Where the machines actually are
Every location on this page is a room we have equipment in, not a partner logo or a resale agreement dressed up as a footprint. Some of these sites we have run since 2019. The newest one is racked and cabled and still waiting on its transit to settle.
Europe
18 SITES| Site | Country | Live since | Uplink | Plan | Stock |
|---|---|---|---|---|---|
| AMS-01 | Amsterdam, Netherlands | 2019 | 400 Gbit/s | RyzenEPYCGPUStorageBare | In stock |
| AMS-02 | Amsterdam, Netherlands | 2022 | 400 Gbit/s | RyzenEPYCStorage | In stock |
| FRA-01 | Frankfurt, Germany | 2019 | 400 Gbit/s | RyzenEPYCGPUStorageBare | In stock |
| FRA-02 | Frankfurt, Germany | 2023 | 200 Gbit/s | RyzenEPYCBare | Low stock |
| HEL-01 | Helsinki, Finland | 2020 | 200 Gbit/s | RyzenEPYCStorage | In stock |
| STO-01 | Stockholm, Sweden | 2021 | 200 Gbit/s | RyzenEPYCStorage | In stock |
| ZRH-01 | Zurich, Switzerland | 2020 | 200 Gbit/s | RyzenEPYCBare | Low stock |
| KEF-01 | Reykjavík, Iceland | 2022 | 100 Gbit/s | RyzenStorage | In stock |
| BUH-01 | Bucharest, Romania | 2019 | 200 Gbit/s | RyzenEPYCStorageBare | In stock |
| SOF-01 | Sofia, Bulgaria | 2021 | 100 Gbit/s | RyzenStorage | In stock |
| KIV-01 | Chișinău, Moldova | 2020 | 100 Gbit/s | RyzenStorage | Low stock |
| RIX-01 | Riga, Latvia | 2021 | 100 Gbit/s | RyzenEPYCStorage | In stock |
| LON-01 | London, United Kingdom | 2019 | 200 Gbit/s | RyzenEPYCGPUBare | In stock |
| PAR-01 | Paris, France | 2020 | 200 Gbit/s | RyzenEPYCBare | In stock |
| MAD-01 | Madrid, Spain | 2022 | 100 Gbit/s | RyzenEPYC | In stock |
| MIL-01 | Milan, Italy | 2021 | 100 Gbit/s | RyzenEPYC | In stock |
| WAW-01 | Warsaw, Poland | 2020 | 200 Gbit/s | RyzenEPYCStorage | In stock |
| PRG-01 | Prague, Czechia | 2022 | 100 Gbit/s | RyzenEPYC | In stock |
North America
05 SITES| Site | Country | Live since | Uplink | Plan | Stock |
|---|---|---|---|---|---|
| NYC-01 | New York, United States | 2019 | 400 Gbit/s | RyzenEPYCGPUStorageBare | In stock |
| MIA-01 | Miami, United States | 2020 | 200 Gbit/s | RyzenEPYCStorage | In stock |
| DAL-01 | Dallas, United States | 2021 | 200 Gbit/s | RyzenEPYCGPUBare | In stock |
| LAX-01 | Los Angeles, United States | 2020 | 200 Gbit/s | RyzenEPYCGPU | Low stock |
| YYZ-01 | Toronto, Canada | 2021 | 100 Gbit/s | RyzenEPYCStorage | In stock |
Latin America
03 SITESAsia-Pacific
06 SITES| Site | Country | Live since | Uplink | Plan | Stock |
|---|---|---|---|---|---|
| SIN-01 | Singapore, Singapore | 2020 | 200 Gbit/s | RyzenEPYCGPUStorage | In stock |
| HKG-01 | Hong Kong, Hong Kong | 2021 | 100 Gbit/s | RyzenEPYC | Low stock |
| TYO-01 | Tokyo, Japan | 2021 | 200 Gbit/s | RyzenEPYCGPU | In stock |
| ICN-01 | Seoul, South Korea | 2022 | 100 Gbit/s | RyzenEPYC | In stock |
| BOM-01 | Mumbai, India | 2022 | 100 Gbit/s | RyzenStorage | In stock |
| SYD-01 | Sydney, Australia | 2022 | 100 Gbit/s | RyzenEPYC | In stock |
Middle East & Africa
02 SITES| Site | Country | Live since | Uplink | Plan | Stock |
|---|---|---|---|---|---|
| DXB-01 | Dubai, United Arab Emirates | 2023 | 100 Gbit/s | RyzenEPYC | In stock |
| JNB-01 | Johannesburg, South Africa | 2023 | 100 Gbit/s | Ryzen | Pre-order |
Thirty-four cities, twenty-nine countries
Latency figures are measured medians from our own probes, not vendor brochures. They move; the page updates when they do.
