GPU Routing

Latency

The network’s round trip from across the US to the nearest of our seven GPU regions, measured place by place.

About 30 ms or less from most of the US, within 50 ms from almost everywhere

Last measured November 17, 2026, from 76 places in the US

Round trip to the nearest region

  • 30 ms or less54
  • 30–50 ms18
  • Over 50 ms4
  • Every place, fastest first

Twice a day, at noon and midnight US Eastern, public probes across the US time a TCP handshake to each of our seven regions. Over the 30 days to November 17, 2026 they measured from 76 places, 31 in data centers and 45 on home and office broadband. Each place is colored by its median round trip to the nearest region that can start a GPU now. It is the network’s round trip alone: no data is sent and no inference runs, and a policy’s own time on its GPU comes on top. Read how a robot is routed