OpenAI unveils jalapeño chip as strong competitor to nvidia's gb300

The OpenAI hardware lead, Richard He, announced during a media briefing that the Jalapeño chip, developed in collaboration with Broadcom, has achieved significant performance improvements over current top-tier inference processors. In the SemiAnalysis InferenceX benchmark, Jalapeño outperformed existing solutions in both single-user token throughput and power efficiency. OpenAI claims that Jalapeño can handle more AI workloads with the same power consumption and deliver faster responses. The chip is designed for AI inference, enabling trained models to execute tasks or drive AI agents. OpenAI tested Jalapeño against NVIDIA's GB200 and GB300 systems, reporting that it achieves 1.5 to 1.9 times more AI work per watt on models such as GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. End-to-end latency is reported to be 28% to 59% lower than competing systems. OpenAI emphasized that full-stack development allows direct optimization of inference bottlenecks, particularly reducing latency in pre-filling and communication stages. The Jalapeño design minimizes data movement and communication delays by keeping model states like KV caches locally accessible. OpenAI plans to deploy Jalapeño in a limited capacity by the end of 2026, with broader deployment expected by 2027. The company has not disclosed specific chip quantities for 2027.