One agent task, 1,000 calls, identical hardware. The only difference is where the GPUs sit.
The task: 20 turns, 50 calls each; every call, model or tool result, is one round trip. And both trends point the same way: agents keep getting longer (METR, 2025: the length of tasks AI can complete doubles about every 7 months) while compute keeps getting cheaper (Stanford HAI, 2025: inference cost fell about 280x in 18 months). Compute shrinks. The wire does not.
The task above is 1,000 round trips:20 agent turns, 50 tool or model calls each. Every call pays the same wire cost no matter how fast the GPU answers,70 ms for a GPU 2,400 miles away, 3 ms for one across town. We dial compute per call from 750 ms down to 0 and hold turns, calls, and hardware fixed, so distance is the only variable left on the clock.
7 months
The length of task an AI agent can complete on its own doubles about every 7 months. Longer tasks mean more calls, not fewer.
280×
Inference cost fell about 280 times in 18 months. Compute keeps getting cheaper. The round trip does not.
Perimeter Compute retrofits Class A real estate into a metro-scale inference network, so agent workloads run across town from the calls that trigger them instead of across the country. The same control plane that routes work on congestion, hardware fit, and grid pricing keeps the wire short too.
Partner with us to launch a pilot Inference Hub in your market and get early compute access as we expand city by city.
We look forward to building the future of metro-edge compute with you.