OpenAI says its first self-developed inference chip Jalapeño outperforms NVIDIA’s GB300: it delivers 1.5 to 1.9 times higher AI output per watt, with latency cut by up to 3.6 times.

Beating AI Insight News Brief: OpenAI has released the first batch of test data for its first custom inference chip Jalapeño, claiming its performance surpasses NVIDIA’s GB300. The chip has reached the Pareto optimal frontier across three publicly available external models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Compared to current best commercial systems, Jalapeño’s peak AI output per watt is 1.5 to 1.9 times higher, end-to-end latency is 1.7 to 3.6 times lower, and its advantage expands to 2.1 to 4.1 times in high-interaction agent scenarios. Jalapeño has a rated power of 700 watts, with measured continuous power not exceeding 550 watts. OpenAI emphasized that under agent workloads, real costs should be measured by "AI workload completed per unit of power consumption" rather than individual chip performance. Jalapeño took only 9 months from design to tape-out, 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 the project to high performance in just two months; AI-generated implementations of some modules were 1.5 to 1.8 times faster than human-written ones. OpenAI plans to deploy Jalapeño to its own computing infrastructure by the end of the year, while continuing to use external accelerators like NVIDIA at scale. This is the first generation in a multi-generation chip roadmap; Gen 2 is already in deep development, and Gen 3 is taking shape.