OpenAI launches Jalapeño custom AI inference chip to cut ChatGPT costs

OpenAI announced its first application‑specific integrated circuit, the Jalapeño chip, designed to run large language models such as ChatGPT more efficiently. The chip was co‑developed with Broadcom and will be fabricated by TSMC using a 3‑nanometre process, with prototype deployments slated for the end of 2026.

Jalapeño features eight stacks of high‑bandwidth memory around a systolic‑array core optimized for the matrix multiplications required by generative AI, promising substantially better performance per watt and up to 50 % lower cost per inference token compared with current GPU solutions. Early internal benchmarks suggest the chip could reduce OpenAI’s reliance on Nvidia hardware and lower operating expenses for its AI services.

The company plans to install the chip in Microsoft‑operated and other partner data centres beginning in 2026, marking a strategic shift toward in‑house silicon to control compute costs and accelerate hardware development cycles.