OpenAI's Jalapeño Chip Shows Strong Inference Performance in Benchmarks
At the Hot Chips conference, OpenAI presented benchmark results for its new AI inference chip, Jalapeño. The system outperformed state-of-the-art processors like Nvidia’s Blackwell on metrics such as tokens per user and throughput per kilowatt. According to Richard Ho, head of hardware at OpenAI, Jalapeño offers both high efficiency in serving multiple users simultaneously and low latency for individual responses. Jalapeño is being developed through a close collaboration with Broadcom and will be deployed in small volumes by late 2026, expanding significantly into 2027. The chip’s design focuses on minimizing delays during the prefill and communication phases of inference processing—common bottlenecks that slow down AI response times. OpenAI aims to make Jalapeño a multigenerational platform with full-stack integration across models, chips, memory, and compute resources.
