Meta is moving parts of its AI stack onto millions of Amazon Graviton CPUs. Amazon said Friday the pact will route more of the social network’s AI spending back to AWS even after Meta’s August 2025, six‑year, $10 billion deal with Google Cloud. Graviton is an ARM-based CPU family Amazon says was built for inference and agent-style workloads; the move underscores how buyers now match AI tasks to the hardware best suited to them.

Why CPUs, not GPUs, for some AI work

GPUs still dominate large-model training because they accelerate parallel matrix math. But once models are trained, many production tasks — answering queries, coordinating agents, routing requests, calling APIs and composing code — involve smaller, varied computations, real-time I/O, and branching logic.

CPUs and GPUs are optimized for different patterns:

  • CPUs (like Graviton): better at branching, serial tasks, many lightweight processes, and orchestration-heavy agent workloads.
  • GPUs: excel at parallel throughput for large-scale model training and some high-throughput inference.

AWS says its latest Graviton chips are designed with inference and agent workloads in mind, prompting some cloud customers to move parts of their serving and orchestration stacks to CPU gear tuned for those patterns.

What the Meta-AWS deal means for the cloud wars

Meta’s decision directs a material chunk of AI compute to AWS even after its six-year, $10 billion Google Cloud agreement in August 2025 and broader ties with Microsoft Azure. The new deal doesn’t negate those contracts but shows large buyers will split workloads across providers based on hardware fit.

For AWS, Meta is a marquee reference for Graviton and a proof point when pitching enterprises that custom CPU silicon can handle parts of modern AI stacks. Cloud competition is increasingly about custom silicon as well as services — from Nvidia’s Vera CPU to Google’s accelerators.

Amazon’s broader silicon strategy

AWS has expanded its chip portfolio: Graviton for general and certain AI-related CPU tasks, and Trainium and other accelerators for training and high-throughput inference. The timing of the Meta announcement also illustrates how cloud vendors use customer wins and product news to influence enterprise perception and sales momentum.

Anthropic, Nvidia and the rush for long-term capacity deals

Amazon’s approach includes multiyear capacity arrangements. Earlier this month, Anthropic agreed to a multibillion-dollar, multiyear commitment to run workloads on AWS.

Related Articles

Amazon said the agreement covers "millions" of Graviton CPUs and was announced on Friday.

This article was created with AI assistance.