Time servers in Europe: who actually runs NTS?

Everyone copies lists of time servers. Nobody measures them. We probed 2935 European sources from 6 different networks, and verified the one thing documentation asserts without proving: who actually serves NTS, the authenticated flavour of NTP (RFC 8915).

The result is harsh. Across the 40 European countries covered, 22 serve NTS. In the other 18, no public reference we could find serves it: Albania, Bosnia and Herzegovina, Bulgaria, Cyprus, Estonia, Greece, Hungary, Iceland, Latvia, Liechtenstein, Lithuania, Monaco, Montenegro, North Macedonia, Portugal, Serbia, Slovakia, Ukraine.

Three questions, three measured figures

The method, because it changes the result

Six vantage points, six autonomous systems

A measurement from a single network lies: an access list, a rate limit or a geographic restriction produces a false negative. So we probed from six distinct ASes:

• RDEM Systems — AS206014: our AS, the only one we operate.
• Free — AS12322: our office ISP. We are a customer; we do not run it.
• Arelion — AS1299: international transit provider.
• IELO — AS29075: French transit provider.
• IPSET — AS199275: French transit provider.
• Contabo — AS51167: German host.

Five of those six networks are not ours. AS206014 is the only one we operate. The other five belong to third parties we do not control: our office ISP (Free), three transit providers (Arelion, IELO, IPSET) and one host (Contabo). A measurement taken only from one's own network would prove very little; it is the convergence of six independent networks that makes the figure defensible.

Across the 5443 sources in this campaign, the 6 networks are unanimous. No disagreement: the vantage_disagreements field in the published file is empty, and that is checkable.

That unanimity is not guaranteed, and we do not present it as such. Vantage points can diverge — typically on anycast addresses, where you do not land on the same machine depending on the network you start from. When it happens, the gap is published in the file, never smoothed away. That is precisely what a single-vantage measurement cannot see.

The rule applied: a server counts as serving NTS if it negotiates from at least four networks out of six. Every row of the dataset carries its exact fraction (nts_from, reachable_from) — so a reader can redo the count with a different threshold than ours.

NTS in Europe, country by country

How to read "none found". It does not mean a country has no NTS server. It means that no public reference we could find and measure serves one, as of 31 July 2026, from our six observation networks. A source that is unpublished, reserved to a national network or filtered by ACL stays invisible to us. If you know of one, write to ntp@rdem-systems.com: we measure it and we correct this page.

Who runs time in the 18 countries where we found no NTS

The table above says a country serves no NTS. It does not say who runs time there — which is the question people actually ask: "Italian NTP, where does it come from?". So here is the metrology institute, research network or operator actually measured — with its real stratum, and an explicit mention of the names that no longer respond.

These countries are not black holes: most have a national institute serving perfectly decent stratum 1. What they lack is NTS — that is, any way for their clients to verify who they are talking to.

8 of these countries have no entry here yet — Albania, Bosnia and Herzegovina, Bulgaria, Cyprus, Liechtenstein, Monaco, Montenegro, North Macedonia. We have measured servers there, but we have not yet done the work of naming who runs them. Saying so is more useful than letting the table look complete.

What this table means: when none of the sources found in a country responds, that is not "no NTS", it is "no reachable public time server from our 6 networks". If you know of one, tell us: this is exactly the kind of gap we would rather see closed by a correction than defended on principle.

The measured inventory, server by server

This table is the inventory itself, not an extract: each row is a machine we queried from six distinct networks, and each column states what was obtained — never what is claimed.

The state has four values because two are not enough. A server can fail to answer us and still serve time: if other servers publish it as their upstream, that is a closed access list, not an outage. And an isolated silence is not a disappearance: it takes three silent campaigns and four months before we write “dead”.