AMD and Meta have further strengthened their collaboration on AI infrastructure by signing a GPU supply agreement worth up to 6 gigawatts. This gigawatt-scale contract, which translates to a power unit for large-scale data centers, demonstrates that the competition to expand global AI infrastructure is intensifying.

Under this agreement, announced on February 24 (local time) in Santa Clara and Menlo Park, California, AMD will supply Instinct GPUs over multiple generations to build Meta's next-generation AI infrastructure. The first 1-gigawatt shipment is scheduled to begin in the second half of 2026.

This agreement is characterized by the expansion of the existing strategic partnership into a multi-year, multi-generational structure. The two companies have agreed to align their roadmaps across semiconductors, systems, and software to build a platform optimized for Meta's large-scale AI training and inference workloads.

The initial deployment utilizes custom AMD Instinct GPUs based on the MI450 architecture. The system is built on 6th generation AMD EPYC processors, known by the codename 'Venice,' the ROCm software stack, and the Helios rack-scale architecture. Helios is an architecture jointly developed by the two companies through the Open Compute Project (OCP) that supports rack-level expansion.

AMD CEO Lisa Su stated that this collaboration will serve as an opportunity to provide high-performance and high-efficiency AI infrastructure by closely aligning the roadmaps of both companies. Mark Zuckerberg, CEO of Meta, also explained that long-term cooperation with AMD is important for diversifying computing infrastructure.

The agreement also included financial mechanisms. AMD issued performance-based warrants to Meta that allow it to acquire up to 160 million shares of common stock. These warrants are set to vest in stages linked to Instinct GPU shipment milestones, with the first tranche being vested upon the initial 1 gigawatt shipment. Subsequently, additional tranches are set to vest when the purchase volume reaches 6 gigawatts.

Meta is a customer that has already adopted MI300 and MI350 series GPUs and multiple EPYC CPUs. This contract is seen as an extension of AMD's strategy to expand its presence in the GPU market amidst surging demand for AI infrastructure. The industry is also paying attention to the possibility that the competitive landscape of the AI accelerator market, currently dominated by Nvidia, will gradually diversify.