Meta moves agentic AI onto tens of millions of AWS Graviton cores instead of GPUs

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 another purchase of GPUs for model training. It is a sign that the market is starting to split into two large directions.

One is training ever larger models on accelerators. The other is serving a huge number of agentic scenarios after training, where peak performance is no longer the only concern and cost, energy efficiency and predictable operation at scale matter just as much. According to an official Amazon statement dated April 24, 2026, the rollout begins with tens of millions of Graviton cores and can be expanded as Meta's needs grow.

Amazon says Meta has already become one of the largest Graviton customers in the world. These are CPUs from the AWS Graviton family, not GPUs: ARM processors designed in-house by Amazon and available through the AWS cloud. Meta plans to run on them part of the infrastructure that supports its AI services and the processing of billions of interactions, where complex multi-step workflows have to be coordinated.

Why it matters: the boom in agentic AI is changing the very structure of hardware demand. GPUs remain indispensable when you need to train large models or run particularly heavy inference. But once agents appear on top of those models, the share of a different type of task rises sharply: real-time reasoning, code generation, search, planning sequences of actions, orchestrating tool calls and managing long chains of steps.

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

First, Amazon gets a very visible external signal to the market: its own chips can be used not only inside AWS but also in one of the most demanding AI infrastructures in the world. Second, it helps win back part of Meta's spending for AWS. In August 2025, Meta signed a six-year cloud agreement with Google Cloud worth more than $10 billion, and against that backdrop the new contract with Amazon looks like a step toward a more diversified compute procurement scheme.

Third, AWS strengthens its position in a new segment: not just a "cloud for models" but a full-stack provider for agentic AI. There is another piece of context. On April 20, 2026, Anthropic announced an expanded collaboration with Amazon and a commitment to spend more than $100 billion on AWS over ten years, including for capacity based on Trainium.

Against this backdrop, the Graviton partnership with Meta shows that Amazon is trying to establish itself across several layers of AI infrastructure at once: on accelerators for training and inference, and separately on CPUs for agentic and service workloads. Amazon's additional argument is economics. According to the company, Graviton5 was built specifically for such scenarios, offering up to 192 cores, a larger cache and up to 25% higher performance than the previous generation.

For customers, this means an attempt to lower the cost of AI operations without giving up scale. The main takeaway is simple: the AI chip race is no longer just about GPU scarcity and Nvidia's dominance. Major players are starting to assemble hybrid compute stacks in which each category of hardware handles its own area: GPUs for training, specialized accelerators for part of inference, and CPUs for orchestration, agentic chains and mass-market application workloads.

The Meta-AWS deal shows that the next fight is over who offers the best price per unit of useful AI work. And if agentic products really do become the main interface to models, demand for such CPU architectures will grow no more slowly than demand for classic AI accelerators.

*Meta is designated an extremist organization and banned in Russia.