More Than Moore published an interview on 4 October with Richard Ho of OpenAI about Jalapeño, the company's own AI chip. The site reports that OpenAI confirmed the project in June 2026. Broadcom is the design partner, and Celestica handles board and rack work. Jalapeño runs trained models rather than training them. It has 216 GiB of HBM4 memory and a 700 W peak power rating. A full system links 2,048 chips. At HotChips 2026, Ho showed first results from working silicon. OpenAI claims 1.5x to 1.9x better performance per watt than NVIDIA's GB200 and GB300 at peak, and up to 104x better at some operating points.
The source says OpenAI's own models helped with chip layout and with tuning software for the hardware. One attention routine went from under one percent to nearly ninety percent of its speed limit in about forty hours. In the interview, Ho said models change fast, so the team kept the chip programmable. The text gives no price, ship date or list of customers. It mentions no independent tests of the benchmark claims. It also cuts off before the Broadcom and memory supply questions are answered.
Teams that buy or run AI inference now have a vendor's efficiency claim against NVIDIA's parts. The source does not say whether anyone outside OpenAI can use Jalapeño. Until outside tests exist, the numbers show what OpenAI says, not what buyers can expect.