Meta Signs Multiyear Deal to Use Amazon Graviton Chips

Meta Platforms signed a multiyear, multibillion-dollar agreement to use Amazon Web Services' Graviton processors for AI infrastructure, according to Bloomberg and WSJ. WSJ reports the deal covers "tens of millions" of Graviton cores and that AWS vice president Nafea Bshara said the contract length is between three and five years. Bloomberg and Quartz (QZ) describe the agreement as giving Meta access to hundreds of thousands of Graviton processors and making Meta one of the larger Graviton customers.

QZ and AWS statements describe the targeted chip as Graviton5, an Arm-based CPU built on 3-nanometer technology; QZ reports AWS saying Graviton5 has 192 cores, a cache five times larger than the prior generation, and delivers up to 25% better performance compared with its predecessor. WSJ and CNBC report that Meta will use Graviton central processing units rather than exclusively relying on GPUs, with coverage framing Graviton as intended for cost-efficient, always-on and CPU-heavy tasks tied to agentic AI workloads.

Multiple outlets place the AWS deal alongside Meta's recent compute commitments, including reported agreements to lease Nvidia GPUs via CoreWeave and Nebius worth roughly $48 billion combined, and a previously reported multibillion-dollar Tensor Processing Unit (TPU) arrangement with Google, per CNBC and QZ. Reporting in TechCrunch and CNBC frames the announcement as part of a broader trend where hyperscalers and AI firms diversify chip suppliers beyond Nvidia GPUs to manage cost, capacity, and task fit.

Bloomberg and other financial coverage note an immediate market reaction: Amazon shares rose on the news, and analysts highlighted that the deal gives AWS another route to compete in AI infrastructure. TechCrunch and CNBC report that AWS markets Graviton and Trainium as alternatives that can lower runtime costs and address inference and agent orchestration workloads where CPUs or custom accelerators may be more cost-effective than GPUs.

Editorial analysis: Agentic and real-time reasoning workloads often shift compute demand away from bulk model training toward higher volumes of inference and orchestration tasks, which typically map well to CPU-based instances and specialized accelerators. Industry observers note that allocating some workloads to CPUs can reduce GPU-hours used for inference, change cost profiles, and increase emphasis on system-level optimization such as caching and low-latency interconnects.

observers should track published utilization case studies or benchmark comparisons from AWS, Meta, and independent testing labs for Graviton5 on agentic workloads; monitor whether Meta expands Graviton usage beyond the initially reported scope; and watch third-party cloud and chip vendors responses, especially in pricing and new instance types. Also monitor contract durations and disclosed core counts in follow-up filings or statements, since WSJ reported a three-to-five-year timeframe but financial terms were not disclosed.

Reported caveats: the companies declined to disclose full financial terms in initial coverage (WSJ, Bloomberg), and public descriptions of performance and efficiency are reported as AWS claims rather than independently verified benchmarks (QZ, CNBC).