OpenAI's Jalapeño chip claims 3.6x lower latency than Nvidia's best — but it won't replace them
openai's jalapeño chip delivers up to 3.6x lower latency than nvidia gb200/gb300 superchips on inference benchmarks, signaling a shift to vertical integration for model deployment efficiency.
openai claims its jalapeño asic, designed with broadcom for ai inference, significantly outperforms nvidia's gb200 and gb300 superchips in internal testing. results show 1.5 to 1.9 times more ai work per watt and 1.7 to 3.6 times lower latency on models like deepseek r1 and kimi k2.5 1t. deployment begins in small volumes by late 2026, scaling significantly through 2027. hardware vp richard ho emphasized that jalapeño offers a latency-throughput balance usually requiring trade-offs, yet openai will continue using nvidia chips rather than replacing its entire ecosystem.