Meta has expanded its strategic partnership with Broadcom to co-develop multiple generations of custom AI silicon, securing the hardware foundation required for its long-term generative AI and recommendation engine goals. The collaboration focuses on the next iterations of the Meta Training and Inference Accelerator (MTIA), a purpose-built chip optimized for the high-scale inference demands of Meta’s social apps and services.
Utilizing Broadcom’s XPU platform and advanced Ethernet technologies, the partnership aims to integrate chip design, specialized packaging, and high-bandwidth networking to support Meta’s rapidly growing compute clusters. The agreement includes an initial commitment exceeding one gigawatt (GW) of power capacity, marking the first phase of a multi-gigawatt rollout intended to deliver real-time AI experiences to billions of users globally.
The shift toward bespoke silicon represents a critical move to optimize performance and total cost of ownership as Meta moves away from off-the-shelf accelerators. By co-designing four new generations of MTIA chips over the next two years, the company is positioning its infrastructure to handle increasingly complex ranking and generative AI workloads.
As part of the expanded deal, Broadcom CEO Hock Tan will transition from Meta’s Board of Directors to an advisory role to provide direct guidance on the firm’s custom silicon roadmap. This vertical integration of hardware and software allows Meta to tailor its compute architecture to its specific AI models, ensuring the efficiency and speed necessary to deploy “personal superintelligence” at an enterprise scale.