A new benchmark shows that chatter between AI agents barely dents the energy bill; it is the extra thinking that chatter causes that actually costs you.
Researchers built a metric called agentic-eCAL to measure the energy cost of AI agent teams that pass tasks back and forth across networks, not just single model calls. It combines an energy model for the compute-heavy prefill and memory-heavy decode steps of each model call with a full accounting of network transport, from radio access to optical backbone. The team validated it against hundreds of GPU benchmark configurations on Nvidia A100 and H100 chips, testing 16 open-weight models across 8 different ways of organizing agent teams. The goal was to answer a question telecom operators building autonomous 5G-Advanced and 6G networks are already asking: where in the edge-cloud pipeline should agent teams actually run.
The headline number: sending text between agents costs just 0.25 percent of total workflow energy, whether it travels over 5G radio access, metro fiber, or optical backbone. That undercuts the common assumption that clustering agents physically close together is what saves meaningful power. The real cost is what a message triggers on the receiving end: another model call, more context to process, more prefill and decode, not the transmission itself.
For an industry still treating edge placement as mainly a bandwidth problem, that is a useful reality check. The paper notes the human brain runs cognition on roughly 20 watts; today's language models do not, and the study suggests the fix has nothing to do with where the wires run.