SemiAnalysis / AI Business Weekly
Important
OpenAIChipsNvidiaHardwareOpenAI Claims Jalapeño Chip Beats Nvidia on Performance-per-Watt
August 25, 20263 min read
OpenAI has published benchmark results for its custom Jalapeño inference chip, claiming higher tokens-per-watt and lower latency than current Nvidia Blackwell/Rubin-class systems on certain open benchmarks, while noting the results are limited in scope.
Why it matters
Even a narrow efficiency win by a frontier lab’s custom silicon intensifies the competitive pressure on Nvidia and validates the strategy of co-designing chips with partners such as Broadcom.
OpenAI has released performance figures for its first custom inference chip, codenamed Jalapeño and co-designed with Broadcom. On selected open benchmarks the company reports meaningfully higher work per watt and lower latency than contemporary Nvidia Blackwell- and Rubin-class systems, while being explicit that the comparisons cover specific workloads and configurations.
The results are among the first public data points suggesting a frontier lab’s in-house silicon can outperform the dominant GPU supplier on efficiency metrics that matter for large-scale serving. Nvidia remains the clear volume leader, but the claims will be closely watched by hyperscalers evaluating total cost of ownership.
OpenAI has framed the numbers cautiously; independent validation and broader workload coverage will determine how seriously the industry treats the early lead.