OpenAI

AI Hardware · Semiconductors · AI Infrastructure

Milestone

OpenAI's Jalapeño chip beats Nvidia GB300 in first benchmarks

Building inference silicon in-house lets OpenAI tune chips, models, and software together instead of waiting on suppliers' roadmaps.

• OpenAI published its first independent benchmark results for Jalapeño, a custom inference chip built with Broadcom, at the Hot Chips conference on August 25, 2026.
• Jalapeño only handles inference (running trained models), not training, and OpenAI says it's a general-purpose accelerator rather than one tuned exclusively to its own models.
• Tested on SemiAnalysis's InferenceX benchmark, Jalapeño delivered 1.5x-1.9x more AI work per watt and 1.7x-3.6x lower end-to-end latency than comparison Nvidia GB200/GB300 systems across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.
• The 700W Jalapeño chip outperformed Nvidia's 1,400W GB300 flagship on throughput-per-watt, and for interactive agent workloads posted up to 4.1x higher performance.
• OpenAI plans to deploy Jalapeño in small volumes by the end of 2026 and ramp meaningfully in 2027, with second- and third-generation chips already in development.
• Jalapeño was first unveiled in June 2026 as an OpenAI-Broadcom partnership; this Hot Chips presentation is the first time OpenAI has released concrete performance numbers.
• Owning inference silicon gives OpenAI leverage against Nvidia's pricing and supply constraints as it races to serve growing model demand more cheaply and reliably.