Meta to Buy Millions of AWS Graviton5 Chips, Cuts 8,000 Jobs

Meta will deploy tens of millions of Amazon Web Services Graviton5 processors over a three- to five-year period under a multibillion-dollar agreement, the company disclosed. The announcement came the same day Meta confirmed it will eliminate 8,000 jobs and leave about 6,000 open roles unfilled.

AWS Vice President Nafea Bshara described the arrangement as a multi-year, multibillion-dollar contract that makes Meta one of AWS’s largest customers for Graviton processors globally. The deployment covers the use of Amazon’s fifth-generation CPU processors, which AWS developed for production AI tasks.

Graviton5 chips include 192 cores per processor and are designed to assign cores to different tasks at once, allowing parallel processing for complex workflows. AWS positions the chips for so-called agentic AI workloads-software that can reason, generate code and manage multi-step tasks without constant human direction. The processors target CPU-heavy production needs such as handling live user queries, generating responses and running models in real time.

Meta’s infrastructure group said it will use Graviton5 to expand the compute base supporting its next-generation AI systems. Meta Head of Infrastructure Santosh Janardhan wrote that diversifying compute sources is a strategic priority and that Graviton will allow the company to run CPU-intensive agentic AI workloads with the performance and efficiency required at scale.

Technology firms have relied mainly on graphics processors for model training, while demand is growing for CPU architectures tuned to run trained models in production. Meta’s agreement with AWS follows a broader industry pattern of combining GPUs for training with CPUs for inference and orchestration as services move into large-scale use.

Meta and AWS framed the deal as part of an ongoing collaboration to run next-generation AI workloads. The companies provided the timeline and financial scale of the contract but did not disclose precise pricing or the exact shipment schedule for the processors.