Meta and Broadcom Partner to Co-Develop Generations of Custom AI Silicon

Building the Foundation for AI at Scale

Meta and Broadcom have announced an expanded strategic partnership to co-develop multiple generations of custom AI chips. This collaboration focuses on Meta Training and Inference Accelerator technology to power AI workloads across apps like Instagram and WhatsApp. The agreement includes an initial commitment exceeding 1 gigawatt of compute capacity, representing the first phase of a massive hardware rollout.

Specialized Expertise and Technical Integration

Broadcom will contribute critical expertise in chip design, advanced packaging, and networking using its proprietary XPU platform. This partnership is essential because Broadcom provides the high-bandwidth Ethernet technologies and 2-nanometer architecture that Meta cannot replicate entirely on its own. While Meta leads in software and AI model design, the physical complexities of silicon fabrication require Broadcom's specialized intellectual property.

Accelerating the Silicon Roadmap

Partnering allows Meta to accelerate its hardware roadmap significantly while avoiding the overhead of building standalone semiconductor manufacturing from scratch. Meta plans to deploy four new generations of MTIA chips within the next two years. These processors are optimized for ranking and generative AI workloads, aiming for a 25-fold increase in raw computing power by late 2027.

Strategic Governance Shifts

To manage the scale of this multi-billion dollar deal, Broadcom CEO Hock Tan will move from Meta’s Board of Directors to an advisory role. This transition addresses potential conflicts of interest while keeping Tan’s architectural guidance accessible to Meta’s infrastructure teams. This ensures Meta maintains access to top-tier semiconductor strategy as it expands its global AI compute clusters. Full details on the infrastructure expansion can be found at the Meta Newsroom.