OpenAI has announced that its first self-developed inference chip, Jalapeño, has surpassed Nvidia's GB300: AI throughput per watt is 1.5 to 1.9 times higher, with latency reduced by up to 3.6 times.

Insight Beat AI News Flash: OpenAI has released the first batch of benchmark data for its custom inference chip, Jalapeño, surpassing the performance of NVIDIA's GB300. The chip has achieved Pareto-optimal results on three externally published models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Compared to the best existing commercial systems, Jalapeño delivers 1.5 to 1.9 times higher peak AI output per watt and reduces end-to-end latency by 1.7 to 3.6 times. In high-interaction AI scenarios, the advantage expands to 2.1 to 4.1 times. Jalapeño has a rated power of 700 watts, with measured sustained power not exceeding 550 watts. OpenAI emphasizes that under AI workload, the real cost should be measured by "AI work completed per unit of power consumed" rather than just chip performance.

Jalapeño went from design to silicon in just 9 months, with AI deeply involved in circuit optimization and validation. Leveraging Codex and GPT-Astra, the team optimized three open-weight models not originally planned within two months to high performance levels. The AI-driven implementation of certain modules is 1.5 to 1.8 times faster than human-coded versions. OpenAI plans to deploy Jalapeño on its proprietary computing infrastructure by the end of the year while continuing to extensively use external accelerators like NVIDIA's. This marks the first generation of a multi-chip roadmap, with Gen 2 already in deep development and Gen 3 taking shape.