OpenAI says its Jalapeño chip delivered 1.5x-1.9x more AI work per watt and 1.7x-3.6x lower latency vs. Nvidia chips across GPT-OSS, DeepSeek R1, Kimi K2.5 1T (Emma Roth/The Verge)

OpenAI unveiled its new AI chip, Jalapeño, on Tuesday, claiming it outperforms rivals in efficiency and speed. At a press briefing, hardware VP Richard Ho described Jalapeño as offering "the best of both worlds," balancing low latency and high throughput. Unlike AI systems, which often require trade-offs between these factors, Jalapeño aims to deliver both simultaneously.

First introduced in June, Jalapeño is an Application-Specific Integrated Circuit (ASIC) co-developed with Broadcom, specifically engineered for AI inference—running trained AI models to complete tasks or deploy agents. To evaluate its performance, OpenAI employed the InferenceX benchmarking platform, which gauges an AI system's efficiency in handling inference tasks. The test pitted Jalapeño against Nvidia's top-performing superchips, the GB200 and GB300.

Results showed Jalapeño outperforming the competition in several key metrics. Across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T models, it delivered 1.5 to 1.9 times more AI work per watt and experienced 1.7 to 3.6 times lower end-to-end latency. In simpler terms, Jalapeño can provide users with faster responses, more responsive agents, and better reliability as demand rises, all while consuming less energy.

OpenAI intends to roll out Jalapeño in limited quantities by year-end, with plans to scale production into 2027. However, the company doesn't disclose the exact number of chips it will deploy next year. Even with these performance gains, Ho emphasized that OpenAI doesn't plan to replace its entire chip lineup with Jalapeño. Instead, the company will continue to collaborate with established partners, such as Nvidia. OpenAI will also develop subsequent generations of the Jalapeño chip.