← Glossary Model
Jalapeño
OpenAI's first custom chip, designed with Broadcom to run finished models cheaply rather than train new ones.
Jalapeño is an inference ASIC, a chip built for a single narrow job: serving answers from models that are already trained. OpenAI designed it with Broadcom, built it on a TSMC N3P production line with HBM 4 memory, and presented it at the Hot Chips conference in August 2026. It draws 700 watts against 900 to 1,150 for Nvidia’s Rubin compute die.
The first published benchmarks put it 1.5 to 1.9 times ahead of Nvidia’s GB200 and GB300 racks on work done per kilowatt. The strategic point matters as much as the numbers: a company with its own credible chip negotiates with Nvidia from a very different position.
Mentioned in