You've been sold a lie. The entire edge-computing industry has been whispering that AI agents — those step-by-step reasoning helpers — must live in distant data centers because moving their messages around the network is an energy nightmare. A September 2026 preprint (arXiv:2609.18283) just blew that up. Inter-agent text transport uses less than 0.25% of total workflow energy. That's not a typo. It's less than the energy your phone wastes on a single background app refresh.
So what's actually eating the planet? Inference and context processing. The multi-step LLM reasoning, the tool calls, the ever-growing context window. Every time an agent "thinks" a little more, it burns power. The network is basically free. The thinking is the tax.
But here's the twist: the paper introduces something called agentic-eCAL, a new way to measure energy in multi-agent systems, and it models placement as a joint energy-memory problem. Memory, not bandwidth, becomes the binding constraint. How much KV-cache and working context can you hold cheaply at each tier? That flips the question from "where is the compute?" to "where can the state afford to live?"
Who wins? Arm. If the problem is memory movement and multi-step reasoning, CPU-adjacent, memory-bandwidth-optimized designs across edge and cloud look smart. Arm is selling both paths. AMD is already pushing local execution, arguing that agentic workloads can consume substantial cloud resources — especially across teams. Cloud-only incumbents have the most to lose, though not fatally. High-frequency, low-complexity agent loops might migrate to local silicon, but frontier-scale models still need big iron.
Network operators lose the framing. The paper notes they lack models to determine if distributed agent communication incurs meaningful energy costs. The answer appears to be: not really. That's awkward for anyone pitching energy-aware network slicing as a monetizable service. The 0.25% figure doesn't kill edge networking, but it removes energy as a justification.
This matters because AI data centers could hit 12% of US electricity use by 2030, according to credible estimates. If we're worried about the grid, we should stop obsessing over moving bits and start obsessing over making agents think less. Cutting unnecessary chatter — not relocating computers — is the smart move.
So here's my prediction: trends suggest that by late 2027, we'll see a wave of "agent diet" tools that prune context and reduce redundant reasoning steps, marketed as energy savers. The companies that win won't be the ones with the fastest networks. They'll be the ones that teach agents to shut up and think efficiently.