The full stack behind abundant intelligence

OpenAI has released performance results from Jalapeño, its first custom inference chip designed to run AI models more efficiently. The chip achieved higher throughput per kilowatt and lower latency than existing commercial systems when tested on multiple model families, demonstrating advantages in speed and energy efficiency. This matters because OpenAI believes that integrating custom chips with software, models, and data centers creates a system where improvements in one area strengthen the others, ultimately delivering more useful intelligence from each unit of compute at lower cost. The company uses this integrated approach alongside partnerships with multiple hardware and cloud providers to balance capability, speed, efficiency, and cost across different AI workloads.