Meta and Broadcom announced an expanded partnership on Tuesday to co-develop custom AI accelerator chips through 2029, with an initial commitment exceeding one gigawatt of computing capacity and plans for a multi-gigawatt rollout. [1]

The numbers deserve a moment of translation. One gigawatt is roughly the output of a nuclear power plant. Meta is committing to deploy that much computing power in custom silicon — its own MTIA (Meta Training and Inference Accelerator) chips designed with Broadcom — before expanding further. The company said in January it expects to spend up to $135 billion on AI infrastructure in 2026 alone, and it is developing 31 data centers. [2]

The deal represents a strategic bet that custom chips, tailored to Meta's specific AI workloads, can deliver better performance-per-watt than off-the-shelf Nvidia GPUs. For Broadcom, the partnership secures a multi-year revenue stream from the world's largest social media company. For Nvidia, it signals that its biggest customers are building alternatives at scale.

The partnership includes chip design, advanced packaging, and optical networking — the full stack required to build AI infrastructure from first principles. Broadcom CEO Hock Tan simultaneously announced his departure from Meta's board, a governance cleanup that separates the customer relationship from the supplier's boardroom.

The power demands alone reshape the conversation about AI infrastructure from a computing question into an energy question.

-- DAVID CHEN, Beijing