The Jalapeño inference accelerator, developed by OpenAI in collaboration with Broadcom and Celestica, was included in the TIME Best Inventions of 2026 list on October 15, 2026. The chip was first unveiled on June 24, 2026.
Technically, the processor is a custom ASIC designed specifically for inference workloads. It provides up to 13.4 petaflops of 4-bit compute, 232 GB of memory, and 15.4 TB/s of bandwidth. The chip operates at a sustained power of 550 W or less, with a maximum rating of 700 W.
Benchmark data from August 25, 2026, indicates that the chip performs 1.5 to 1.9 times more AI work per watt compared to Nvidia systems. OpenAI aims to reduce cost-per-token and per-kilowatt metrics by approximately 50% through this hardware.
The partnership between OpenAI and Broadcom, involving CEOs Sam Altman and Hock Tan, was announced in October 2025 with a goal of reaching 10 gigawatts of capacity by 2029. Small-volume deployment of the Jalapeño chip is scheduled for late 2026, with a transition to larger volumes expected in 2027.
