Meta announced it has secured a deal for millions of AWS Graviton chips -- Arm-based CPUs, not GPUs -- to handle its growing AI agent compute workloads. The deal landed at Google Cloud Next, a deliberate competitive signal from AWS.
The story under the headline is a workload split that is becoming structural. GPU dominance was built on training and dense batch inference. AI agents run differently: continuous reasoning loops, real-time code execution, tool coordination, and sequential token generation rather than parallelized batch jobs. That workload profile favors lower-latency, cost-efficient CPUs over high-throughput GPUs whose utilization craters when token generation cannot be batched. Graviton is Arm-optimized for exactly this -- high single-thread performance, low cost per core, and AWS-native integration. The deal is not a statement against GPUs; it is a statement that not all inference is the same.
The concrete downstream effect is silicon procurement decisions. If your inference fleet is doing agentic reasoning loops rather than batch embedding or image processing, the GPU-first assumption deserves a real benchmark, not a default. Anthropic's $100B Trainium commitment at AWS is for training; what Meta signed for is a different stage of the pipeline entirely.
Nvidia is not sitting still -- Vera, their new CPU, is also Arm-based and explicitly positioned for agentic inference. Nvidia building a CPU is an admission that the workload split is real. For hardware teams architecting AI inference systems, the decision tree just gained a branch that did not exist two years ago.