Is the Nostr Relay Network Shrinking? What Two Years of Monitoring Data Shows
We counted Nostr relays using NIP-66 monitoring data. There is no sign the network is shrinking, but churn is heavy: over half of 2024’s relays have gone quiet.
Every so often someone asks whether Nostr is dying, and the evidence offered is usually a relay that has stopped working. We have our own example. In September we removed three dead relays from our relay directory, and one of them had been unreachable for about seven months while still listed. One relay going dark tells you very little about the network, so we went looking for a population-level answer.
The short answer: no sign of shrinking. Five of nostr.watch’s long-running monitors each counted 786 or 787 relays in July 2024. On 23 September 2026 the same five reached between 1,119 and 1,139. But the same data cannot tell a new relay from a newly discovered one, so the honest claim is “no sign of shrinking”, not “growing”. What it shows very clearly is churn: only 32% of the relays answering in July 2024 were still answering on that date.
How relays are monitored
NIP-66 lets anyone run a relay monitor: a program that connects to relays on a schedule and publishes what it finds as ordinary signed Nostr events (kind 30166). A report says, in effect, “I reached this relay at this time, it took this long to answer, and here is what it says it supports”. The largest fleet of monitors is run by nostr.watch, from eight locations on six continents, from Johannesburg to Seoul. Most of them check every relay on their list every four hours.
A monitor only publishes a report when it actually reaches a relay. So a recent report means a relay is answering, and a missing one means the monitor has not reached it lately, or has stopped trying.
The catch: the data remembers departures, not arrivals
NIP-66 reports are addressable events. Each monitor keeps exactly one report per relay, and every new report overwrites the last. That has an awkward consequence for anyone trying to measure a trend.
- A relay that answers on a given day has a report dated that day.
- A relay that stopped answering in March 2025 still has a report, dated March 2025: the last time it answered.
- A relay that first appeared last month looks exactly like one that has been running since 2023. Nothing records when a relay arrived.
So the monitoring data can count the relays answering now, and it can date when relays went quiet. It cannot count arrivals, which means it cannot on its own tell you whether the network grew or shrank.
There is one lucky exception. Until the end of July 2024 nostr.watch published its reports under an older draft event kind, then switched to the current one. Nothing overwrites the old kind any more, so each monitor’s final round of old-style reports is frozen: a count of the relays it could reach in July 2024. We checked how complete that count is. Six monitors on five continents each reached 786 or 787 relays in their final round, and any two of them share at least 98% of their lists, which is what a full sweep looks like. Of the 52 relays whose last contact came in the five weeks after the switch, and so were very probably answering at the time, 48 (92%) appear in it. Both checks are against nostr.watch’s own list: a relay it did not know about is in neither count.
What we counted
Of the 609 accounts that have published relay reports, 24 report on a hundred relays or more; 537 of the rest report on only one or two. Of those 24, 8 were reporting both before August 2024 and on 23 September 2026, and all of them are nostr.watch’s. Reporting at both ends is not the same as reporting throughout: three of them have long holes in their record, Sydney with reports 204 days apart (29 March 2025 to 19 October 2025), New York with reports 202 days apart (29 March 2025 to 17 October 2025) and Johannesburg with reports 116 days apart (11 November 2025 to 7 March 2026). Leaving those three out moves none of the monthly averages below by as much as one relay.
On 23 September 2026, 1,207 relay addresses answered at least one of those 8 monitors in the latest four-hour round. They ran on 1,083 distinct servers (some servers answer on many addresses). Counting every active monitor over the 24 hours to 15:51 UTC that day gave 1,444 addresses on 1,250 servers.
The fairest comparison with 2024 is monitor by monitor, one round against one round:
| Monitor | July 2024 (final old-style round) | September 2026 (latest round) | Change |
|---|---|---|---|
| Johannesburg | 787 | 1,124 | +43% |
| Silicon Valley | 786 | 1,119 | +42% |
| São Paulo | 786 | 1,131 | +44% |
| New York | 786 | 1,121 | +43% |
| India | 786 | 1,139 | +45% |
| Netherlands (checks hourly) | 787 | 700 | −11% |
| Sydney (2024 round cut short on 23 July) | 745 | 1,126 | +51% |
| Seoul (2024 round cut short on 23 July) | 542 | 1,124 | +107% |
Five monitors with a complete final round in 2024 each reached between 42% and 45% more relays on 23 September 2026. The exception is the monitor that checks every hour, which reached 11% fewer. That fall looks like its own watch list rather than the network. Of the 1,198 relays then answering the five four-hourly monitors, 387 had never been reported by the hourly one, and it had stopped reporting 111 more that the others still reached.
