OpenAI publishes the first Jalapeño inference benchmarks and a multigeneration roadmap
OpenAI publishes Jalapeño's first InferenceX results on three open models: 1.5–1.9× higher peak performance per watt and 1.7–3.6× lower end-to-end latency.
What happened
Event details
OpenAI tested Jalapeño with GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T against commercial GB200 or GB300 systems. It reports 2.1–4.1× performance under high-interactivity loads, a 700 W nameplate power rating and no more than 550 W sustained power in the tests. Deployment in OpenAI infrastructure is planned to begin by the end of 2026; Gen 2 is in advanced development and Gen 3 planning has started. The reported 53.7–104.3× figures at fixed existing time-between-tokens are concurrent throughput per kilowatt, not single-user or Sol speedups, and OpenAI has not published quantified results for running Sol on Jalapeño.
Assessment
Why it matters
The first quantified results move Jalapeño from an engineering sample to comparable working silicon and show OpenAI trying to convert external ultra-low-latency inference into first-party, scalable inference economics.
Availability
Access notes
The chip remains in production qualification, software maturation and scaling preparation. Internal OpenAI deployment is planned to begin by the end of 2026; no external sale has been announced.