Meta has committed to deploying tens of millions of AWS Graviton5 cores in a multibillion-dollar agreement, positioning the social-media giant as one of the largest customers for Amazon's custom CPU infrastructure. The deal includes the option to expand capacity as Meta's AI workloads grow, signaling major enterprise validation for Graviton as a CPU platform for agentic AI.
Graviton5, launched in June 2026, packs 192 cores per chip on TSMC's 3nm process, with a 5x larger L3 cache, up to 33% lower inter-core latency, and DDR5-8800 memory. AWS positions the chip squarely at agentic workloads—reasoning, real-time code generation, and multi-step task orchestration—where CPU orchestration is the bottleneck, not GPU utilization. Meta will join Uber and Snowflake in scaling Graviton for agent infrastructure.
For architects, this validates a critical infrastructure shift: agentic AI systems are CPU-bound orchestrators, not GPU-bound trainers. Graviton's 192-core density and lower latency directly address the coordination overhead of persistent agent environments. Graviton now accounts for over 50% of AWS's new CPU capacity additions in the past three years and has crossed a $20 billion annual run rate—enterprise-grade infrastructure, not experimental.