OpenAI and Broadcom Unveil ‘Jalapeño’ AI Inference Chip to Boost Compute Efficiency

OpenAI officially unveiled 'Jalapeño' on June 24, 2026, its inaugural custom-designed AI inference chip developed in strategic partnership with Broadcom. The Application-Specific Integrated Circuit (ASIC) is specifically optimized to handle the intensive computational demands of running large language models (LLMs) on previously unseen data, allowing the startup to enhance the reliability and efficiency of its AI models while reducing its historical reliance on Nvidia GPUs.

The development of Jalapeño reflects OpenAI's commitment to building a full-stack infrastructure to achieve higher performance-per-watt metrics than existing state-of-the-art processors. By designing hardware tailored to its proprietary serving systems, memory movement patterns, and kernels, the company aims to push the boundaries of LLM inference performance. Broadcom provided critical implementation and networking technology, including Tomahawk silicon, to support the platform, positioning the collaboration as a key milestone in scaling data centers for future AI workloads.

This move underscores a broader industry pivot among hyperscalers—including Amazon, Google, and Microsoft—to develop internal silicon to combat the rising costs and scarcity of premium AI training hardware. While early tests indicate that Jalapeño operates near theoretical peak performance, OpenAI plans to publish a detailed technical report regarding its benchmark figures later this year. The platform is designed for flexible integration across various LLMs, and its deployment is expected to scale significantly as part of OpenAI's massive 10-gigawatt compute capacity initiative scheduled for later in 2026.