OpenAI unveils Jalapeño chip, outperforms nvidia gb300 in efficiency

OpenAI has unveiled its first custom AI inference chip, Jalapeño, which it claims outperforms leading market systems in terms of energy efficiency. The chip, developed in collaboration with Broadcom and manufactured by TSMC using 3nm process technology, was showcased on August 25, 2026. OpenAI used the InferenceX benchmarking tool from SemiAnalysis to evaluate Jalapeño against NVIDIA’s GB200 and GB300 systems, using models such as GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. The results showed that Jalapeño achieves 1.5 to 1.9 times higher AI throughput per watt compared to the benchmark systems, with end-to-end latency ranging from 28% to 59% of the competing systems. When running GPT-OSS 120B, Jalapeño’s peak throughput per kilowatt was 1.9 times higher than GB200. On DeepSeek R1, it achieved 1.7 times higher performance per watt than GB300, and on Kimi K2.5, it was 1.5 times better. Jalapeño operates at a rated power of 700 watts but consumes less than 550 watts during testing. The chip is an ASIC designed for large language model inference, not training, and is not available for external sale. OpenAI’s hardware lead, Richard Ho, previously worked on Google’s Cloud TPU project. The development of Jalapeño took approximately nine months, marking one of the fastest ASIC development cycles in high-performance semiconductor history. OpenAI’s own AI models were integrated into the chip design process to accelerate optimization. The second generation of Jalapeño is in late development, with third-generation design already underway. OpenAI plans to deploy Jalapeño in small-scale by the end of 2026 and expand deployment in 2027. It noted that Jalapeño has not yet been tested against NVIDIA’s latest Vera Rubin chip, but emphasized its continued partnership with NVIDIA and a multi-supplier strategy to meet its computing needs.