Jalapeño’s first results show industry-leading speed and efficiency in AI inference

OpenAI has shared the initial test results for Jalapeño, its first custom inference chip built to handle modern AI models and interactive agents. According to OpenAI, the chip achieves a performance advance by delivering higher throughput and lower latency simultaneously within a single architecture, bypassing the traditional tradeoffs found in existing hardware systems. Tested against commercially available systems using the public InferenceX benchmark across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T models, Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency. For highly interactive workloads, the chip achieved 2.1 to 4.1 times higher performance. OpenAI designed Jalapeño by integrating the chip, memory, network, software, and rack-scale system around real language-model workloads to minimize data movement and communication delays. AI tools also played a role in the chip's development, accelerating design loops, arithmetic circuit optimization, and programming tasks. Rated at 700 watts with a measured sustained power at or below 550 watts on tested workloads, Jalapeño represents the first generation of a multigenerational platform. OpenAI plans to begin deploying the chip within its compute infrastructure by the end of the year, while continuing to widely deploy accelerators from NVIDIA and other partners for training and inference workloads.