Pick the law first, the milliseconds second
Almost every location guide on the internet opens with a latency map. Ours opens with the courts, because in five years of doing this the customers who regretted their choice have never once regretted the ping.
Latency you can fix by adding a second site. Jurisdiction you cannot.
The order of operations matters. Work out which legal system you are content to operate under, shortlist the sites inside it, and only then compare round-trip times between whatever survives. Do it the other way round and you end up with a server that is forty milliseconds closer to your users, sitting in a country whose notice-and-takedown regime you have not read.
This is not advice about breaking the law anywhere. It is advice about knowing, before you pay, which set of rules you have agreed to live under, how a complaint reaches us, and what has to happen before anything is taken offline. That answer is different in Chișinău than it is in Frankfurt, and both of those are different again from Panama City.
Once the shortlist is legal rather than geographic, latency is usually an easy call. Within Europe every site is inside forty-five milliseconds of every other; across an ocean the physics stops negotiating and nothing you buy will help.
The site is the jurisdiction
A machine in Riga answers to Latvian law and a machine in Panama City answers to Panamanian law, regardless of what else is true about the account that pays for it. That is the entire reason we operate thirty-four of them instead of four.
Latency is a floor, not a setting
Light in glass covers roughly two hundred kilometres per millisecond, and routes are never straight. Amsterdam to Singapore is one hundred and sixty-eight milliseconds because the cable is long, not because someone forgot to optimise it.
Twenty milliseconds is invisible, two hundred is not
For a web application, the difference between Prague and Frankfurt will not be noticed by anyone. The difference between Frankfurt and São Paulo will be noticed on the first page load, and again on every round trip your protocol makes after it.
Transit cost varies far more than distance
A terabyte leaving Bucharest and the same terabyte leaving Sydney are separated by an order of magnitude in what the upstream charges us. If your workload is egress-heavy, that gap will dominate your bill long before core count does.
Stock belongs in the decision
Six sites are marked low or pre-order at any given time, and the ones people fight over are rarely the obvious ones. Zurich and Hong Kong sell out of EPYC capacity faster than we can rack it.
What the numbers on each site page mean
Every location page carries the same five figures. They are worth two minutes of your attention, because three of them are commonly published as something they are not.
All measured, none rounded in our favour.
Round-trip time
A thirty-day median from our own probes to four reference hubs: Amsterdam, New York, Singapore and São Paulo. Medians, not best-case samples taken at four in the morning. They drift when upstream paths change, and the page changes with them.
Uplink
Capacity into the site, not capacity sold to you. A hundred-gigabit site can still hand you a forty-gigabit port; a four-hundred-gigabit site simply has more headroom before anybody notices a neighbour having a bad day.
Scrubbing
Filtering capacity available at the edge in front of that site, between four and twelve terabits per second. It is always in path. There is nothing to enable and nothing to page us about when an attack starts.
Live since
The year the first production instance was billed there, not the year we signed a contract or announced an intention. Four sites carry 2019 and they have been running continuously ever since.
Stock
In stock means order it now. Low means the popular configurations are thin and a large order may need a few days. Pre-order means the hardware is racked but not yet selling normally.
None of these figures are marketing rounding. If a site is doing badly this week, the status page says so and the latency table moves, which is the whole reason we publish medians instead of a pretty map with dots on it.
Identical everywhere, and the parts that are not
A great deal of what we sell is deliberately the same in all thirty-four rooms, so that moving between them is a decision about law and distance rather than about relearning how anything works.
The platform is one platform. The physics and the law are local.
The same silicon rules
Zen 5 and Zen 4 only. Nothing older than the current generation is in production at any site, and a core you buy in Sofia is the same dedicated core you buy in Tokyo.
The same panel and the same API
One account, one console, one API token, every site. Rebuilds, snapshots, RDNS and out-of-band access behave identically whichever city you are pointed at.
The same policy
An email address at sign-up and nothing else, everywhere. No site has a different threshold, a stricter verification step or a local exception buried in its own terms.
