Meta and Broadcom Sign a Multi-Generation MTIA Silicon Deal

What the Broadcom agreement actually covers

Meta announced an expanded partnership with Broadcom to co-develop multiple generations of its MTIA chips — the Meta Training and Inference Accelerator that Meta describes as purpose-built for inference and recommendation at scale.

The scope is broader than a chip order. According to the announcement, Broadcom will work with Meta across three layers: chip design, advanced packaging, and networking. The design work is built on Broadcom's XPU platform for custom AI accelerators, while Broadcom's Ethernet technologies handle high-bandwidth connectivity across Meta's compute clusters.

This is a follow-on to a plan Meta says it already disclosed: developing and deploying four new generations of MTIA chips within the next two years, aimed at ranking, recommendations, and generative AI workloads. The Broadcom agreement is framed as accelerating that existing roadmap rather than starting a new one.

The 1GW-plus commitment and what it signals about scale

The agreement carries a commitment that Meta says exceeds 1GW, described explicitly as the first phase of a sustained, multi-gigawatt rollout. Meta measures the deal in power capacity rather than chip count or dollars, which reflects how the constraint on this kind of infrastructure is increasingly electrical and thermal, not just procurement.

Zuckerberg's framing reinforces the phased structure: 'As we roll out more than 1GW of our custom silicon to start and then multiple gigawatts over time.' The word 'start' matters — the announced number is a floor, not a target.

MTIA as one piece of a portfolio, not a full replacement

Meta is careful to position MTIA within what it calls a portfolio approach to AI silicon, matching the right accelerator to each workload for the best mix of performance and total cost of ownership. MTIA is the accelerator optimized for inference and recommendation — the workloads that actually run Meta's apps for billions of users — not a general-purpose training chip meant to displace everything else.

That distinction is the strategic core here. Recommendation and ranking are Meta's highest-volume, most latency-sensitive workloads, and they run continuously. Custom silicon tuned to those specific jobs is where total-cost-of-ownership gains compound most directly, which explains why Meta is committing to multiple generations rather than a single design.

The governance change buried at the bottom

The announcement includes a corporate-governance detail that is easy to overlook: Broadcom CEO Hock Tan will transition off Meta's Board of Directors and move to an advisor role on Meta's custom silicon roadmap. Meta notes he has served on the board for the last two years.

Given the scale of this expanded partnership, Broadcom President and CEO, Hock Tan, will transition off of Meta's Board of Directors and will move to an advisor role for the company.Montana Labs

Meta frames this as a consequence of the deal's scale. A board seat and a multi-gigawatt commercial supply relationship create an obvious conflict, and moving Tan to advisor keeps his silicon expertise available while separating it from board-level fiduciary oversight of a partner Meta is now paying at scale.

Why buyer-led custom silicon is Meta's chosen dependency

For teams watching AI infrastructure, the specific implication is that Meta is choosing to own its accelerator roadmap while depending on a partner to execute it. The XPU platform gives Meta a customizable base, but Broadcom controls the design, packaging, and networking capabilities that turn that base into deployable hardware.

That trade is deliberate. Meta gets silicon tuned to its own recommendation and inference workloads across multiple generations; in exchange it locks into Broadcom for design and Ethernet networking through the multi-gigawatt buildout. The board-to-advisor move for Tan is the tell that both sides expect this relationship to be large enough, and long enough, to need formal separation from Meta's governance.