OpenAI Unveils In-House 'Jalapeño' Chip Benchmarks Outperforming Nvidia Blackwell

OpenAI on Tuesday published benchmark results for its first in-house artificial intelligence processor, codenamed Jalapeño, developed in partnership with Broadcom. The custom application-specific integrated circuit (ASIC) is engineered specifically for AI model inference rather than general-purpose training. Evaluated on SemiAnalysis’s InferenceX platform across public models including GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, Jalapeño achieved 1.5 to 1.9 times higher output per watt than prevailing commercial systems. On DeepSeek R1, the hardware reduced end-to-end latency by 1.7 to 3.6 times and produced over 700 tokens per second for low-concurrency user workloads. OpenAI confirmed that internal AI tools assisted in the chip’s physical layout, bring-up, and code optimization. Rated at 700 watts TDP with sustained drawing below 550 watts, deployment into primary infrastructure is scheduled for late 2026 alongside continued use of external accelerators.