OpenAI has made a big announcement with a significant agreement with Broadcom to design and supply its own chips in its data centers. The deal is meant to provide OpenAI with increased influence over the hardware to operate its large models. The project is scheduled to have a compute capacity of about 10 gigawatts built within a few years with the use of custom accelerators. It is one of several attempts by major tech companies to lessen reliance on the existing chip vendors and to customize hardware to their particular AI tasks.
Deal size and timeline
The collaboration will target a deployment that begins in the second half of 2026 and continues through to the end of 2029. OpenAI and Broadcom expect to build racks of systems powered by the new chips and to scale to about 10 gigawatts of total custom AI accelerators. That amount of capacity is very large and is designed to support both research and production-level AI services. OpenAI said the new hardware will come from a close co-design process, so the chips and systems reflect lessons learned from building frontier models.
Technical goals and features
OpenAI plans to embed model-specific innovations directly into the hardware. The goal is to speed up training and inference while cutting power and cost per unit of work. The custom chips will focus on matrix math and memory movement that dominate modern model compute. The systems will be rack scale and optimized for throughput. Broadcom will use its chip design and systems expertise to deliver hardware that matches OpenAI needs. The effort aims to unlock new levels of capability and to let OpenAI tune hardware and software together.
Industry reaction and impact
The deal sits alongside other recent compute partnerships that OpenAI has made. OpenAI is also working with Nvidia and AMD on separate supply and investment arrangements. The Broadcom collaboration moves OpenAI further into the realm of vertically integrated AI infrastructure. For the wider market, the news highlights how leading AI companies are building tailored supply chains. The shift may bring new business opportunities to the chip makers and systems vendors, but also increased competition. To end users, the change can translate to increased speed in the model features and reduced service costs in the long run.
OpenAI and Broadcom framed the project as a step toward more efficient and capable AI infrastructure. The pace and scale of deployment will shape how quickly the custom chips affect model performance and the cloud market. OpenAI said it will continue to use multiple vendors as it scales compute in the years ahead.