Pesquisa
Where the matching engines live, and why most of crypto now sits in Tokyo
A trading server's most important setting is its location. Venues do not advertise where their matching engines run, but enough is public to draw the map. As of autumn 2026 it has one centre of gravity.
Os artigos de pesquisa são publicados em inglês.
For a latency-sensitive strategy, the distance between your server and the venue’s matching engine sets a floor that no amount of code can lower. Light in fibre covers about 200 km per millisecond; a round trip between Frankfurt and Tokyo is around 230 ms, between Singapore and Tokyo around 70 ms, inside one cloud region well under a millisecond. Choosing the region is therefore the first engineering decision, and it has to be made from incomplete information, because venues treat the location of their matching engines as semi-confidential.
This note collects what is public. It does not include our own measurements; those belong to our trading operation, and in any case your measurements from your candidate region are the ones that matter.
What is public, by venue
Binance. Binance’s futures and spot APIs resolve directly, without a CDN in front, to addresses in Amazon Web Services’ Tokyo region (ap-northeast-1). The venue has never published an official statement, but the fact is well known in the market-making community and consistent with every public measurement.
OKX. OKX announced in 2026 that it was moving its core trading infrastructure from Hong Kong to Tokyo, completing the move at the end of July. Before that, Hong Kong was the right region; since then, Tokyo.
Bybit. Bybit ran for years from AWS Singapore (ap-southeast-1) and said so in its own co-location material. In September 2026 it announced a move to AWS Tokyo. If you place a server for Bybit, check the current state before you commit to a region; this is the venue where the answer changed most recently.
Gate, Bitget, KuCoin, MEXC. Public measurements and the venues’ own developer communications place their gateways in Tokyo, mostly inside AWS. None publishes the availability zone.
Hyperliquid. The validator set runs in Tokyo; the venue’s documentation discusses Tokyo explicitly for operators who want to minimise latency. The public API is served through a CDN, which hides the origin from a naïve traceroute but not from a round-trip measurement.
Coinbase Exchange. US-based; its documentation points institutional clients to AWS us-east-1 (Northern Virginia) for lowest latency.
Deribit. Equinix LD4 in Slough, west of London, stated in Deribit’s own co-location documentation.
Kraken, Bitfinex, HTX. Not published in a form we would quote. Measure.
What the map means
Nine or ten of the venues that matter for crypto market making are, as of autumn 2026, within a few milliseconds of one another in Tokyo. That has three consequences.
One region covers most of the book. A single well-placed server in Tokyo can quote on Binance, OKX, Gate, Bitget, KuCoin, MEXC and Hyperliquid with sub-10-millisecond round trips to all of them. Five years ago that set was spread across Tokyo, Singapore and Hong Kong and needed three machines and a state-synchronisation problem.
Hedging across venues got easier. A hedger that buys on one venue and sells on another cares about the sum of two legs. When both venues are in the same metro area, the second leg’s latency drops from tens of milliseconds to a few, and the loss on the hedge leg, the slippage between deciding and filling, drops with it.
Inside the cloud versus next to it. Within Tokyo there is still a choice: a server inside the same cloud provider as the venue, ideally in the same availability zone, or a bare-metal machine at a co-location provider in the same metro area connected through public peering. The first can be under half a millisecond away; the second is typically two to four. Whether that difference matters depends on the strategy: it is decisive for a hedger racing other hedgers and irrelevant for a strategy that rebalances once a minute.
How to find out for yourself
- Rent the cheapest virtual machine in each candidate region for an hour.
- From each, run the same measurement: open a connection to the venue’s API host, send a dozen requests, keep the warm median. The procedure in How we measure exchange latency is sufficient.
- The region with the lowest warm round trip is where the gateway is, or at least where the venue wants you to be. A round trip under two milliseconds means you are in the same metro; under one, probably the same cloud.
- Repeat quarterly. Venues move, as the two announcements above show, and they do not send a letter.
Do this before signing anything for a year. The cost of the exercise is a few dollars; the cost of a server in the wrong region is every trade.
Escrito pela equipe de engenharia da Avtim. Os números citados da documentação das exchanges estavam atualizados na data da redação; as plataformas mudam limites sem aviso, portanto verifique-os antes de construir algo com base neles. Nada aqui constitui recomendação de investimento.