OpenAI's Jalapeño Chip Smokes Nvidia's Blackwell in Inference

OpenAI Jalapeño: Better than Nvidia Blackwell

At Hot Chips, OpenAI unveiled Jalapeño, an inference chip built with Broadcom that outperforms every Nvidia, AMD, and Google chip in SemiAnalysis's InferenceX benchmarks. Despite a rapid 16-month development cycle, the chip delivers industry-leading performance per watt, even beating Nvidia's upcoming Rubin on token throughput per megawatt. SemiAnalysis details the architecture, software, and benchmark results, noting that Jalapeño is a general-purpose inference chip, not specialized for OpenAI models, and that its performance is achieved without speculative decoding.