Meta Expands AI Capacity with AWS Graviton Chips

Meta signed a multiyear agreement with Amazon Web Services to deploy tens of millions of AWS Graviton cores, making Meta one of the worlds largest Graviton customers. The deployment will power CPU-intensive agentic AI workloads that require large-scale orchestration and real-time reasoning, and the deal is reported to run multiple years and be worth multibillion dollars. "This isnt just about chips; its about giving customers the infrastructure foundation, as well as data and inference services," said Nafea Bshara, Vice President and Distinguished Engineer at Amazon.

The contract centers on Graviton5, Amazons purpose-built CPU family. Graviton5 key architectural points and performance claims include:

• •192 cores per chip, delivering parallelism for high-concurrency workloads.
• •Cache five times larger than the prior generation, which Amazon says reduces inter-core communication delays by up to 33%.
• •Integration with the Nitro System for I/O, security, and availability improvements.

Meta will run these cores alongside other architectures and cloud services such as Amazon Bedrock, using Graviton for tasks that are more CPU-bound than GPU-bound. The company emphasizes that agentic AI workflowsreal-time planning, multi-step orchestration, code synthesis, and search-heavy reasoning pipelinesare often dominated by latency-sensitive, branching compute patterns where CPU throughput, memory bandwidth, and cost-efficiency matter more than raw matrix-multiply GPU FLOPS.

This deal is a strategic pivot in how hyperscalers and AI builders assemble compute stacks. For several years, GPUs dominated the headlines because of model training. Now, as systems shift toward agentic and interactive AI at scale, CPU-centric workloads regain prominence. Meta is explicitly pursuing a diversified compute portfolio: custom ASICs and data centers for certain model training and inference, GPUs from vendors like Nvidia for dense matrix work, and cloud-based Graviton capacity for high-concurrency, orchestration, and post-training operations.

The agreement also validates AWSs push to commercialize custom silicon beyond commodity CPUs. For AWS, securing a marquee customer like Meta strengthens Gravitons credibility for large-scale AI deployments and creates pricing and capacity leverage against competitors. For Meta, the multicloud and multi-architecture approach reduces single-supplier risk and lets it map workload characteristics to the most cost-effective hardware.

The deal could pressure pricing and deployment strategies across the ecosystem. Expect closer integration between Graviton instances and AI orchestration stacks, broader use of Amazon Bedrock-style managed services at Metas scale, and potential follow-on agreements with other hyperscalers. Monitor how this affects GPU procurement cycles, Metas on-prem versus cloud mix, and whether other large AI consumers adopt purpose-built CPUs for agentic workloads.