Together with Broadcom, OpenAI has presented the first benchmark results for its own inference chip “Jalapeño” – and it beats Nvidia’s Blackwell systems on energy efficiency. The figures were independently verified by the analysis firm SemiAnalysis.
1.5 to 1.9 times more efficient
Across three public models – GPT-OSS 120B, DeepSeek R1 670B and Moonshot Kimi K2.5 1T – Jalapeño delivers 1.5 to 1.9 times more AI compute per watt than Nvidia’s Blackwell (tested against the GB200 and GB300 systems), according to OpenAI. On latency the chip is even 1.7 to 3.6 times ahead. It is optimized specifically for inference – running AI models in production – and was designed partly with AI assistance. It was unveiled at Hot Chips 2026.
Small volumes at first
The rollout begins in “very small volumes” at the end of 2026, with the major scale-up planned for 2027. OpenAI stresses that it will keep buying Nvidia hardware – its own chip is a complement, not an immediate replacement.
For Nvidia the news is nonetheless awkward: more and more major customers are developing their own silicon, putting pressure on the market leader’s high margins.
Sources: OpenAI, Tom’s Hardware, CNBC.
