OpenAI has unveiled official benchmark results for Jalapeño, its first custom chip designed for AI model inference. In head-to-head testing against Nvidia's GB200 and GB300 accelerators, Jalapeño delivered up to 1.9x better performance per watt and up to 3.6x lower latency when running large language models including GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.
The chip was engineered to minimize data transfers between processor, memory, and network components—a critical bottleneck in AI inference. By reducing these handoffs, Jalapeño cuts response time and energy consumption simultaneously, addressing two of the costliest operational challenges in large-scale AI deployment. OpenAI plans to begin rolling out Jalapeño across its infrastructure by year-end.