OpenAI Secures Best Invention Honors for Custom AI Chip

What is Jalapeño?

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.

Specifications

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.

Partnership and Goals

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.

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.

Milestones to Watch

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.