As AI workloads shift from simple chatbots to autonomous agents, CPUs are becoming increasingly important infrastructure. Experts from Arm and Google explain that agent tasks — orchestrating tools, calling APIs, managing memory, and executing code — are fundamentally CPU workloads. Google's Axion (Arm-based) processors and GKE Agent Sandbox are highlighted as solutions, with the sandbox capable of spinning up 300 isolated environments per second per cluster. Small models in the 8B parameter range can run efficiently on CPUs, and Google claims 30% better price-performance versus competing cloud providers for these workloads.
Table of contents
The workload shifted from answering to actingWhy agents need sandboxes, and lots of themThe efficiency pitch94 Impressions