Meta Inks Long-Term GPU Deal with AMD

- The deal is valued at up to $60 billion over five years and includes a performance-based warrant for 160 million shares of AMD common stock, potentially giving Meta up to a 10% equity stake in AMD if certain shipment and stock price milestones are met. - The first deployment, starting in the second half of 2026, will use a custom AMD Instinct GPU based on the upcoming MI450 architecture, specifically optimized for Meta's inference workloads. - This hardware will be delivered in the AMD Helios rack-scale architecture, a system co-developed with Meta through the Open Compute Project to standardize and scale AI infrastructure efficiently. - Meta is also adding AMD's Instinct MI300X accelerators to its data centers and will be a lead customer for the 6th Generation EPYC CPUs, codenamed "Venice" and "Verano". - The agreement hinges on the maturity of AMD's ROCm open-source software stack, a key alternative to Nvidia's dominant CUDA platform, which has historically maintained a performance advantage due to deep software and framework optimization. - This move is part of Meta's broader portfolio strategy to diversify its AI compute, which combines sourcing from partners like AMD and Nvidia with developing its own custom silicon, the Meta Training and Inference Accelerator (MTIA). - The deal reflects an industry-wide trend among hyperscalers to reduce vendor lock-in and optimize costs through custom silicon, illustrated by Microsoft's Maia 200, Google's TPUs, and Amazon's Trainium chips. - Mark Zuckerberg specifically highlighted the goal of deploying "efficient inference compute," targeting the recurring and rapidly scaling operational costs of running AI models, which can often surpass the initial one-time training expense.