AWS Graviton5 goes GA with EC2 M9g/M9gd instances, claiming 25% gains for agentic workloads

AWS made Graviton5 generally available, launching it through EC2 M9g and M9gd general-purpose instances. The company positions Graviton5 as purpose-built for the agentic AI era, optimized for the complex and long-running tasks associated with advanced AI agents, and claims up to 25% better compute performance than Graviton4-based instances along with improved energy efficiency.

Amazon CEO Andy Jassy framed the launch in the context of an 11-year custom-silicon journey: the Annapurna team's first Graviton CPU has grown into a chip 'well-loved by our AWS' customers. The release lands alongside the broader AWS AI push — Claude Fable 5 arriving on Bedrock and a stated plan to deploy more than 1 million NVIDIA GPUs across global regions this year spanning Blackwell and Rubin architectures.

Graviton5's relevance to AI is primarily on the orchestration and inference-serving side rather than training: Arm-based CPUs increasingly handle the glue work of agentic pipelines — tool calls, routing, and serving — where energy efficiency directly cuts the cost-per-token that has become a top customer complaint. The competitive backdrop is Microsoft's Cobalt 200 and Google's Axion Arm chips, all racing to lower the operating cost of always-on agents. The question for buyers is whether 25% headline gains translate into real bill reductions once workloads are migrated.