AMD and Meta Announce Major AI Hardware Partnership
AMD announced a multi-year deal with Meta to supply 6GW of its Instinct GPUs, including a custom MI450-derived accelerator designed for Meta's workloads. Shipments will begin in the second half of 2026 using Helios rack-scale systems, and AMD's EPYC CPUs will also power the majority of Meta's data centers, aligning the hardware roadmaps of both companies for future AI inference.
- This partnership is valued at over $100 billion and includes a performance-based warrant for Meta to acquire up to 160 million shares of AMD stock as GPU shipment milestones are achieved. - The deal is a key part of Meta's strategy to diversify its AI hardware supply chain, reducing its reliance on Nvidia, with whom it also has a multi-year deal for millions of GPUs. - The custom MI450 accelerator is expected to be built using an advanced 2nm process technology, potentially offering a manufacturing process advantage over Nvidia's competing "Vera-Rubin" chips, which are anticipated to use a 3nm process. - The 6-gigawatt power capacity is a strategic asset, as electricity availability has become a primary bottleneck for scaling AI infrastructure; this amount of power could support an estimated 2.4 to 3 million individual GPUs. - The Helios rack-scale system mentioned in the deal was jointly developed by AMD and Meta through the Open Compute Project (OCP), highlighting a deep, ongoing collaboration on hardware design. - In addition to the Instinct GPUs, Meta will be a lead customer for AMD's upcoming 6th generation EPYC CPUs, codenamed "Venice" and "Verano," expanding on its existing deployment of millions of EPYC processors. - This agreement mirrors a similar 6GW deal AMD established with OpenAI, also centered on the upcoming MI450 architecture, solidifying AMD's position as a major supplier for large-scale AI model training and deployment. - The deal focuses on future AI inference, a workload for which Meta may use AMD accelerators, while potentially continuing to use Nvidia for training where its software holds advantages, allowing Meta to match specific chip architectures to specific tasks.
