Meta Locks 1GW+ MTIA Deal With Broadcom Through 2029, Targets Multi-Gigawatt Buildout
Meta has committed to deploying more than one gigawatt of its custom MTIA AI accelerators in an expanded multi-year partnership with Broadcom that runs through 2029. The initial 1GW-plus deployment is described as the first phase of a “sustained, multi-gigawatt rollout,” with plans to scale to multiple gigawatts in 2027 and beyond.
The Deal
The partnership extends an existing co-design arrangement covering chip design, advanced packaging, and network switching. Broadcom is supplying the technology stack for Meta’s Training and Inference Accelerator programme, including its Ethernet networking fabric to connect AI clusters at scale.
The first-generation chip is already shipping. Three further MTIA generations are planned through 2027, targeting inference workloads — the process by which deployed AI models respond to user queries. Unlike Google or Amazon, which make their custom silicon available to cloud customers, Meta’s MTIA chips are used exclusively for internal workloads: AI recommendations, ranking, advertising, and generative features across WhatsApp, Instagram, and Threads.
The new MTIA chips will be the first AI accelerators built on a 2nm process node, according to Broadcom’s statement.
Hock Tan Exits the Board
Broadcom CEO Hock Tan is stepping down from Meta’s board and moving to an advisory role on the custom silicon programme. He joined the board in 2024. The departure is a direct consequence of the partnership’s scale: as the commercial relationship expands toward multi-gigawatt volume, his continued board presence raised independence concerns under standard corporate governance rules.
Tan’s exit removes a semiconductor industry veteran from Meta’s governance at the moment Meta is making its largest hardware bet. His shift to an advisory role keeps the institutional knowledge inside the relationship without the formal conflict.
Broadcom shares rose approximately 3% on the announcement. Meta stock was flat.
Context
Meta committed in January to spending up to $135 billion on AI in 2026 — the largest single-year AI capex commitment by any company. The Broadcom deal is the supply-side counterpart to that number: without a guaranteed chip roadmap at scale, the capex commitment would run into hardware delivery constraints.
The partnership also signals a structural shift in Meta’s GPU strategy. The company has ordered tens of thousands of Nvidia H100s and is continuing to invest in Nvidia hardware for training. But at inference scale, the economics of custom silicon become compelling. MTIA chips are optimised for Meta’s specific workloads, run at lower total cost of ownership than general-purpose GPUs, and are not subject to export controls or Nvidia’s allocation queues.
Mark Zuckerberg framed the deal in terms of the end goal: “building out the massive computing foundation we need to deliver personal superintelligence to billions of people.”
Key Numbers
- Initial deployment: 1GW+ of MTIA chips
- Multi-gigawatt target: 2027 and beyond
- Partnership extends: through 2029
- Process node: 2nm (industry first for AI accelerators)
- Meta 2026 AI capex commitment: up to $135B
- Broadcom shares on announcement: +3%
- MTIA use case: inference only, internal Meta workloads