World map of NTP and NTS: coverage, operators, concentration

This map does not say what the state of NTP is in the world. It says what we measured of NTP in the world — not the same thing, and the difference is written wherever it matters. Every figure comes from a dated campaign, probed from six networks, and can be recomputed from the published CSV.

This page is a macro view: the world, the continents, concentration. It does not replace the detail — each server's name, the gap between what an operator documents and what it actually serves, the campaign history live on the continent pages. It recommends no server either: to choose a French source and configure it, go to the France guide.

The world, continent by continent

The queued column counts servers already in our inventory that we have not probed yet. It exists so that a zero row stays readable: a continent can show zero measured machines because we have not gone there yet, which has nothing to do with an absence of servers. Today 0 machines in 0 countries are waiting for a campaign.

These rows are aggregates of the same measurements. "World" includes global anycast and machines whose location we could not establish — which therefore belong to no continent.

2 entire continents have no measured authentication endpoint at all: Oceania and Africa — 175 machines serving time without a single one able to prove it is the server it claims to be. This is not a gap in our measurement: those machines answer, we reached them, and none of them serves NTS.

Stratum distribution

A country with forty servers all at stratum 3 depends on elsewhere; a country with five stratum 1 holds its own time reference. That is what this count shows, and nothing else — a stratum is a logical distance, not a quality.

France first

We are French, the question concerns us, and the answer is not flattering. Out of 267 French machines measured, 23 serve NTS — and 43.5 % of those are ours.

This section says what we measured in France. It does not say which source to use: for that there is the guide to French NTP servers — choice by use case, access policies, chrony and w32tm examples.

Country by country

The four counts are kept separate, never merged: a server answering only over IPv4 and a server answering only over IPv6 are not the same fact, and that is exactly what a coverage map must show.

This table gives the volumes. It does not give the names, nor what each operator documents versus what it actually serves, nor what has moved since the previous campaign: that takes prose, country by country, and lives on the continent page. The right-hand column points there when it exists.

The pool, stratum by stratum

This map counts servers published as time services. Alongside them lives another, far larger world: the machines volunteers lend to pool.ntp.org. They do not join the registry — “who publishes a service” and “who lends a machine” are two questions, and mixing them would sink the NTS share for a reason that has nothing to do with NTS. But their stratum breakdown says something the registry does not: what everyone's time actually rests on.

3578 machines lent to the pool, 3473 answered, 1106 reachable over both IPv4 and IPv6. Machines are counted once even when we know both of their addresses. Harvest of 2026-08-05. A pool zone does not mean the country: the project tops up a sparse national zone with servers from neighbouring countries, so a zone says where the operator registered, not where the machine sits.

Full data, per continent and per zone: pool-strata.json (CC BY 4.0).

The method, and what it does not say

Is your server missing?

Declare it in our open repository — one YAML file, a single required field, the hostname — or by e-mail to ntp@rdem-systems.com. It enters the next campaign — and that is how a "queued" row in the table above eventually becomes a measurement. That repository is not a directory of recommended servers: an operator's declaration is worth existence and permission to measure, not a recommendation.

The data

What reproduces this page exactly are the sealed files of the campaign shown — signature, two authority timestamps and OpenTimestamps. Count the rows and the total lands exactly — but each file has its own spelling, and a recipe that does not say which one cannot be replayed:

• CSV — counted=yes and nts_measured=yes (yes/no values);
• JSON — counted and nts (true/false booleans).

Both return the same set of machines: it is the same campaign, projected twice.

Two caveats, so the check succeeds first time. The counts header of that same file is not obtained that way: it applies the carry-forward rule, which keeps the machines the campaign did not probe with their previous reading, and so states 3,709 machines and 200 NTS where the rows give 3,680 and 198 — see the erratum. And the files offered below are those of the campaign you are reading, not a “latest known” URL: their names change with every campaign, and that is what makes them citable.

All these files are published under CC BY 4.0:

• the sealed measurement of 7 August 2026 (JSON) — one row per server; this is the file that reproduces this page, and its seals sit beside it
• the inventory of the same campaign (CSV) — sealed as well
• the current inventory (CSV) — a rolling URL, replaced at every campaign: handy for following, unfit for citing. The aggregates no longer have a rolling URL: each campaign carries its own, under its own name.
• the time series (JSON) — each point carries the method verdict in force on that day

The pages that take over from this one:

• Analyse the European measurements — who serves NTS in Europe, under which name, and what the documentation does not say.
• Choose a French source — the decision and configuration guide, not a count.
• Understand NTS — what RFC 8915 actually authenticates.
• Walk up the dependency tree — which server follows which, and on which clock. Clock concentration is not operator concentration.