Google Cloud says AI agents must have far more CPUs than chatbots

Mark Lohmeyer, a vice president at Google Cloud, told EE Times that AI agents change what data centers must have. A chatbot uses mostly GPUs or TPUs (processors for AI). An agent also uses APIs, visits websites and operates code, and these are CPU tasks. He tells that the CPU share of AI work grows. The best ratio is not known at this time.

How agents differ

Agents are bursty. An agent can do nothing for a day or two, and then work hard. Its data must stay safe for days or weeks. Lohmeyer tells that the answer is to keep idle agents in storage and start them again fast.

What Google does

Google Cloud uses microVMs (small, light virtual computers) to keep agents apart on one server. Lohmeyer tells that the same hardware can operate 1,000 microVMs, not 100 traditional virtual machines. He also names agent gateways that control traffic out of the network.

What is not known

These are the claims of one Google manager. The source gives no test results, no costs and no ratio of CPU to GPU. Lohmeyer tells that no single CPU is correct for all agents.

Insight: Agent work is a CPU and storage problem

GitHub told on 29 October 2026 that agents cause faults in services that worked correctly before. It adds 4 million cores. Google showed a microVM for each agent on 22 October 2026. This reads as cloud firms with one answer to a problem that GitHub shows in public. Agent work looks like a CPU and storage problem, not only a GPU problem. Tell your host to state the price of idle agents.