Ibex, at an average of around 112,000 tokens per run. The company says the two cores it has designed so far are now available as open-source artefacts, as are three new benchmarks for chip work: HWBench, HLSBench and DFTBench.

The next step is to put a model directly inside the tools chip engineers use. OpenAI is building GPT-Synopsys, a model designed to use Synopsys electronic design automation tools and run workloads inside them, as part of a partnership with Synopsys.