Key Takeaways
1. Meta and AMD have partnered to deploy up to 6 gigawatts of AMD Instinct GPUs for Meta’s AI infrastructure.
2. The first 1-gigawatt deployment is set to begin in the latter half of 2026, using customized MI450 GPUs.
3. The deployment will integrate MI450 GPUs with 6th generation EPYC CPUs and ROCm software.
4. AMD’s performance-based warrant could lead to a 10% ownership stake in AMD if fully exercised.
5. The partnership aims to enhance Meta’s AI infrastructure with greater flexibility by collaborating with multiple suppliers.
Meta and AMD have announced a long-term partnership that will see the deployment of up to 6 gigawatts of AMD Instinct GPUs in Meta’s upcoming AI infrastructure. This agreement is part of Meta’s strategy to broaden its AI computing resources and ensure better alignment between its GPU/CPU technology, systems, and software.
Deployment Timeline
AMD plans to start shipping the first 1-gigawatt deployment in the latter half of 2026. This deployment will utilize a customized Instinct GPU based on AMD’s MI450 architecture. It will be integrated into the Helios rack-scale architecture, which the two companies have previously discussed in the context of the Open Compute Project.
Technical Specifications
The initial deployment will combine the MI450-based GPU with 6th generation EPYC CPUs, codenamed “Venice,” running on ROCm software. A noteworthy aspect of AMD’s announcement includes a performance-based warrant for up to 160 million AMD shares, which will vest as shipment milestones are achieved, beginning with the first gigawatt and scaling towards the full 6 gigawatts. There are also additional technical and commercial requirements that must be met.
Reports from Financial Times and AP suggest that if the warrant is fully exercised, it could equate to a 10% ownership stake in AMD.
Strategic Goals
Meta emphasizes that this partnership aims to enhance its AI infrastructure by fostering greater flexibility and resilience through collaboration with multiple partners, instead of depending solely on a single supplier. Meanwhile, AMD characterizes the deal as a long-term ramp that encompasses GPUs, EPYC CPUs, and rack-scale systems tailored to meet Meta’s specific workloads.
As major AI customers continue to secure long-term capacity amidst ongoing demand for accelerators and the expansion of new data centers, this agreement comes at a crucial time. The Financial Times has reported extensively on the scale of Meta’s anticipated AI infrastructure investments and its initiative to diversify its supplier network.
Source:
Link