OpenAI's custom inference chip Jalapeño outperforms Nvidia GB200 and GB300

On August 25, OpenAI unveiled its first self-developed AI inference chip, Jalapeño, developed with Broadcom and manufactured by TSMC using a 3nm process, designed specifically for large language model inference. Tests show that when running GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 models, Jalapeño delivers 1.5 to 1.9 times the throughput per watt of Nvidia's GB200 and GB300 systems, with end-to-end latency just 28% to 59% of the compared systems.