Meta Taps AWS Graviton Chips as Agentic AI Push Raises CPU Demand

Meta is deepening its AI infrastructure strategy with a new agreement to deploy millions of AWS Graviton processors, adding Amazon’s custom CPU line to the compute mix behind Meta’s expanding AI systems. The deal underscores a shift in the AI buildout: GPUs still dominate training, but agentic AI is making high-performance CPUs more important for production workloads.

That matters because agents do not only generate text or images. They coordinate steps, manage memory, route tasks, call tools, handle state, and run inference-heavy workflows that can place major pressure on general-purpose compute.

Agentic AI Is Changing the Compute Mix

Under the agreement described in Amazon’s announcement, Meta will use AWS Graviton processors as part of its broader AI expansion. Amazon says its latest Graviton chips have a cache five times larger than the previous generation, a design choice meant to improve data processing speed and bandwidth for demanding workloads.

The timing is important. Large language models created the first wave of demand for AI accelerators, especially GPUs used for training and high-throughput inference. Agentic AI adds another layer. These systems often need CPUs to support orchestration, memory management, task coordination, and the glue work around model calls.

In other words, the compute race is no longer only about who can buy the most GPUs. It is also about who can assemble the most efficient blend of accelerators, CPUs, networking, storage, and custom chips to keep AI systems running at scale.

Meta Wants More Compute Sources

Santosh Janardhan, Meta’s head of infrastructure, framed the AWS agreement as part of a diversification strategy. As Meta scales the infrastructure behind its AI ambitions, expanding to Graviton gives the company another source of compute for CPU-intensive workloads behind agentic AI.

That is a practical concern for Meta. The company is building AI into its social platforms, creator tools, assistants, recommendation systems, and long-term model research. Serving those systems to billions of users requires not just raw training capacity, but reliable production infrastructure that can keep inference and agent workflows efficient.

Nafea Bshara, a vice president at Amazon, described the deal as more than a chip arrangement, saying it gives customers an infrastructure foundation for AI systems that can understand, anticipate, and scale efficiently. That is the sales pitch Amazon wants to make with Graviton: not a replacement for AI accelerators, but a key part of the wider AI infrastructure stack.

Custom Chips Are Becoming Strategic Leverage

The Meta-AWS agreement fits into a wider rush among major AI companies to secure compute supply and reduce dependence on any single vendor. Earlier this month, OpenAI and Anthropic expanded partnerships with Amazon tied to Trainium, Amazon’s in-house AI accelerator. Meta has also struck large infrastructure deals with AMD and Nvidia, while expanding work with Broadcom on AI-specific chip design.

The broader trend is that hyperscalers and AI leaders are treating chips as strategic leverage. A TechTarget explainer on custom cloud hardware lays out why cloud providers are increasingly investing in custom silicon: control over cost, efficiency, performance, and supply can become a competitive advantage.

For Meta, Graviton is another piece in that puzzle. The company still needs GPUs and AI accelerators for frontier model work, but agentic systems require a broader infrastructure base. If those agents become a core interface for users and businesses, the supporting CPUs, memory systems, and orchestration layers become just as important as the model itself.

The AI Infrastructure Race Is Getting Wider

The lesson from the deal is not that CPUs are replacing GPUs. It is that AI infrastructure is becoming more specialized and more diversified at the same time. Training, inference, orchestration, memory, retrieval, and agent control loops each stress different parts of the stack.

That is why Meta’s AWS chip deal is worth watching. It shows one of the largest AI builders preparing for a future where AI workloads are not isolated model calls, but persistent systems that reason, act, coordinate, and run across enormous user bases. The winners in that environment may be the companies that can match each workload to the right compute, not simply the ones that buy the most of any one chip.