OpenAI’s first homegrown AI chip has picked up some hardware of a different kind. Jalapeño, the inference accelerator OpenAI built with Broadcom, was named to TIME’s Best Inventions of 2026 list on October 15.
What Jalapeño actually is
OpenAI first unveiled Jalapeño on June 24, 2026. The company describes it as its debut “Intelligence Processor,” developed alongside Broadcom and manufacturing partner Celestica.
Technically, it’s a custom ASIC, or application-specific integrated circuit. The silicon is designed around a single workload — inference — rather than adapted to it after the fact.
The spec sheet, translated
Jalapeño delivers up to 13.4 petaflops of 4-bit compute. It pairs that with 232 GB of memory and 15.4 TB/s of bandwidth.
On power, the chip runs at or below 550 W sustained. Its maximum rating is 700 W.
Benchmark results published on August 25, 2026, showed Jalapeño doing 1.5 to 1.9 times more AI work per watt than Nvidia systems.
The chip reportedly went from design to tape-out in nine months.
The people and the plan
The project brings together OpenAI CEO Sam Altman, OpenAI hardware VP Richard Ho, and Broadcom CEO Hock Tan.
Their partnership went public in October 2025. The stated goal was to reach 10 gigawatts of capacity by 2029.
Jalapeño is the first concrete hardware to come out of that arrangement. Small-volume deployment is planned for late 2026, with a ramp to larger volumes expected in 2027.
OpenAI’s target is to improve cost-per-token and per-kilowatt metrics by approximately 50%.
Why OpenAI wants its own silicon
Nvidia has long dominated AI hardware. A custom chip gives OpenAI more control over both cost and performance — the same logic that has pushed large tech companies like Google, Amazon, and Microsoft to design their own processors rather than rely entirely on off-the-shelf parts.
OpenAI isn’t abandoning GPUs overnight. Jalapeño targets inference specifically, and a small-volume launch is a pilot, not a wholesale replacement.
What this means
The gap between 1.5 times and 1.9 times better efficiency is worth watching. Real-world workloads will determine where in that range the chip actually lands.
The milestones to watch are straightforward: whether small-volume deployment starts on schedule, whether independent results confirm the per-watt claims, and how quickly capacity grows toward the 2029 target of 10 gigawatts.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.