How we measured
In each city we started the smallest server (1 vCPU, 1 GB RAM) and sent 11 requests to each exchange's public server-time method over one connection. We kept the median of requests 2 to 11. That number is the time a request takes to reach the exchange and come back, without DNS lookup, TCP handshake or TLS setup, which a running bot pays only once per connection.
We ran the whole set twice, a few minutes apart, and kept the lower number for each pair. The two runs never differed by more than 20 ms. Then we destroyed all 21 servers, which cost a few cents in total.
One limit to keep in mind: the public time method shows where an exchange serves its REST API. Order gateways and WebSocket feeds can sit somewhere else, so before you commit, repeat the check from your own server against the endpoint your bot actually uses.
The results by exchange
| Exchange | Closest city | Round trip | From Frankfurt |
|---|---|---|---|
| Binance | Tokyo | under 10 ms | about 230 ms |
| Bybit | Tokyo | under 10 ms | about 230 ms |
| OKX | Tokyo | about 25 ms | about 255 ms |
| Hyperliquid | Tokyo | about 15 ms | about 240 ms |
| Coinbase | Chicago | about 30 ms | about 110 ms |
| Kraken | London | about 15 ms | about 45 ms |
Four of the six exchanges answer fastest from Tokyo. Singapore is the runner-up for all four, at 75 to 95 ms. Sydney, Mumbai and Bangalore land between 105 and 170 ms.
Coinbase is a US East exchange: Chicago about 30 ms, Atlanta about 50 ms, New Jersey about 50 ms, Toronto about 45 ms. Kraken is the European one: London about 15 ms, then Amsterdam, Nuremberg and Frankfurt between 30 and 45 ms.
From our eight US cities, Binance and Bybit returned a refusal from the nearest edge node instead of an answer, so there is no latency to report. Both exchanges do not serve US addresses on these API hosts. If your bot trades there, a US server is not an option at any latency.
What the numbers mean for your bot
A bot that trades on candle closes, timers or signals does not care whether a request takes 10 ms or 250 ms. Its decisions come seconds or minutes apart, and a quarter of a second disappears in that. For such bots, pick the city you like for other reasons and spend the effort on keeping the process alive instead. Our post on picking the region for a trading bot covers that side.
A bot that races other traders for fills is different. Market making, arbitrage between venues and liquidation hunting all lose money to whoever is 200 ms closer. For those, the table above is the whole decision:
- Trading only on Binance, Bybit, OKX or Hyperliquid: Tokyo.
- Trading only on Coinbase: Chicago.
- Trading only on Kraken: London.
- Trading on Binance and Coinbase at once: no city is close to both. From Tokyo Coinbase is about 185 ms away, from Chicago Binance does not answer at all. Run one server near each exchange and let them talk to each other, or put the bot near the exchange where speed decides the trade.
Checking it yourself
The measurement fits in four lines of shell, and it is on the exchange latency page ready to copy. Rent the smallest server in two or three candidate cities for an hour, run it against your exchange, keep the city with the lowest median and destroy the rest. On VPS for trading bots the smallest plan is billed by the hour, so an hour of testing in three cities costs a few cents.
Repeat the check every few months. Exchanges move their infrastructure, and a table from November is a starting point, not a guarantee. We plan to rerun ours monthly and update the page.