Overview
- OpenAI unveiled Jalapeño this week, saying internal tests show the Broadcom‑made chip delivers more AI work per watt and lower end‑to‑end latency than Nvidia’s GB300/GB200‑class systems.
- Jalapeño is built only for inference, uses roughly 700 watts of power, and is not designed for model training or yet tested publicly against Nvidia’s newest Vera Rubin family.
- An independent firm, SemiAnalysis, tested Jalapeño inside OpenAI’s labs and reported superior tokens‑per‑megawatt numbers in those on‑site trials.
- OpenAI says it used its own generative models to speed chip design, moving from concept to tapeout in about nine months, and it is finishing a formal tape‑out while developing second and third generations.
- OpenAI expects the chip to lower operating costs for interactive AI services but will continue to rely on external suppliers for large‑scale capacity, and the move accelerates a wider industry shift toward specialized AI silicon.