Reaching users in mainland China from a Server Room server

China Telecom sells two ways into its network from abroad. The ordinary one is ChinaNet, the “163” backbone (AS4134), which every international carrier reaches. The premium one is CN2 (AS4809), and CN2 GIA is the grade that keeps both directions of the journey on it. It costs a hosting company several times what ordinary transit does, and it is what “optimised for China” means on other hosts' pages.

Our traffic into China Telecom crosses Cogent and enters ChinaNet. We do not buy CN2 in either direction, so neither the way in nor the way back is CN2 GIA. We checked this on 27 September 2026 from our routers in New York and Bucharest, and with traceroutes from San Francisco, Miami and Amsterdam: every path enters China Telecom on 202.97.x.x, the ChinaNet backbone, and none crosses 59.43.x.x, which is CN2.

What that means in practice: the route works, and at quiet hours its latency is simply the distance. During China's evening peak, ordinary international links into the country congest, so expect higher latency and some packet loss then. Avoiding that congestion is exactly what CN2 GIA is sold for. If your users need it avoided every evening without exception, we are not the right host for that part of your service, and we would rather say so now.

Distance is the part of latency nobody can buy away, and San Francisco is the closest of our five datacenters to China: traffic crosses the Pacific directly instead of crossing the United States first. These are the round trips we measured from each datacenter on 27 September 2026, at 17:30 Beijing time, averaged over five pings.

That is one afternoon and one address on each network, so trust the order more than the exact figures. Two things stand out. China Mobile answered Amsterdam faster than San Francisco that day, so if most of your users are on China Mobile, test both. And China Unicom's addresses did not answer pings from abroad at all, which is why it has no column: for Unicom, the test in the next section is the only honest measurement.

The direction your users feel is the way back: from their line in China to your server. That is the half CN2 GIA is about, and it is the half our own tools cannot measure, because it is chosen inside China. So test it from where your users are, on the carrier they use, at the hour they are busiest.

Every datacenter has a public speed test, and its address is a real server in that datacenter's network. Open it in a browser on a China Telecom, Unicom or Mobile connection, or run ping and traceroute against it from a machine there (on Windows the second command is tracert):

• San Francisco: speedtest-sf.serverroom.net
• New York: speedtest-ny.serverroom.net
• Miami: speedtest-mia.serverroom.net
• Amsterdam: speedtest-ams.serverroom.net
• Bucharest: speedtest-buc.serverroom.net

ping speedtest-sf.serverroom.net
traceroute speedtest-sf.serverroom.net

For the other direction, the looking glass runs ping, traceroute and BGP lookups from our New York and Bucharest routers toward any address you give it, including one in China.

Then test with a real machine. Order a ready server, run your own workload through your users' evening, and if the route does not suit you, ask for a refund within three days of activation. The refund is pro-rated, less a 15% fee on the refunded amount; the full terms are in our refund policy. Three days of real traffic from real users beats any figure on this page.

A server that serves China will sooner or later meet the national firewall, and it blocks by address. The shape is recognisable: from inside China, ping still answers but every TCP port times out, while the same server works normally from everywhere else. That is not a block of ours and not a fault on your machine.

What we do about it is replace the address. You get another one from the same datacenter and the same range, so the netmask and gateway stay as they are, and the old address is released. There is no charge while the number of addresses on the server stays the same. You change the address on the interface and in your DNS; the machine is not reinstalled and does not move. Ask in live chat or open a ticket, and tell us which address is blocked.

Two limits, stated plainly. A swap clears a block on an address; if whatever drew the block runs on the new address, the firewall can block that one too, and a new address cannot change how the firewall sees your traffic. And we do not sell DDoS protection on any line, so if what you run attracts attacks, put a scrubbing service or a CDN in front of it.

A blocked address does not call for a new machine; the swap above is faster and costs nothing. If you want a different machine anyway, for more memory or another city, it is a new order, and there is no migration fee on top of it. A ready server ordered as it stands has no setup fee, except the entry M700 cartridge at $13 once; a custom build carries a one-time $29. You pay for the new machine from the day it goes live and for the old one until you cancel it, which takes five business days' notice, so allow a short overlap while you copy your data across. Copying it is yours to do.

Within one datacenter we can usually move your addresses to the new machine; ask at the time rather than assuming it. Between cities we cannot, because each city is announced from its own address ranges.

Service is month to month with no contract. Paying three, six or twelve months ahead takes 5%, 10% or 15% off, and twelve months is the longest term. The out-of-band console, HPE iLO, Dell iDRAC or the board's own BMC, is included on every line that has one, with nothing extra to buy.