OpenAI unveils Jalapeño, a custom chip it says beats Nvidia

OpenAI unveiled Jalapeño, a custom 700-watt AI inference chip co-developed with Broadcom that reportedly outperforms Nvidia's top Blackwell accelerators on efficiency and latency in early benchmarks. The chip targets inference workloads with high memory bandwidth using HBM4, signaling a potential shift in AI inference hardware and increasing pressure on scarce high-bandwidth memory supply chains.

This Wire brief sits within Fusion42's coverage of AI & ML, and 6 sources have reported it between 25 Aug 2026 and 27 Aug 2026.