New Study: Where AI Runs Matters Less Than How Much It Thinks

AI "agents" are programs that don't just answer one question — they plan, look things up, run tools, and hand work to other agents. Phone companies want to run these agents across their networks for 5G and 6G. The big question: should each agent sit on your device, on a nearby cell tower computer, or in a giant data center far away? Until now, nobody had good numbers to answer that.

A research team ran hundreds of tests on high-end AI chips (NVIDIA A100 and H100), using 16 open AI models and 8 different ways of arranging the agents. They measured two things: how much electricity each step burns, and how much energy it takes to ship the agents' messages across network links — cell towers, city networks, and long-distance fiber.

The headline result is surprising. Sending text between agents costs almost nothing: about a quarter of one percent of the whole job's energy. The expensive part is what happens after the message arrives — the receiving agent has to read it, think it over, and generate a reply. For comparison, a human brain runs on roughly 20 watts, like a dim light bulb, while today's big AI models need far more. Communication is cheap; comprehension is not.

What does this mean for you? Phone and network companies can stop obsessing over where to place their AI computers and start focusing on making agents talk less. Fewer back-and-forth messages means less energy, lower costs, and cooler, longer-lasting devices. It also means the AI assistants coming to your phone may work fine without shipping everything to a distant server — better for battery life, speed, and privacy. The catch: this is early lab research, not yet tested on live commercial networks.