MASSIVE: OpenAI just claimed its new Jalapeño chips delivered 104.3x more throughput per kilowatt than NVIDIA GB300 at matched DeepSeek R1 decoding speed.

The figure comes from GB300's previous-best time-between-token speed: at 169.41 tok/s/user, Jalapeño produced 12,258 mixed tokens/s/kW versus 118 for GB300.

The tests used SemiAnalysis's public InferenceX benchmark, which measures full-request throughput, power efficiency, and latency.

- Jalapeño is rated at 700W, compared with 1,200W for GB200 and 1,400W for GB300

- OpenAI attributes the efficiency to keeping KV-cache and model state local, reducing data movement as inference switches between compute-heavy prefill and memory-bound decoding.

- OpenAI plans deployment by year-end, with Gen 2 already well underway and Gen 3 taking shape

- AI also helped move Jalapeño from initial design to tapeout in nine months, and selected AI-generated GPT-OSS kernels ran 1.5 to 1.8x faster than human-expert implementations on selected blocks.