The figure is a rate, not a total
Monthly transfer caps punish the shape of your traffic rather than its volume. We publish a sustained rate instead, measured as the ninety-fifth percentile of five-minute samples across the calendar month, on whichever direction is busier.
| Port | Fair use, 95th percentile | Equivalent over a month |
|---|---|---|
| 10 Gbit/s | 1 Gbit/s | about 324 TB |
| 25 Gbit/s | 2.5 Gbit/s | about 810 TB |
| 40 Gbit/s | 5 Gbit/s | about 1.6 PB |
The percentile is the part that matters. Discarding the top five percent of samples discards thirty-six hours of every thirty-day month, so a release day at line rate, a migration, or a fortnight of nightly backups running flat out never register. Sustained saturation does.
Measuring your own side
cat /sys/class/net/enp1s0/statistics/rx_bytes /sys/class/net/enp1s0/statistics/tx_bytesFor anything beyond a spot check, install vnstat and let it keep the history:
apt install -y vnstat
systemctl enable --now vnstat
vnstat -i enp1s0 -mOur numbers and yours will differ by a percent or two. We count at the switch port, which includes framing your kernel never sees.
Crossing it
You get an email. Not an invoice, not a throttle, not a suspension: an email from someone who has looked at the graph and can say which direction is saturated and since when. From there the answer is usually one of three things.
- A dedicated 40 Gbit/s port, available as an add-on on EPYC, storage and metal plans.
- A storage node, which comes with a wider port and is cheaper per terabyte if the traffic is bulk.
- A second instance somewhere else, if the traffic is one-directional and cacheable.
Bandwidth on its own has never been a reason to suspend an account. Ignoring six weeks of email about it eventually would be.
What does not count
- Attack traffic dropped at the edge. It never reaches your port, and it was never yours.
- Traffic between two of your own instances at the same site, over the internal network.
- Copies we make on your behalf: site migrations, off-node backup runs, snapshot writes.
Ports are dedicated
The figure applies to your port rather than to a share of the rack. Nobody else’s Friday evening changes your Tuesday. Sit at line rate for an hour and the only consequence is a graph that is flat along the top.
Why not simply meter it
Metered billing makes a hosting bill unpredictable in exactly the month you can least afford a surprise, and it turns every traffic spike into a conversation about money. A published rate is a number you can design against. That is the whole argument.
Reading the graph
Five-minute samples, ninety days of history, both directions, with the percentile line drawn across it. The raw series, if you would rather plot it yourself:
curl -s "https://paragonvps.com/api/v1/instances/<instance-id>/traffic?period=month" -H "Authorization: Bearer $Paragon_TOKEN" | jq .