Home / Costs & subscriptions / Tokyo, Singapore, Frankfurt: Latency and Price for Cloud Servers
Tokyo, Singapore, Frankfurt: Latency and Price for Cloud Servers
The same specs can cost twice as much across three regions. What latency, bandwidth and compliance each buy you.
The most commonly overlooked point when choosing a region: where your readers are matters more than where your datacenter is. Build a site for European and American readers and put the server in Tokyo, and you've given away 100ms of latency for nothing.
The actual differences between the three locations
| Region | To mainland China | To Europe | To US East | Typical monthly price (2C2G) |
|---|---|---|---|---|
| Tokyo | 40–90ms | 250ms+ | 150ms+ | $5–10 |
| Singapore | 60–120ms | 280ms+ | 200ms+ | $5–12 |
| Frankfurt | 200ms+ | 20–40ms | 90ms | $4–8 |
The price column is only an order of magnitude; the specifics depend on promotions and bandwidth caps. What really determines the experience is the distribution of your target readers.
Bandwidth hits the bottleneck before CPU does
Load on a small site is almost entirely static requests. 2 cores / 2G is enough to serve a few thousand daily actives, but a 1TB monthly traffic cap gets eaten up by images.
How to judge: work out "average page size × daily visits × 30". A 200KB page with 5,000 daily visits is about 30GB a month; if the first screen carries three large uncompressed images, that number goes up fivefold.
Compliance and filings
- For European and American readers: Frankfurt/Amsterdam + Cloudflare, the least hassle
- For Asia: Tokyo/Singapore, low latency but expensive bandwidth
- If you need stable access from mainland China: this often conflicts with the goal of going overseas, so decide the priorities first
My current combination
A small Frankfurt instance runs the main site (readers are mostly in Europe), and the Tokyo instance only serves as a jump host and backup. Both use object storage for images, so bandwidth isn't maxed out by static assets.
One reminder
The cost of picking the wrong region isn't money, it's that you'll spend half a year optimizing the wrong metric — say, repeatedly tuning image compression on a Tokyo server while your readers are actually in Germany.
The actual differences between the three regions
| Region | Latency (from mainland China) | Bandwidth cost | Suited for |
|---|---|---|---|
| Hong Kong | 30–60ms | High | Mainland China + overseas |
| Singapore | 70–120ms | Medium | Southeast Asia + global |
| Japan/Tokyo | 60–100ms | Medium | Japan, North America relay |
A 50ms latency difference is barely perceptible for a static site, and very obvious for interactive applications (login, payment callbacks).
Cost has to be calculated over three years
A $5/month machine is $180 over three years; but if you have to migrate once a year (because of overselling, the provider vanishing, or a price hike), the time cost of migrating far exceeds the price difference. Look at the renewal price, not the first-term price.
A practical rule of thumb
- Static site + CDN only: pick the cheapest, the region doesn't matter
- Running a database: pick low latency and NVMe disks
- Running long-lived connections/proxies: look at the route, not the specs
One-line summary
The most expensive item in a VPS's price is never the monthly fee, it's "can this machine run stably for two years".
A final reminder: before buying, test the return route once (mtr to a city you commonly serve). Some cheap regions have great outbound routes but a roundabout return path, and the latency curve turns into a roller coaster at peak hours.
An order of decisions
- First set a latency ceiling (within 100ms for interactive, irrelevant for static sites)
- Then set bandwidth (estimate from peak PV × average page size)
- Finally compare prices, and only look at the renewal price
Do it the other way round — compare prices first — and you usually end up buying another relay machine because the cheap one took a roundabout route.
The real trade-offs with three machines
A common combination on the same budget: one Hong Kong machine for the main site, one Singapore machine for backup and relay, and one cheap US machine for batch jobs. The main site needs latency, the backup needs "not in the same datacenter", and batch jobs need cheap. Splitting machines by role saves more money than splitting by specs.
One more thing: every machine has to be independently rebuildable. Write your own deployment scripts and data backups, so that switching datacenters doesn't turn into learning everything over again.