Quarterly Observation Report, Q3 2026

We, a community based in Taiwan, collect two kinds of observation data every day. The Tor Relay Watch tracks how many volunteer-run Tor (onion routing network) relays operate in Taiwan and nearby regions, and OONI Coverage tracks which networks the censorship measurements of OONI (Open Observatory of Network Interference) come from. Both pages show the last 60 days. Every three months the quarterly report fixes the numbers in place, looks at the quarter as a whole, and adds our reading of them. This is the first report, covering 1 July to 30 September 2026. The glossary at the end explains the fields and how they are calculated.

Highlights

• Taiwan had 11 running Tor relays at the end of the quarter, two fewer than at the start. That is about 0.4 relays per million internet users, the same as South Korea, against 3.2 in Hong Kong and 15 in Singapore
• 8 of the 11 relays run on Chunghwa Telecom's HiNet (AS3462), and the top two networks carry 82% of them, the highest concentration among the ten regions we track
• Two Tor releases appeared in September, and by the end of the quarter seven in ten relays in Taiwan had upgraded, eighth among the ten regions
• OONI recorded about 2.76 million measurements in Taiwan this quarter, and the networks with measurements cover 95% of internet users. Mobile networks serve 45% of users and produced 4% of the measurements
• The community's OONI Run link 10328 sent 39,116 measurements this quarter, 2.2 times the quarter before

Tor relays

Taiwan started the quarter with 13 running relays and ended it with 11. Observed bandwidth fell from 60 MB/s to 57 MB/s, and the number of exit relays held at four. The chart below includes the previous quarter, and the count stayed between 10 and 16 throughout.

Pulse stopped collecting from 8 July to 23 August, so the blank stretch in the chart is missing data rather than relays going offline. 45 of the quarter's 92 days have data. The comparison between the start and the end of the quarter is unaffected, but this report gives no quarterly averages.

Relays that joined and left

Of the 13 relays running at the start of the quarter, 10 were still running at the end and 3 had left, while 1 new relay had joined. 16 relays appeared at some point during the quarter, and 7 of them were running on at least nine in ten of the 45 days with data. With numbers this small, one or two relays coming or going moves the total noticeably.

Upgrades to new releases

Two Tor releases appeared this quarter. Version 0.4.9.12 was first seen among the ten regions on 8 September, and 0.4.9.13 on 23 September. A relay in Taiwan was running 0.4.9.12 on the day it first appeared, and by the end of the quarter 7 of Taiwan's 11 relays ran one of the two. That ranks eighth among the ten regions, while Vietnam and Hong Kong were both above 85%. New Tor releases usually carry security fixes, and operators can subscribe to the Tor Project's tor-announce mailing list to hear about them as soon as they ship.

Network distribution

Taiwan's 11 relays at the end of the quarter ran on four networks: 8 on HiNet, and one each on the Taiwan Academic Network (TANet), Taiwan Digital Streaming and Hoshin Multimedia. The top two networks carry 82% of the relays, the highest of the ten regions, against 23% in Japan. With relays concentrated on one provider, an outage at that provider, or pressure on it, takes most of them down together. Tor relays on campus puts relays on the academic network, which is one way to spread them out.

Comparison with nearby regions

Taiwan has 0.4 relays per million internet users, the same as South Korea and slightly fewer than Japan's 0.5. Hong Kong and Singapore have far fewer internet users, so their relay counts work out higher per person. The Netherlands and Germany host more relays than anywhere else and are shown for reference.

OONI measurement coverage

OONI recorded about 2.76 million measurements in Taiwan this quarter, a fifth more than the quarter before, and the number of networks with measurements rose from 35 to 44. By the estimates of APNIC (the Asia Pacific Network Information Centre), the networks measured cover 95% of internet users, about the same as last quarter. The measurements also became more concentrated, with the top two networks' share of measurements rising from 53% to 65%, mostly because HiNet went from 28% to 44%. The second-largest network accounts for 0.05% of users but a fifth of the measurements, which suggests test devices running automatically around the clock.

