OpenAI unveils Jalapeño ASIC promising faster, more efficient AI inference

In a blog post and subsequent briefing, OpenAI introduced its Jalapeño application-specific integrated circuit, designed for AI inference tasks. Partnered with Broadcom, the chip was measured on the InferenceX platform and beat Nvidia's top-tier GB200/GB300 superchips, delivering between 1.5 and 1.9 times more AI work per watt on models such as GPT-OSS 120B, DeepSeek R1 and Kimi K2.5. Latency improvements ranged from 1.7 to 3.6 times lower, which the company says will enable quicker responses and more reliable access as demand rises.

OpenAI intends to ship Jalapeño in small volumes before the end of the year and scale up production through 2027, while still relying on partners like Nvidia for other compute needs. Development of second- and third-generation versions is already underway.