Story perspectives

OpenAI, Broadcom’s Jalapeño ASIC boosts AI inference per watt

8/26/2026

Summary

Story summary

OpenAI and Broadcom unveiled the Jalapeño ASIC, built exclusively for AI inference.
InferenceX benchmarks show Jalapeño delivers 1.5–1.9× more AI work per watt than Nvidia.
The chip delivers 1.7–3.6× lower latency and up to 4.1× better performance on interactive tasks.
Richard Ho, head of hardware, said Jalapeño minimizes data movement and keeps the KV cache local.
OpenAI will ship Jalapeño by end-2026, scale in 2027, and tape-out next-gen chips soon.

History

Historical Context

After Google bought DeepMind for $400 million in 2014 and began developing its own AI chips, the move signaled a shift toward specialized hardware. Building on that trend, OpenAI’s 2024 Jalapeño ASIC, co-developed with Broadcom, aims to cut inference power, echoing Google’s strategy that leveraged roughly 2.5 million servers in its data centers.

1 of 2

Story summary

OpenAI and Broadcom unveiled the Jalapeño ASIC, built exclusively for AI inference.
InferenceX benchmarks show Jalapeño delivers 1.5–1.9× more AI work per watt than Nvidia.
The chip delivers 1.7–3.6× lower latency and up to 4.1× better performance on interactive tasks.
Richard Ho, head of hardware, said Jalapeño minimizes data movement and keeps the KV cache local.
OpenAI will ship Jalapeño by end-2026, scale in 2027, and tape-out next-gen chips soon.

1 / 2