【#Tech24H】On August 25, OpenAI published a blog post showcasing its first inhouse developed Jalape？o chip, claiming that it surpasses the current toptier inference processors on the market in both peruser token throughput and kilowatthour throughput. In its public testing of Jalape？o, OpenAI used its own relatively small opensource model, GPTOSS 120B, as well as thirdparty models from DeepSeek and Moonshot AI. On Moonshot’s Kimi K2.5 1T model, Jalape？o demonstrated even more pronounced advantages. Jalape？o is OpenAI’s first selfdeveloped AI inference chip, codeveloped with Broadcom and manufactured by TSMC on a 3nm process. It was first publicly demonstrated on June 24, 2026. The chip is an ASIC (applicationspecific integrated circuit) featuring a systolic array architecture and equipped with HBM highbandwidth memory, designed specifically for large language model inference workloads, not for model training. Neither the chip nor its accompanying server systems will be sold externally. They are for internal OpenAI use only, aiming to reduce the company’s GPU procurement and computational infrastructure operating costs.[ By Zhang Liyan | Tang Ruohan ]