AI Pulse · Frontier Labs Watch

OpenAI’s custom inference chip, Jalapeño, has posted its first published benchmark results against Nvidia’s GB300 system. In independent InferenceX tests across three large models, including GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, the 700W chip delivered 1.5 to 1.9 times more throughput per kilowatt. It also showed up to 3.6 times lower latency than Nvidia’s 1,400W GB300, according to the published report.

Plain English, OpenAI may be able to run certain AI workloads faster while using materially less power. That matters because inference, the cost of running models in production, is where AI bills build up. Training a model is expensive, but serving thousands or millions of requests every day is where power, hardware capacity and response time become operational issues.

There are important limits. Jalapeño is only targeting low-volume production in late 2026, and it has not been tested against Nvidia’s upcoming Vera Rubin generation. These results don’t mean Nvidia has lost its lead. They do show OpenAI’s chip effort is looking like a credible performance option, rather than simply insurance against relying on one supplier.

For operators, the immediate lesson is to keep watching inference cost and speed, not just which model scores highest on a benchmark. This is the kind of thing we build into an AI command centre, tracking where the practical economics of AI are shifting for the business.