AMD and Meta partner on massive 6GW GPU deployment for AI workloads

AMD and Meta have finalised a multi-year, multi-generation agreement to deploy 6 gigawatts of AI compute power. This partnership focuses on scaling Meta's infrastructure through the deployment of AMD Instinct GPUs and EPYC CPUs. The first gigawatt of this deployment will begin in the second half of 2026.

The collaboration centres on a custom AMD Instinct GPU based on the MI450 architecture. AMD tailored this silicon specifically for Meta's AI models and inference requirements. This move signals Meta's strategy to diversify its compute sources as it develops what CEO Mark Zuckerberg calls "personal superintelligence."

The Helios architecture and hardware stack

The deployment will use the AMD Helios rack-scale architecture. Developed through the Open Compute Project, Helios allows for standardised, scalable AI clusters at the rack level. The hardware stack for the initial phase includes:

• Custom MI450 GPUs: Tailored for Meta’s specific software and workload needs.

• 6th Gen AMD EPYC CPUs: Codenamed "Venice," these processors will handle orchestration and general compute tasks within the AI stack.

• ROCm Software: AMD’s open-source software platform will provide the programming environment for the deployment.

Meta is also a lead customer for "Verano," a next-generation EPYC processor designed to optimise performance-per-watt for specific workloads. This emphasises the growing role of CPUs in managing the complexity of large-scale GPU clusters.

Performance-based financial structure

The deal includes a financial structure that ties AMD's incentives to hardware delivery. AMD issued Meta a performance-based warrant for up to 160 million shares of common stock. These shares will vest in tranches as Meta reaches specific milestones for Instinct GPU shipments.

The first tranche vests when the initial gigawatt of hardware ships. This alignment ensures both companies focus on technical execution and commercial targets. AMD CFO Jean Hu noted that the structure supports long-term revenue growth and value creation for shareholders.

Diversifying the AI compute market

This agreement positions AMD as a primary provider in the global buildout of AI infrastructure. For Meta, the partnership reduces reliance on a single hardware vendor. Zuckerberg stated that diversifying computing is an important step for the company's long-term infrastructure goals.

The two companies are currently working to harmonise their silicon and software roadmaps. This integration aims to produce AI platforms that run Meta's services for billions of users with higher energy performance. Shipments for the 2026 launch will mark the first major step in this 6-gigawatt roadmap.