Technology
OpenAI presents benchmark results for its Jalapeño inference chip
Developed with Broadcom, the specialized chip is scheduled for initial small-scale deployment in late 2026.
The short version
- OpenAI presented benchmark details at the Hot Chips conference for Jalapeño, a custom Application-Specific Integrated Circuit built with Broadcom for running AI inference.
- The company stated that tests showed higher throughput per kilowatt and lower latency compared to existing systems like Nvidia's Blackwell chips.
- Initial deployment in very small volumes is projected for late 2026, ahead of broader deployment in 2027.
- Competitor performance may advance before the chip reaches full-scale deployment.
Key facts
- OpenAI released benchmark results for its Jalapeño chip on the SemiAnalysis InferenceX benchmark during the Hot Chips conference.[TechCrunch]
- The Application-Specific Integrated Circuit was developed in collaboration with Broadcom specifically to handle AI inference tasks.[TechCrunch · The Verge]
- OpenAI's hardware lead Richard Ho stated the chip will deploy in very small volumes at the end of 2026, followed by larger deployment in 2027.[TechCrunch]
- The chip architecture is designed to minimize data movement and latency bottlenecks during the prefill and communication phases of model processing.[TechCrunch]
What remains uncertain
- Whether Jalapeño's benchmark advantages over Nvidia's Blackwell system will persist by the time the chip reaches full deployment in 2027 as competing hardware evolves.[TechCrunch]