Meta Signs Multibillion-Dollar Graviton5 Agreement with AWS

Meta signed a multi-year agreement with Amazon Web Services to deploy "tens of millions" of Graviton5 CPU cores, making Meta one of the largest Graviton customers worldwide, according to DataCenterDynamics and Bloomberg. The companies say the initial deployment is intended to support CPU-intensive, agentic AI inference workloads, including multi-step orchestration and real-time reasoning, per Axios and Economic Times. DataCenterDynamics and Tom's Hardware report the deal represents a "multibillion-dollar" arrangement over several years, though exact financial terms were not disclosed.

AWS has stated that Graviton5 packs 192 Arm Neoverse V3 cores on a 3nm process with roughly 180 MB of L3 cache, and that the chip delivers about 25% higher compute performance versus its predecessor plus up to 33% lower inter-core latency, as reported at AWS re:Invent and summarized by Tom's Hardware and DataCenterDynamics. Tom's Hardware also reports AWS vice president Nafea Bshara confirming the contract runs at least three years and that the majority of the capacity will be deployed in the U.S.

Editorial analysis: Public reporting frames this agreement as part of a broader shift in AI infrastructure where CPU capacity is growing in importance for inference and agentic workloads. Axios and Bloomberg note that leading cloud providers are promoting in-house CPU and accelerator designs as GPU availability and Nvidia supply constraints tighten. Tom's Hardware cites Amazon CEO Andy Jassy describing agentic AI as "becoming almost as big a CPU story as a GPU story," an observation reported during the announcement.

Editorial analysis: Engineers and ML infrastructure teams should view this as evidence that large-scale inference, orchestration, and agent-style pipelines are increasing demand for high-core-count, cache-optimized CPUs in addition to GPUs. Organizations designing production inference stacks will likely need to benchmark CPU-bound components such as search, retrieval, tool orchestration, and light-weight model inference across both GPU and CPU instances to estimate cost-performance tradeoffs.

Editorial analysis: Monitor three indicators: 1) public capacity and reservation limits for Graviton instances at major clouds, which may reveal supply tightness; 2) pricing movements for high-core-count CPUs and for GPU spot/ondemand instances, which will affect workload placement decisions; 3) vendor disclosures about customer commitments or multi-year reservations, which can foreshadow longer-term shifts in procurement patterns. Observers should also watch whether other hyperscalers or chip vendors announce increased production or similar large-scale customer agreements.

Meta's AWS Graviton5 agreement is a high-profile example of growing CPU demand for agentic inference workloads, and reporting from multiple outlets places the deal in the context of constrained GPU supply and cloud providers pushing homegrown silicon. The announcement includes specific technical claims from AWS on Graviton5 performance and capacity commitments quoted by company officials, while financial terms remain described as "multibillion-dollar" by several outlets without line-item disclosure.