Fastest VPS location for visitors in mainland China: 10 rounds of pings from China Telecom, Unicom and Mobile home lines (9–14 October 2026)

From home broadband in mainland China, only six of the 53 data-centre test hosts we ping stayed under about 115 ms on China Telecom, China Unicom and China Mobile alike: Tencent Cloud Singapore, Tokyo and Hong Kong, and AWS Lightsail Tokyo, Seoul and Singapore. Several popular budget picks were fast on one carrier and two to five times slower on another. In the evening, packet loss across all hosts more than doubled.

These numbers come from our own latency map, which pings every test host twice a day from probes inside China, Hong Kong, Tokyo, Singapore and other cities through the free Globalping network. The map shows the latest round. This note pools ten rounds and keeps only the probes on home carrier networks. Most of our city probes in China sit in cloud data centres, and they flatter the result.

What we measured

• Window: 10 rounds from 9 Oct 2026 20:02 to 14 Oct 2026 12:09 China Standard Time (UTC+8). Morning rounds ran roughly 08:00–13:10 and evening rounds 20:00–02:10 (CST).
• Home-line probes: China Telecom (AS4134) in Dongguan and Guangzhou, plus one probe in Nanjing; China Unicom (AS4837) in Changsha, Guilin and Nanning; China Mobile (AS9808) in Beijing, Kunming, Nanning and Taishan. Globalping hands out whichever probe is online, so the set rotates between rounds. The Nanjing probe is flagged as a data centre.
• Tests: 1,576 home-line tests, of which 1,329 finished. The rest found no probe online or failed. Each test sends 10 packets. We report the median round-trip time per carrier and the average packet loss.
• Targets: the providers' own public test hosts (looking-glass, speed-test, object-storage or API endpoints in each region), listed with sources on the latency page. Lightsail hosts are measured with TCP connects to port 443, everything else with ICMP ping.

Each carrier cell below rests on 3 to 10 tests from a handful of probes. Read it as a measured snapshot for one week, not as a promise about your readers' lines.

The six that held up on every carrier

Median round-trip time in milliseconds, per carrier, 9–14 Oct 2026. Sorted by the slowest of the three carriers.

The next host on the same "slowest carrier" ranking was Vultr Tokyo at 184 ms, so the gap after these six is large.

The Lightsail zeros need care. A TCP handshake retries a lost packet, which shows up as a slower connect rather than as loss, so 0% there is not directly comparable with ICMP loss. The ranking by time still holds.

The same city, three different answers

Singapore and Tokyo are the usual advice for a site with readers in China. Our home-line data shows why that advice works for some people and fails for others: the carrier mattered more than the city.

Loss is shown where it averaged 10% or more. A reader on China Mobile would call DigitalOcean or Linode Singapore fast. A reader on Unicom would see the same server at five times the delay.

The traceroutes we run once a week point to the routes behind this:

• China Mobile: all 54 home-line traces in weeks 41 and 42 left through CMI (AS58453), China Mobile's international network. That fits Mobile's low numbers to Singapore and Tokyo across providers.
• China Unicom → Tencent Cloud: all eight traces to Tencent's Hong Kong, Tokyo, Seoul and Singapore hosts went through AS10099/AS9929, Unicom's premium international routes. Traces to the other providers mostly did not.
• China Telecom: we saw no CN2 (AS4809) path in either week. Only 7 of 42 Telecom traces could be labelled as the ordinary 163 backbone; the rest crossed networks we could not classify.

Route labels come from the AS numbers in each trace and are a hint, not proof of what a provider sells. A test host can also sit on a different network segment from the virtual machines you rent.

Evenings are worse, and loss is the bigger problem

Splitting all 1,329 finished home-line tests by round:

The median moved by about 14 ms. Loss more than doubled, and nearly half of the evening tests dropped at least one packet. Even the top six slowed in the evening: Tencent Cloud Singapore went from a 60 ms morning median to 92 ms, and Lightsail Singapore from 68 to 90 ms. Two hosts stood out for loss: OVHcloud Singapore lost 37.6% of packets on average across all home-line tests, and Hetzner Singapore 19.3%. A web page survives some loss but slows down, and SSH sessions and game servers feel it at once.

One result we do not trust

Alibaba Cloud's Hong Kong storage endpoint answered in about 1 ms from a Hong Kong probe, yet took 249–356 ms from all three carriers' home lines. Hong Kong is next door, so a long detour or ICMP handling at that endpoint is more likely than a slow region. We treat it as a quirk of that test host, not a verdict on Alibaba Cloud Hong Kong servers. It is a good example of why one test host should not decide a purchase.

What the fast regions cost

Entry prices for the regions above, from our price tracker on 14 Oct 2026 (provider list prices, before tax, monthly):

Tencent Cloud and Alibaba Cloud are not in our price tracker, so we do not quote them. Lightsail's half-transfer rule for some Asia-Pacific regions does not apply to Tokyo, Seoul or Singapore; we covered that and other regional allowances in the 7 October note on sites with readers in Asia. The prices page shows every tracked plan. None of these list prices changed in our tracker between 9 and 14 October.

How to use this before you buy

1. Find out which carrier your readers use. If you serve one province or one company network, that carrier's column matters more than the average.
2. Readers on all three carriers: start with the six regions in the first table. Tokyo or Singapore on Lightsail, or Tencent Cloud's overseas regions, held up across carriers this week.
3. Mostly China Mobile readers: DigitalOcean and Linode in Singapore and Tokyo were among the fastest options, at $4–$5 a month.
4. Check the evening. Re-open the latency map after 20:00 China time, and weigh the loss column as much as the milliseconds.
5. Test the server you actually rent. Most providers bill hourly, so you can spin up the region, ping it from your readers' networks, and delete it the same day.

When you move a site to a new region, two of our free tools help with the switch. DNS Lookup queries your domain's A and AAAA records over DNS-over-HTTPS and shows the TTL, so you know when the old address stops being served. IP Address Tools shows the public address your own traffic leaves from, which helps you confirm a VPN or proxy is off before you run your own tests, and includes a CIDR calculator for firewall rules.

Limits

• A small probe pool. 3–4 home probes per carrier, in a few southern and central cities plus Beijing. Shanghai, Chengdu or northern home lines may look different.
• One week. Routes in and out of China change, and premium routes such as CN2 GIA, CMI or 9929 are often sold as separate products at higher prices. Our test hosts sit on whatever network each provider uses for its standard service.
• Ping is not page speed. Throughput, TLS, a CDN and filtering at the border all affect what a reader sees. For a content site, a CDN in front of the server can matter more than the region.
• Not a ranking of providers. These are test hosts in specific regions. A provider's slow region here says nothing about its other regions.