Meta and Broadcom Commit to One Gigawatt of Custom AI Chips Through 2029
What Happened
Meta and Broadcom extended their partnership through 2029 to co-develop multiple generations of Meta's MTIA (Meta Training and Inference Accelerator) chips, starting with an initial one-gigawatt deployment. The chips will use Broadcom's 2-nanometer process — the first custom AI silicon at that node. Broadcom CEO Hock Tan is stepping down from Meta's board to move into an advisory role. Meta has committed over $135 billion in capital expenditures for fiscal 2026.
My Take
One gigawatt of custom silicon is not a hedge against Nvidia — it's a declaration that Meta intends to own its AI compute stack end-to-end. The 2nm process detail matters: Meta will have custom chips on the most advanced node before most hyperscalers get there with general-purpose GPUs. This is the "build vs. buy" decision that defines who wins the AI infrastructure decade. Companies still renting GPU hours from cloud providers should watch this closely — when the platform you advertise on builds its own chip fab, your compute costs are their competitive moat.
Read Original Source