OpenAI’s Jalapeño Benchmark Promises Faster, Cheaper AI
OpenAI says its Jalapeño chip delivered up to 1.9 times more performance per watt than Nvidia Blackwell systems in its first published benchmark tests. The post OpenAI’s Jalapeño Benchmark Promises Faster, Cheaper AI appeared first on TechRepublic .
OpenAI has unveiled its custom Jalapeño inference chip, claiming it delivers up to 1.9 times greater performance per watt than Nvidia's Blackwell systems in benchmark tests. The chip, co-developed with Broadcom, was presented to OpenAI's leadership in June 2026. Designed to work across various models, Jalapeño was tested with GPT-OSS 120B and non-OpenAI models DeepSeek R1 670B and Kimi K2.5 1T.
The chip addresses different phases of AI inference, such as prefill and decode, to reduce latency and improve responsiveness. A key feature is a localized KV cache that keeps model data close to compute resources, minimizing data movement during inference. OpenAI plans to deploy Jalapeño within its infrastructure by the end of 2026, with potential benefits for businesses, including faster AI responses, increased service capacity, and reduced inference costs.
The chip is intended to complement existing accelerators from Nvidia and other partners, rather than fully replacing them.
Written by urgent.news from TechRepublic's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