By network type

All three mobile networks had measurements this quarter, so their coverage is 100%, but the volume differs widely. Mobile networks serve 45% of users and produced 4.3% of the measurements. Chunghwa Telecom's mobile network alone serves 17% of users, and produced 0.3% of the measurements on 50 days of the quarter. The "other" group serves 2.6% of users yet produced nearly 40% of the measurements, and includes fibre providers with very few users as well as cloud, hosting and VPN networks. OONI assigns a country by IP address, so measurements sent through a VPN exit in Taiwan count toward Taiwan.

Network types are tagged by hand from network names and public information, and only networks with at least 0.1% of users whose type we could confirm are tagged.

The ten networks with the most users

Networks with no measurements this quarter

Nine networks with at least 0.1% of users had no measurements at all this quarter, about 3.4% of users combined. All three ASNs of UNION BROADBAND NETWORK (UBB) are on the list, and one of them, AS24164, had 494 measurements last quarter and none this quarter.

Changes in the networks measured

Of the 18 networks measured this quarter but not last, the cable providers, So-net, Taiwan Fixed Network and National Yang Ming Chiao Tung University are local networks. Most of the rest are hosting and VPN networks.

Daily measurements

The busiest day was 13 August, with 47,042 measurements. That day seven counties and cities in northern Taiwan held a mobile-network throttling drill, and the community asked people to record it with OONI Probe (see the drill results on the docs site).

Of the 2.35 million website tests (Web Connectivity) this quarter, 21 were marked as confirmed blocking, spread across seven domains and five networks, with at most eight on any one domain. Nothing points to blocking concentrated on a particular network. Domains with a high anomaly rate were mostly anomalous on every network, which looks more like the site itself being unreachable. How OONI decides a site is blocked explains how to read such results.

The community OONI Run link

The community's OONI Run link 10328 sent 39,116 measurements this quarter, 2.2 times the 17,518 of the quarter before and 1.4% of all measurements in Taiwan. Nine tenths came from Taiwan Mobile's mobile network, while Chunghwa Telecom's mobile network and Far EasTone together sent fewer than 200.

Events this quarter

• 8 July to 23 August: Pulse stopped collecting for 47 days
• 13 August: mobile-network throttling drill in seven northern counties and cities, the day with the most OONI measurements this quarter
• 8 and 23 September: Tor 0.4.9.12 and 0.4.9.13 first seen in the regions we track

Help needed next quarter

• OONI measurements on mobile networks: Chunghwa Telecom's mobile network and Far EasTone together serve three in ten users and produced under 1% of the measurements. Install OONI Probe on a phone using mobile data and open the community's OONI Run link 10328. The OONI Run v2 guide covers the risks to weigh first
• Cable broadband: UNION BROADBAND NETWORK, VEE TIME and New Kaohsiung Broadband had no measurements all quarter, and a home connection on any of them adds a network
• Tor relays outside HiNet: university networks and other providers both help; see the Tor Relay Setup Guide and Tor relays on campus
• Relay upgrades: if you run a relay, move to 0.4.9.13 or later

Glossary

The dashboards have glossaries of their own covering more fields: Tor Relay Watch and OONI Coverage.

