OpenAI’s Custom AI Chip ‘Jalapeño’ Surpasses Nvidia in Efficiency Benchmarks

OpenAI, in collaboration with semiconductor titan Broadcom, has introduced Jalapeño—OpenAI’s inaugural custom AI inference chip. Preliminary benchmarks indicate that it surpasses Nvidia’s leading Blackwell systems in AI work per watt. SemiAnalysis‘s in-depth analysis reveals that Jalapeño’s B0 stepping offers 13.4 petaFLOPs of MXFP4 performance at a mere 700W, in contrast to the 900–1,150W requirement of Nvidia’s Rubin architecture. Furthermore, it achieves over 700 tokens per second per user on DeepSeek R1.

The chip’s design, from the initial concept to manufacturing tape-out, was co-created in a record-breaking nine months. This is approximately half the usual ASIC development timeline of 1.5 to 2 years. OpenAI’s own AI models played a significant role in expediting parts of the design and optimization process. The chip, manufactured on TSMC’s N3P (3-nanometer) process and coupled with HBM4 memory at 15.4 TB/s bandwidth, is specifically engineered for large language model inference workloads.

OpenAI and Broadcom aim for initial deployment by the end of 2026. They aspire to eventually power gigawatt-scale data centers in collaboration with Microsoft and other partners. This move signifies OpenAI’s strategic thrust towards vertical hardware integration and increased control over its rapidly escalating infrastructure costs as it gears up for a highly anticipated IPO.

Source: CNBC — OpenAI and Broadcom reveal Jalapeño, first AI chip in partnership