How far is your exchange? Mapping crypto API latency around the world

Every CCXT tutorial ends the same way: await exchange.fetchOrderBook(symbol). What none of them tell you is that the cost of that one line — how many milliseconds pass before the bytes come back — depends less on the exchange than on where your code is running when it asks.

We measured it. Ten venues — seven spot exchanges, a perpetuals DEX (Hyperliquid) and, now that CCXT supports prediction markets, a prediction market (Polymarket) — each hit at its public order-book endpoint from eight cities on six continents. The spread is not subtle: the same call that returns in 6 ms from the right city takes over a second from the wrong one.

Pick a venue above and watch where the packets want to go. A few patterns jump out immediately.

Exchanges have a home, and it shows

The ten venues split cleanly into two camps by where they answer fastest:

• Western-hosted — Coinbase (10 ms from New York and London), Kraken (16 ms from London) and, interestingly, Polymarket (17 ms from London) live in US/EU infrastructure and get progressively slower the further east you measure.
• Asia-hosted — Binance, Bybit, Gate.io, OKX, KuCoin, Bitget and the Hyperliquid perps DEX all answer fastest from Tokyo, Singapore or Hong Kong (Hyperliquid hits 6 ms from Tokyo, the fastest number in the whole study) and pay a 100–350 ms tax when called from the Americas or Europe.

There is no globally "fast" venue. There is only fast from where you are. Run a Coinbase strategy from Singapore, or a Binance strategy from São Paulo, and you have quietly signed up for a quarter-second handicap on every request — before rate limits, before matching-engine time, before anything you can optimise in your own code.

The same line of code, 10–100× apart

Holding the venue fixed and changing only the vantage point:

Coinbase from Sydney is the standout: over a second to first byte, and it's not a fluke — across five rounds the median held at ~1.1 s. Coinbase has no Asia-Pacific point of presence for its API, so a request from Australia hairpins across the Pacific to US infrastructure and back. If you poll an order book once a second, that request never finishes before the next one is due.

"But don't they have servers near me?"

A fair objection: big exchanges run global infrastructure — surely there's a Binance box near Frankfurt? To answer it, we timed the TCP handshake (how long to connect) separately from time-to-first-byte (how long until the order-book data arrives). Binance:

The answer is both, and the distinction is the whole story. The TCP handshake is 1–3 ms from everywhere — there is an edge point-of-presence near you, terminating your connection locally (an anycast/CDN front door). You are not doing a TCP round trip to Tokyo. But an order book is live, per-request data — it can't be served from an edge cache — so the front door just forwards your request to the origin in Asia and relays the answer back. That forwarded hop is the last column, and it scales cleanly with distance. The local PoP saves you the connection-setup cost; it does not save you the data-fetch cost. For the thing you actually care about, Binance effectively always answers from Tokyo. (Cacheable endpoints like exchangeInfo behave differently — those can come from the edge and look fast everywhere. It's the live reads that pay the origin round trip.)

It's not only speed — it's access

Latency assumes you can connect at all. From our New York probes, two venues aren't slow, they're walls:

• Binance returns HTTP 451 ("Unavailable For Legal Reasons") — the well-known US geo-block. It was blocked on 3 of 5 samples; api.binance.com will not serve most US IPs, at any latency.
• Bybit returns HTTP 403 from datacenter IP ranges (4 of 5 samples, on DigitalOcean/M247-type networks). This one is network-dependent rather than a blanket geo-block — a different US probe occasionally got a 200 — but if you deploy to the wrong US cloud range, you're refused.

These show up as blocked on the map and table. "Which region should I deploy in?" is sometimes answered for you by the venue's compliance rules, not by a ping time. (Amusingly, Polymarket — which blocks US traders — happily served its order book to our New York probe in 87 ms. The API host and the trading permission are two different gates.)

How we measured this (and how to redo it yourself)

The whole study is free and reproducible — no cloud bill, no fleet of VMs to spin up.

The probe network. We used Globalping, jsDelivr's free, community-run network of measurement probes, via its public HTTP API (no key required). It runs an HTTPS GET from a real probe in each target city and returns per-phase timings: DNS, TCP handshake, TLS, and time to first byte (TTFB) — the number we report.

The request. For each venue we hit the public order-book (depth) endpoint that a CCXT fetchOrderBook maps to under the hood — for example api.binance.com/api/v3/depth, api.exchange.coinbase.com/products/BTC-USD/book, and clob.polymarket.com/book (with a live token id resolved at run time). Two need a footnote: Polymarket's book is parameterised by a rotating market token, so we fetch a current one each run; Hyperliquid's data API is POST-only, so we time a GET /info that its origin answers with a 405 — still a real round trip, just not an order-book payload.

The numbers. Each venue × city pair was sampled over five rounds; we report the median TTFB over valid round trips, which smooths out jitter. A cell where the majority of samples were refused (the 451s and 403s) is marked blocked rather than given a latency.

The collector is a single Node script committed alongside this post in the CCXT repo: docs/website/scripts/collect-latency.mjs. The snapshot it produced is committed too (latency-data.json), so you can check our numbers as well as regenerate them. Clone the repo and re-run it for a fresh reading — no dependencies beyond Node 18+:

git clone https://github.com/ccxt/ccxt.git
cd ccxt/docs/website
node scripts/collect-latency.mjs 5   # 5 rounds -> src/data/latency-data.json

Honest caveats

• TTFB includes server processing, not just the network. It's the number that actually matters to you — total time until data arrives — but it isn't a pure network RTT.
• Long-haul cells are volatile. Trans-Pacific routes bounce around: Coinbase→Sydney ranged from 328 ms to 1119 ms across rounds. The map shows the median; hover a table cell for the range.
• Exchanges sit behind CDNs and load balancers, so you're often measuring the nearest edge PoP, not a single origin box (see the TCP-vs-TTFB section above). These are access latencies, not co-location distances.
• The probes are datacenter-hosted — a feature here: it mirrors where trading bots actually run (AWS/GCP regions), not home Wi-Fi.
• It's a point-in-time snapshot. Routing and PoP placement change. Treat the shape of the map as durable and the exact milliseconds as a reading.

What to do with this

1. Deploy near your venue, not near yourself. The single biggest latency win available to most bots isn't code — it's picking the cloud region closest to where your venue answers. For Binance or Hyperliquid that's Tokyo; for Coinbase, US-East; for Polymarket, Europe.
2. Multi-venue? Pick your compromise deliberately. If you trade Coinbase and Binance from one box, no region is good for both. Accept the split, spread workers across regions, or route order-book reads over WebSockets so you pay the round trip once at subscribe time instead of on every poll.
3. Measure from your actual deployment region. Your laptop's ping is not your production ping. Run the collector (or CCXT itself) from the box you'll deploy on.
4. Confirm access before you optimise. A geo-block or a 403 is not a latency problem you can tune away. Check that your venue serves your region at all.

CCXT gives you one API across 100+ exchanges — now including perps DEXes and prediction markets — so you can move a strategy between venues by changing a single identifier. This map is the other half of that story: the code is portable, but the physics isn't. Know where your exchange lives before you pick where your bot does.