Heavy churn
In July 2024, 805 relay addresses answered at least one of the long-running monitors. On 23 September 2026:
- 257 (32%) still answered a monitor.
- 450 (56%) had gone quiet: no monitor had reached them for more than 30 days, and their last contact was not in one of the bursts described below.
- 81 (10%) were last reached during one of the bursts of simultaneous departures described below, when the monitors’ own lists appear to have changed, so their fate is unknown.
- 17 had gone silent in the month before that date, too recently to call.
Counting servers rather than addresses gives much the same picture: 252 of 730 servers (35%) still answered, and 53% had gone quiet outside the bursts.
The rate at which relays go quiet has gone up. Counting servers, the long-running monitors saw an average of 56 a month go quiet from March to December 2025 (between 35 and 70) and 96 a month in the first seven months of 2026 (between 75 and 117). January and February 2025 are left out because the record has gaps in them (see below); counting them, the 2025 average is 53. That is not evidence of decline on its own. With no record of arrivals, a larger population turning over faster looks just like a network losing ground, and on 23 September 2026 the monitors saw more relays, not fewer.
Things that look like collapse and are not
A monitor coming and going. One large monitor reported from 15 March 2026 to 15 June 2026 and then stopped. It reported on 6,370 relay addresses, 4,280 of which no other monitor ever reported, and it stopped reporting most of those while it was still running: only 335 were last reported on its final day. Counted naively, relay addresses going quiet in April, May and June 2026 were 2,183, 1,847 and 956. Restricted to the monitors reporting since before August 2024 and still reporting, they were 167, 83 and 95.
Housekeeping. Between 12:05 and 14:22 UTC on 12 April 2026, 852 relay addresses were last reported. Independent relays do not fail together like that. These addresses sat on just 111 servers, and 81 of those servers were still answering on other addresses on 23 September 2026, which points to the monitors’ own list being cleaned up.
Not every burst is so clear. Between 16:33 and 17:27 UTC on 17 September 2024, 114 relays on 114 different servers were last reported, and none of those servers answered on 23 September 2026. All 114 were last reached by the hourly monitor. The four-hourly monitors had not reported any of them since the July 2024 switch, although 74 had been in those monitors’ July count and the hourly monitor was still reaching them. So the four-hourly monitors’ lists dropped at least 74 relays that were still running, and the hourly monitor’s list seems to have caught up in September. Whether they were still running by then, the data cannot say.
We set aside every burst like these, rather than count it as ordinary churn: 3 bursts, holding 1,009 addresses on 245 servers.
What this cannot tell you
- Whether relays were born or found. The count rose, but it counts what the monitors know about. More relays and better discovery look identical.
- Whether anyone uses them. A personal test relay counts the same as a busy public one. NIP-66 says nothing about users or traffic, which are better measures of health than a relay count.
- Anything independent of nostr.watch. Every long-running monitor is nostr.watch’s. Several monitors agreeing is several vantage points, not several opinions.
- When relays went quiet in late 2024. Almost all of the history before 2026 survives on a single relay that stores these reports, and it has holes: between 6 October 2024 and 4 February 2025 there are 7 stretches with no report from any long-running monitor, each between reports a week to a month apart, the longest between reports 31 days apart. Relays that went quiet in those months cannot be dated to the month.
- Exact figures. The 2024 and 2026 counts come from the same monitors but different versions of their software, and their watch lists changed at the switch itself (see the September 2024 burst above).
What it means if you use Nostr
Relays come and go, and that is how Nostr is meant to work. Your identity does not live on any one relay, which is why clients spread your posts across several of them and publish a relay list so others know where to find you. Given that more than half of the relays answering in July 2024 have since gone quiet, it is worth checking that list now and then, and not relying on a single relay you set up once and forgot. Our guide to how relays work covers the basics, and our relay roundup is a place to start.
It also means relay directories go stale quickly, ours included. We now check our relay list against the same monitoring data used here.
How we measured
On 23 September 2026 we collected every relay report we could find from five relays that store NIP-66 data, keeping the latest report from each monitor about each relay: 55,142 in all. Every figure dated 23 September 2026 counts what the monitors reported in one of two windows that end at 15:51 UTC that day, when we collected the reports: their latest four-hour round, or the 24 hours to that moment. We normalised addresses (monitors disagree about trailing slashes), counted both addresses and servers, treated a relay as gone after 30 days without a report, and compared all monitors against only those reporting since before August 2024 and still reporting. Because the monitors overwrite their own history, each count we keep becomes part of the record: repeating it over time is what would finally show arrivals as well as departures.