Meta and Broadcom Commit 1 Gigawatt to Custom AI Silicon — First Chips on a 2nm Process

Meta and Broadcom announced a sweeping extension of their MTIA (Meta Training and Inference Accelerator) chip partnership, committing to multiple generations of custom AI silicon through 2029. The initial deployment exceeds 1 gigawatt of Broadcom-designed compute — a commitment that signals Meta is serious about structurally reducing its dependence on Nvidia.

What 1 gigawatt of custom silicon actually means

One gigawatt of compute deployed in a data center context is an enormous infrastructure bet. And Meta’s 1GW commitment is explicitly the first phase — the announcement describes the broader rollout as “multi-gigawatt.” MTIA chips already power AI across Meta’s apps: ranking and recommendation in Facebook and Instagram feeds, Reels inference, generative AI features, and the models behind Meta AI. Scaling that infrastructure on custom silicon instead of Nvidia GPUs compresses cost per inference at volume.

Broadcom will handle chip design, advanced packaging, and networking for MTIA’s next generations. That’s comprehensive ownership of the entire silicon stack — from die design to the data center interconnect fabric.

The 2nm milestone

The next-generation MTIA will be the first AI accelerator built on a 2nm process node, according to Broadcom. This is a meaningful manufacturing milestone. TSMC’s N2 process delivers substantial density and power-efficiency improvements over N3. Moving AI inference workloads to 2nm silicon cuts the energy cost per inference operation, which matters enormously when you’re running billions of model calls per day.

For reference: most current AI accelerators, including Nvidia’s H200, are manufactured on TSMC’s N4 (4nm-class) process. Meta and Broadcom are skipping ahead by two full nodes.

Hock Tan steps down from Meta’s board

Alongside the deal announcement, Broadcom CEO Hock Tan confirmed he is leaving Meta’s board of directors. The separation is structured to avoid conflicts of interest as the commercial relationship between the two companies deepens into a long-term chip supply agreement. Tan joined Meta’s board in 2022 when the original MTIA partnership was announced.

The strategic picture

This deal is a direct bet that custom inference silicon beats commodity GPUs at scale. Meta already demonstrated with Llama 4 that it can train at the frontier. If MTIA delivers the efficiency gains Broadcom is projecting, Meta’s AI infrastructure cost structure improves significantly — and its exposure to Nvidia’s pricing and supply allocation for Blackwell and future Rubin GPUs shrinks with it.

Google has TPUs. Microsoft has Maia. Amazon has Trainium and Inferentia. Apple has the Neural Engine. Now Meta has a locked-in supply chain for custom AI silicon through the end of the decade. The hyperscaler custom chip playbook is fully mainstream.