Meta Buys Tens of Millions of AWS Graviton 5 CPU Cores — a New Front in the AI-Infrastructure Arms Race

Meta and Amazon disclosed on Friday that Meta will purchase tens of millions of AWS Graviton 5 processor cores — Amazon's fifth-generation in-house ARM CPU — to power agentic AI workloads inside Meta's data centers. The deal makes Meta one of the largest Graviton customers in the world, on par with the very largest AWS internal services, and it is structured to be expandable as Meta's needs grow. Meta will run agentic-AI orchestration workloads on the Graviton 5 cores initially, with the option to migrate to its own ARM-based silicon — a project called the "AGI CPU" that Meta unveiled with ARM in March — as that program matures. Neither company disclosed pricing or contract duration.

The CPU Side of the AI Stack

The conventional narrative of the AI infrastructure boom is a GPU story: Nvidia, AMD, and a long tail of custom accelerators competing to train and serve frontier models. Friday's deal is a reminder that agentic AI — systems that plan, call tools, coordinate sub-agents, and run for minutes-to-hours of wall-clock time per task — generates a different bottleneck. The orchestration layer (planning, tool dispatch, memory management, context construction, retrieval) is dominantly CPU-bound, and ARM cores designed for cloud workloads are a better fit for that workload than either GPUs or general-purpose x86 server chips on a per-watt and per-dollar basis. Meta's combined CPU strategy now spans three vendors: Nvidia's Grace ARM CPUs (deployed in February), AWS Graviton 5 (this deal), and Meta's own AGI CPU with ARM (in development with no shipping date yet). The company is hedging across all three because none of them is a perfect fit and because supply on any single vendor is constrained.

What This Says About Meta's Agent Roadmap

The scale of the Graviton 5 commitment — tens of millions of cores — is consistent with Meta planning a very large internal deployment of agentic systems, likely supporting both consumer-facing assistants across Meta's apps and internal automation workloads. CPUs at that scale also imply Meta expects most agentic compute to run on commodity hardware rather than on the GPU clusters reserved for frontier-model training. For the broader market, the deal is a significant validation of AWS's chip strategy: Amazon has spent eight years building Graviton against Intel and AMD on price-performance, and landing Meta as a flagship external customer at this scale puts Graviton on the same competitive footing as Nvidia's Grace and Google's Axion in the agentic-AI infrastructure race.