Andrew Ng: More Agents Demand More Compute

This argues that autonomous agents will dramatically increase demand for compute, storage, and networking because agents perform many more automated actions than humans. Examples include a single paper-reviewing agent making roughly 5-10K web-search calls per day and consumer chatbots triggering multiple searches per prompt. Analytics and coding agents will generate far more data, queries, and transactions than manual workflows, forcing companies to rearchitect systems - just as online banking required new infrastructure when consumer transactions multiplied. Across industries, agents will multiply transaction volume and long-term storage needs, creating urgent demand for many more data centers.

The piece highlights a less obvious bottleneck: tool calls and data retrieval, not just LLM token throughput. Agents can quickly synthesize code or queries, but executing searches, scraping pages, and processing analytics often takes far longer than generating prompts. That means optimizing the entire software stack and scaling backend services is as important as model speed. Expanding data-center capacity is essential, and those facilities are presented as energy- and water-efficient for the scale of work. Political resistance to new data centers is slowing build-outs, but the expectation is that recognition of AI’s benefits will ultimately unlock the necessary infrastructure expansion.