According to Beating AI News, OpenAI has released the first batch of test data for its first custom inference chip, Jalapeño, claiming it outperforms NVIDIA's GB300. The chip achieves Pareto best performance on three publicly available external models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Compared to the best existing commercial systems, it offers 1.5 to 1.9 times higher peak AI output per watt and 1.7 to 3.6 times lower end-to-end latency; in highly interactive agent scenarios, the advantage expands to 2.1 to 4.1 times. Jalapeño has a rated power of 700 watts, with tested continuous power consumption not exceeding 550 watts. OpenAI emphasizes that under agent loads, true cost should be measured by "AI workload per unit power consumption," rather than single-chip performance.
Jalapeño was developed from design to tape-out in just nine months, with AI deeply involved in circuit optimization and verification. Leveraging Codex and GPT-Astra, the team optimized three open weight models not originally planned for this chip to high performance in just two months, with AI-generated implementations for some modules being 1.5 to 1.8 times faster than human handwritten implementations. OpenAI plans to deploy Jalapeño on its own computing infrastructure by the end of the year, while continuing to utilize external accelerators such as NVIDIA on a large scale. This is the first generation of a multi-generation chip roadmap; Gen 2 is already under deep development, and Gen 3 is taking shape.