• Tor relay: A server, usually run by a volunteer, that forwards encrypted traffic for Tor users. The report counts only running relays, those listed in the day's last Tor network consensus
• Start and end of the quarter: Pulse records the state of every relay once an hour, and the report uses each day's last record. The start of the quarter is the one from its first day, the end the one from its last day
• Observed bandwidth: The highest rate a relay reports sustaining over the past few days. It is a ceiling on forwarding capacity, and actual traffic is usually lower. MB/s is megabytes per second; multiply by 8 for roughly Mbps
• Exit relay: The last relay in a connection, which reaches ordinary websites on the user's behalf. Websites see the exit's IP address, so exit operators handle more abuse complaints
• Stayed, left, new: Relays are matched by their identity fingerprint at the start and end of the quarter. A relay in both counts as stayed, though it may have paused in between. One only at the start counts as left and may come back later. One only at the end counts as new
• This quarter's releases, n days later: Tor versions first seen in the observed regions this quarter. "n days later" is how many days after the earliest region the release first appeared in that region, which is not necessarily the Tor Project's release date
• Autonomous system (ASN): A group of networks run by one organization under a single routing policy, such as a telecom, a university or a cloud provider. Each has an AS number; Chunghwa Telecom's HiNet is AS3462. A telecom's fixed and mobile networks are often separate ASNs
• TANet: Taiwan Academic Network, which connects schools and research institutions
• Top two networks' share of relays: The share of relays on the two ASNs with the most relays. The higher it is, the more relays depend on a few networks, and the more a problem on those networks affects them
• Relays per million internet users: Relays at the end of the quarter divided by APNIC's estimate of internet users, which counts people online rather than the population. Hong Kong and Singapore have few users and many data centres, so their figures run high
• OONI Probe, measurement: OONI Probe is OONI's open-source app for testing, on a phone or computer, whether websites and services can be reached. Each result for one site or service is one measurement, and a single website test produces hundreds, so the number of measurements is not the number of people testing
• Web Connectivity: OONI Probe's main test. It connects to each website on a list and compares the result with a control connection OONI makes from another network
• OK, anomaly, confirmed, failure: OK means the result matched the control. An anomaly means it did not, which can be blocking but also an unstable connection or a problem with the site itself. Confirmed blocking is an anomaly matching a known blocking pattern, such as a redirect to a block page. A failure is a test that did not complete
• APNIC: The Asia Pacific Network Information Centre, which allocates IP addresses and ASNs in the region and estimates each network's users from sampled measurements in web ads. Estimates cover only the last 60 days, so each report uses the one available when its numbers were produced
• Coverage, users covered: The share of internet users on networks with measurements this quarter. A single measurement in the quarter counts all of a network's users, so whether a network was measured enough shows in its share of measurements and its days. Coverage by network type is the same share within each type. Networks missing from APNIC's estimates, such as some academic networks, count as zero users
• Users, measurements (%), days: Users is the network's share of internet users, measurements its share of all measurements this quarter. The further apart they are, the less the measurements match the number of users. Days counts the days with measurements out of the quarter's 92; on the dashboard it is out of the last 30
• Top two networks' share of measurements: The share of measurements from the two ASNs with the most measurements. The higher it is, the more measurements come from a few networks. It is a different figure from the relay share in the Tor section
• pp (percentage points): The difference between two percentages. The top two networks' share of measurements rising from 53% to 65% is written +12 pp
• Network type: Networks are labelled by hand as mobile, fixed and cable broadband, or academic, from their names and public information. "Other" covers fibre providers with few users, cloud, hosting and VPNs. OONI assigns a country by IP address, so measurements sent through a VPN exit in Taiwan count as Taiwan, which is why "other" has so many
• OONI Run link: A link that tells OONI Probe what to test; opening it makes the app run the list it contains. The community's link is 10328, and the report counts the measurements sent through it separately

Data and method

• The period is 1 July to 30 September 2026 (UTC), compared with 1 April to 30 June
• Tor relay data comes from Pulse, run by the community, which collects from Onionoo, part of Tor Metrics, every hour, and each day uses its last snapshot. The start and end of the quarter are the snapshots of 1 July and 30 September, and "running on nine in ten days" counts only the 45 days with data
• The snapshot of 7 April captured only a few relays and is treated as missing data
• A release's first-seen date is the first day it appeared in any of the ten regions, which is not necessarily the Tor Project's release date
• OONI figures come from OONI's aggregation API and include every test type
• Users per network are APNIC's estimates as of 8 October 2026. APNIC only publishes its latest estimate, so this quarter and the previous one use the same figures, and changes in coverage reflect only changes in which networks were measured
• The network type tags are in data/reports/asn-types.toml, and the raw numbers behind every chart in this report are in data/reports/2026-q3.json