The sequence
Warning emails as you get close. When the balance covers about 72 hours of running cost, then 24, then 6, you get an email at each step. Three at most, not a daily campaign. If you top up, the steps reset and you will be warned again the next time you get close.
The hour it runs dry, the server stops. Billing is hourly, so the moment there is not enough balance to pay for the next hour, the server is suspended rather than charged into a negative balance. It is powered off, not deleted.
Then it is archived, and this is the part that matters. Instead of destroying the disk, we take a full snapshot of it and then release the instance. The data is kept; the machine is not, because a stopped instance still occupies hardware we pay for.
The snapshot is held for seven days. Top up during that window and the server is rebuilt from the snapshot automatically, no ticket, no button, no asking. The rebuild starts as soon as the balance covers an hour.
After seven days the snapshot is deleted and the server is gone for real. This is the hard deadline and there is no recovering past it.
The part that is not in your favour
Three honest caveats, because a page like this is worthless without them.
GPU and bare metal servers are not archived. A snapshot of a GPU disk would cost more to store than the server earns, and bare metal has no snapshots at all. When the balance runs out on one of these, it is deleted and its data is not kept. The warning emails say so; if you run one, keep headroom.
You get a new IP address. The restore builds a fresh instance from your snapshot, and a fresh instance gets a fresh IPv4. Your files, packages and configuration come back exactly as they were; the address does not. If anything points at that IP, DNS, a firewall allowlist somewhere else, a licence tied to the address, you will need to update it.
Seven days is seven days. The snapshot is deleted on schedule, not when we get round to it. If you are going to be away, top up before you go rather than planning to fix it later.
And the policy behind all of this, stated plainly: top-ups are final. Balance is prepayment for hosting here. It does not expire and it sits on the account until you spend it, but it is not refundable and cannot be paid back out. That is in the terms, and it is the reason the sensible pattern is to fund what you are about to use rather than months ahead.
Why it is built this way
Hourly billing cuts both ways. It means you are never paying for a server you stopped needing, and it means the balance drains continuously rather than on a date you remember. Something has to catch the gap between those two facts, and deleting people's work is the lazy version.
Snapshotting instead costs us the storage for a week and costs you nothing. It also means the common real-world case, you meant to top up, life happened, you came back on Thursday, ends with a working server rather than a support ticket about data you cannot get back.
What this does not replace
Snapshots on suspension are a safety net for the billing edge case. They are not a backup strategy.
They are taken when you run out of money, which is not when you need a restore point. The thing that saves you from a bad deploy or a rm -rf in the wrong directory is your own backups, taken on your own schedule, stored somewhere that is not the server. There is a separate walkthrough for doing that properly with restic or borg, and it is worth the hour.