Meta Shifts Agent AI to Tens of Millions of AWS Graviton Cores Instead of GPU

Meta is effectively betting on a new layer of AI infrastructure: the company has signed an agreement with AWS to deploy tens of millions of Graviton cores for agentic AI workloads. This is not just another GPU purchase for model training, but rather a sign that the market is beginning to split into two major directions. One involves training increasingly large models on accelerators.

The second involves serving a massive number of agentic scenarios post-training, where peak performance matters less than cost, energy efficiency, and predictable performance at scale. According to an official Amazon announcement from April 24, 2026, the rollout begins with tens of millions of Graviton cores and can be expanded as Meta's needs grow. Amazon notes that Meta has already become one of the world's largest Graviton customers.

We're talking about AWS Graviton CPUs, not GPUs: these are ARM processors developed by Amazon, available through AWS cloud. Meta plans to run part of its infrastructure supporting its AI services and processing billions of interactions on these cores, where complex multi-step workflows need to be coordinated.

Why this matters: the agentic AI boom is changing the very structure of hardware demand. GPUs remain essential when training large models or running particularly heavy inference. But once agents are layered on top of these models, the proportion of different types of tasks rapidly increases — real-time reasoning, code generation, search, action sequence planning, tool call orchestration, and managing long chains of steps.

Such workloads often hit not just accelerators, but also CPU, memory, inter-node communication, and the cost of each request. For companies at Meta's scale, this is no longer a technical detail but an economic question for the entire AI platform. For AWS, the Meta deal is several wins at once.

First, Amazon gets a very visible market signal: its own chips can be used not only within AWS but also in one of the most demanding AI infrastructures in the world. Second, it helps bring some of Meta's spending back to AWS. In August 2025, Meta signed a six-year cloud agreement with Google Cloud for over 10 billion dollars, and against this background, the new Amazon contract looks like a step toward a more diversified computing purchasing scheme.

Third, AWS strengthens its position in a new segment: not just "cloud for models," but a supplier of a complete stack for agentic AI.

There is another context as well. On April 20, 2026, Anthropic announced an expansion of its partnership with Amazon and a commitment to spend over 100 billion dollars on AWS over ten years, including capacity based on Trainium. Against this background, the Meta-Graviton partnership shows that Amazon is trying to establish itself in multiple layers of AI infrastructure: on accelerators for training and inference, and separately on CPUs for agentic and service workloads.

Amazon's additional argument is economics. Graviton5, according to the company, was created specifically for such scenarios, offering up to 192 cores, a larger cache, and up to 25% performance improvement over the previous generation. For customers, this means an attempt to reduce the cost of AI operations without sacrificing scale.

The main conclusion is simple: the race for AI chips is no longer just about GPU shortages and Nvidia's dominance. Major players are beginning to assemble hybrid compute stacks, where each category of hardware handles its own domain: GPUs for training, specialized accelerators for some inference, CPUs for orchestration, agentic chains, and mass application workloads. The Meta-AWS deal shows that the next competition is over who will provide the best price per unit of useful AI work. And if agentic products truly become the main interface to models, demand for such CPU architectures will grow no less rapidly than demand for classical AI accelerators.