OpenAI publishes first benchmarks for its Jalapeño chip

OpenAI published first benchmark numbers for Jalapeño, its Broadcom-built inference chip, claiming up to 1.9x the throughput per kilowatt of Nvidia's GB300 racks. The gap matters because Jalapeño runs on 700 watts against GB300's 1,400. Nvidia's flagship racks draw twice the power for numbers OpenAI says its chip already beats on throughput and latency.

Tom's Hardware's benchmark report measured Jalapeño on the SemiAnalysis InferenceX suite. OpenAI's numbers there show 1.5x to 1.9x more throughput per kilowatt and 1.7x to 3.6x lower end-to-end latency than Nvidia's GB200 and GB300 rack systems. Jalapeño's single-token output per megawatt also edges past the multi-token figures Nvidia and CoreWeave published for Vera Rubin in July.

Per the companion OpenAI post, Jalapeño went from design to silicon in about nine months. Broadcom handled physical implementation and networking, and Celestica built the boards and racks. It targets prefill and inter-chip communication, the two bottlenecks OpenAI names as dominant in serving. Nine months is a product cycle, not a chip program. If OpenAI can repeat it, that changes who can plausibly build custom inference silicon.

The chip is still at engineering-sample stage while Rubin already ships to customers. Total cost per token between the two comes out roughly even. OpenAI hasn't run larger models like DeepSeek V4 Pro or Kimi K3 on Jalapeño at all. The workload mix behind the headline ratios is narrower than it looks.

OpenAI's senior data center executive left the company on August 25, according to a TechCrunch report on the departure. OpenAI said the move followed a reorganization of its infrastructure group "to support the scale and pace of our work." Turning nine months of engineering samples into racks that customers can order needs that organization intact.