Here's an edge case, measured. The question: does a cloud datacenter back up faster than home broadband — and does the region closest to the storage win? You can't reason your way to the answer; the map lies. So I raced it with disposable instances and let the network tell me where it actually goes.
Spin up throwaway instances in three regions — Newark, Seattle, Fremont — plus a home box on 1 Gb symmetric fiber as a fair baseline. Run the same benchmark on each: one 128 MiB incompressible payload, written then read back over 8 parallel streams, wall-clock timed and integrity-checked. The mover is the everyday backup toolkit — rclone, rsync and restic, speaking S3 and WebDAV — driven at high concurrency, because a real backup saturates the link with many parallel uploads, and it's aggregate throughput, not any single stream's latency, that picks the winner. Two object stores as targets. Then every box destroys itself. Launch → measure → tear down — the whole sweep costs pennies.
| region | result |
|---|---|
| Newark · us-east | fastest to both stores |
| Fremont · us-west | middle of the pack |
| Seattle · us-west | slowest datacenter |
| Home · 1 Gb fiber | crushed at first; on wired ethernet, 2nd of four on read |
The store nearest Seattle lives on the US west coast. Seattle should have won on distance. It came last.
Geographic distance didn't predict throughput. The transit path — backbone, peering, hop count — did. The closest region on the map was the slowest on the wire.
Network paths are an empirical question. Disposable instances turn "guess the best region" into a measurement you run in minutes and throw away.
Home on gigabit fiber isn't a strawman — it held its own on restore even as datacenters won the upload. "Cloud vs home" isn't one answer; it's a per-leg map.
None of this is worth doing if a vantage costs a day and a support ticket to stand up. It's worth doing because a region has become a disposable primitive. On Linode a box boots from a saved disk image in well under a minute — the same image every time, so every vantage runs the identical harness and the only variable left is the network. You spin it up, assess the number, stay in control of exactly what's running, then rip the box out — no residue, no monthly line item, no attachment to a machine you were always going to throw away.
spin up → assess → control → rip out → repeat
What makes the loop fast is the part you don't see: provisioning that's quick enough the box is ready before you've finished the command, an image workflow simple enough the whole sweep is a short script, and — the reason the map lost to the wire — lightning-fast peering between cloud providers. That last one is the whole game here: the disposable box isn't the destination, it's a well-connected vantage point onto the paths between clouds, and a provider that peers hard is a provider that lets you see the real network instead of a congested on-ramp. The disk image makes each run reproducible; the per-region, per-minute footprint makes running it everywhere cost pennies. Cheap enough to measure means cheap enough to keep measuring — this sweep is one cron line away from a standing sensor.