Meta Expands AI Chip Partnership With Broadcom

Meta and Broadcom announced an expanded, multi-generation strategic partnership that extends through 2029, committing an initial deployment of more than 1GW of power to support Meta's custom AI accelerators, MTIA. The companies said the collaboration will co-develop successive MTIA chips, leverage Broadcom's XPU platform, and use Broadcom's advanced Ethernet technologies to remove networking bottlenecks across large AI compute clusters. As part of the agreement, Broadcom CEO Hock Tan will leave Meta's board and assume an advisory role focused on Meta's custom silicon strategy. Mark Zuckerberg framed the deal as critical to building "the massive computing foundation we need to deliver personal superintelligence to billions of people."

MTIA is Meta's Training and Inference Accelerator portfolio, built to match purpose-built accelerators to specific workloads such as ranking, recommendation, and inference for generative features. The announced first product in the MTIA family, the MTIA 300, is already deployed to accelerate ranking and recommendation systems. Meta has published a roadmap for at least four chips, with subsequent generations optimized more heavily for inference and generative workloads. Broadcom positions its XPU platform as the co-design substrate that tightly couples logic, memory, and high-speed I/O to meet Meta's scale demands. Key engineering focal points include:

• •chip architecture and logic co-design across generations
• •advanced packaging to increase density and thermal efficiency
• •high-bandwidth Ethernet and networking to enable scale-up, scale-out, and cross-cluster communication

The press materials also reference an "industry's first 2nm AI compute accelerator" in the combined roadmap, signaling ambitions to move to leading-edge process nodes to improve compute-per-watt. The initial 1GW commitment is framed as the first phase of a sustained multi-gigawatt rollout through 2029, implying aggressive data-center expansions, power provisioning, and system-level integration work.

Hyperscalers are accelerating custom silicon programs to control cost curves and performance trade-offs as demand for model inference and real-time generative features explodes. This deal reinforces three industry dynamics:

• •hyperscalers are moving from general-purpose GPUs to heterogeneous, workload-specific accelerators
• •networking and packaging become first-class design constraints as clusters scale to gigawatts
• •large infrastructure vendors like Broadcom are becoming system integrators, not just component suppliers. For Broadcom, the partnership cements momentum after recent collaborations with other hyperscalers and positions its XPU and networking stack as a competitive alternative to GPU-dominant stacks. For Meta, deeper vertical integration with a hardware partner reduces exposure to third-party GPU pricing and supply variability while enabling tighter co-optimization between models and silicon

Deploying multi-gigawatt custom silicon requires synchronized advances: foundry availability at cutting nodes, datacenter power and cooling upgrades, firmware and systems software to manage heterogeneous accelerators, and sustained supply-chain coordination. Any delays at the foundry or networking layer could slow the rollout or raise integration costs. Financially, the move trades capital expenditure and integration complexity for lower long-term TCO and greater control over feature performance.

Expect technical disclosures on the next MTIA generations, concrete timelines for multi-gigawatt deployments, and signals from foundries about 2nm availability. Monitor operational updates on datacenter power provisioning, Broadcom's product qualification cycles, and how this affects Nvidia's market share and hyperscaler procurement strategies.