OpenAI and Broadcom unveil LLM-optimized inference chip

OpenAI and Broadcom have introduced Jalapeño, OpenAI's first custom Intelligence Processor designed specifically for large language model inference. According to the companies, the chip was built from the ground up to improve performance, efficiency, and scale across AI systems, departing from general-purpose accelerators adapted from earlier workloads. Developed in nine months through software-hardware co-development with Broadcom and Celestica, the accelerator program utilized OpenAI models to assist in the design and optimization process. Engineering samples are currently running machine learning workloads in labs at production target frequency and power, including GPT-5.3-Codex-Spark. While final performance measurements are ongoing, early testing indicates that Jalapeño will deliver performance per watt substantially better than current state-of-the-art alternatives. The architecture is designed to reduce data movement and balance compute, memory, and networking resources. Broadcom provided silicon implementation and networking technologies, such as Tomahawk networking silicon, to support large-scale production. Jalapeño is the initial product in a multi-generation compute platform created by the companies. Deployment is planned at gigawatt scale with data center partners, including Microsoft, beginning in 2026. The hardware effort aligns with OpenAI's broader strategy to expand its full-stack platform, integrating its product experience, models, kernels, and custom chip architecture to make advanced AI faster, more reliable, and more accessible.