Meta announced a multibillion-dollar, multi-year deal to run hundreds of thousands of AWS Graviton5 CPUs for agentic AI workloads. Graviton5 is built on 3nm process with 192 cores, a cache five times larger than Graviton4, and 33% lower inter-core latency. The deal makes Meta one of the largest Graviton customers in the world and was timed against Google Cloud Next -- not a coincidence.
GPUs are the right substrate for model training and large batch inference. For agentic workloads -- real-time reasoning, multi-step orchestration, code generation, search -- the access patterns are sequential, latency-sensitive, and state-heavy. They are not parallelizable in the way that drives GPU throughput. Meta running its $135B capex budget into a hard wall on GPU efficiency for agent serving is not a surprise internally. What is notable is the decision to formalize it at this scale and announce it publicly. That is a signal about where the AI infrastructure mix is heading, not just where Meta is today.
The specific properties that matter in Graviton5 are the 5x cache increase and the 33% inter-core latency reduction. Agent orchestration involves many short, dependent compute bursts with significant state passing between steps. Larger cache reduces the number of times that state has to touch main memory. Lower inter-core latency shrinks coordination overhead when multiple cores handle different branches of an agent graph. These are the workload characteristics that make ARM server CPUs competitive against x86 and GPU alike for this application class.
For hardware teams building diagnostic agents, embedded test orchestration, or AI-assisted ATE decision systems: profile before assuming GPU is the right substrate. The compute pattern for real-time embedded agent tasks -- sequential reasoning over instrument state, multi-step debug inference, decision trees over measurement data -- looks much more like Graviton than H100. The signal from Meta is that heterogeneous AI compute is now confirmed at hyperscaler scale. Budget the profiling exercise before infrastructure lock-in.