Meta Books Hundreds of Thousands of AWS Graviton Chips for Agentic AI

Meta has signed a multi-year agreement with Amazon Web Services to bring hundreds of thousands of AWS Graviton chips into its data centers, the two companies announced on April 24. The deal covers tens of millions of Graviton cores over at least three years and makes Meta one of the largest customers of Amazon's homegrown Arm-based silicon.

Most of the conversation about AI infrastructure has been about GPUs, the parallel-processing chips that train frontier models. The Meta-AWS deal points at the other half of the bill. Agentic systems, the kind that plan and chain multiple steps across many tools, lean heavily on CPU work: orchestration, code generation, search, and the connective tissue between model calls.

What Graviton brings

Graviton5, the chip generation underpinning the new arrangement, packs 192 cores per processor and a cache about five times larger than the previous generation, according to AWS. Amazon says the larger cache cuts inter-core latency by up to 33 percent, which matters for agent workloads where cores need to talk to each other constantly.

For Meta, the appeal is performance and economics. CNBC reports the deal sits in the multibillion-dollar range, and Graviton typically costs less to run than traditional x86 silicon for the workloads Amazon designed it for. Pairing AWS CPU capacity with its own GPU build-out gives Meta more room to maneuver as the AI compute bill keeps climbing.

Why it matters

The agreement is a useful signal of where compute demand is heading. Training a single model is one expense. Running a fleet of agents that wake up, look at tickets, fan out queries, and report back is a different one, and that second category is what the hyperscalers are now planning around.

It also tightens an existing relationship rather than replacing Meta's own infrastructure work. AWS has been a Meta cloud partner for years, and GeekWire notes that the deal extends, not redirects, that history.

Sources: Meta Newsroom, About Amazon, CNBC, GeekWire.