The same routed /64
IPv6 is delegated at every location, at no cost, on every plan. Custom ISO installation is available at every location too, though it is a manual job and takes longer than the forty-seven second median.
The same refund window
Seven days, no reason required, in the asset you paid with. Ordering in an unfamiliar city is therefore cheap to test and cheap to abandon.
| What differs by site | Range across the fleet | Why |
|---|---|---|
| Product lines available | One to five of the five | GPU floors and bare metal need power density and space that not every room has |
| Uplink into the site | 100 to 400 Gbit/s | Upstream capacity is bought locally and priced locally |
| Scrubbing capacity | 4 to 12 Tbit/s | Filtering sits at regional edges, not in every building |
| Egress cost to us | Roughly tenfold, worst to best | Transit in Bucharest and transit in Sydney are not the same market |
| Applicable law | Twenty-nine legal systems | This is the reason the list is long |
| Lead time on unusual builds | Same day to two weeks | Some rooms hold spares locally, others are served from a regional depot |
The five regions
The fleet is grouped into five regions, which is how the ordering flow, the looking glass and the off-node backup targets all think about placement. Backups always land in a different site, and by default in a different country.
Eighteen sites in Europe. That ratio reflects our customers, not our ambitions.
Europe · eighteen sites
The dense part of the fleet, and the widest spread of legal systems on offer. Amsterdam and Frankfurt carry the volume, London holds the shortest transatlantic path, and the Baltic and Balkan sites exist because people want an EU-adjacent jurisdiction that is not Germany. Everything here is inside forty-five milliseconds of everything else.
North America · five sites
New York anchors the east, Los Angeles the west, Dallas the middle, and Toronto sits fourteen milliseconds from New York under a different flag. Miami is in this region on the map and in the Latin American one in practice. Metal and GPU are easiest to get in Dallas.
Latin America · three sites
Panama City for jurisdiction, São Paulo for the users, Santiago for the southern cone. Bandwidth is dearer here than anywhere except Australia, and the honest answer for many Latin American projects is still a Miami origin with a local edge in front of it.
Asia-Pacific · six sites
Singapore is the default and Tokyo is the stable one. Hong Kong holds the shortest path to the mainland, Seoul has the lowest jitter we measure anywhere, Mumbai is the cheap way into South Asia, and Sydney is expensive because Australian transit is expensive for everyone.
Middle East and Africa · two sites
Dubai sits on the hinge between Europe, East Africa and the subcontinent, which makes it a useful third leg on an anycast setup. Johannesburg is the newest room we have and is currently taking pre-orders while its transit arrangements settle.
“We spent a month arguing about whether to be in Frankfurt or Bucharest and about ninety seconds on the latency difference, which turned out to be twenty-seven milliseconds and completely irrelevant to us.”
How a site opens, and how one gets closed
New locations are not announced because a region looks underserved on a slide. They open because enough existing customers keep asking for the same city, and they open slowly.
We have retired two sites. Both times the customers knew before the public page did.
Retirement runs in reverse and is deliberately slow. If a facility becomes untenable, whether through a landlord change, an upstream we no longer trust or a legal environment that has shifted under us, affected customers are emailed before anything is published, with a target date at least ninety days out.
Migration to any other site is free in that situation, including the data transfer and including a fresh IP allocation, and we will do the move at whatever hour suits you rather than at whatever hour suits us. Nobody has ever been given a shorter notice period than ninety days. We do not intend to start.
- 01
Demand, counted rather than guessed
Every ticket and pre-sales message asking for a city we do not have is tallied. A city needs to be asked for consistently across at least two quarters before anyone goes looking at rack space in it.
- 02
Space, power and upstreams
We take space in facilities that will sell us diverse transit rather than resell us their own, and we want at least two independent upstreams plus local peering before signing. This part takes months and kills roughly half the candidates.
- 03
Hardware in, network up
Nodes are built and burned in at a regional depot, shipped, racked and cabled. The site joins the internal network and the looking glass first, and sits there being measured while nothing is sold on it.
- 04
Thirty days of quiet
A month of synthetic load, deliberate failure testing and probe data before a single customer instance exists. If the numbers are unconvincing, the site stays shut and we eat the cost, which has happened.
- 05
Pre-orders, then general availability
The location page goes live, usually in pre-order state, and stock opens when the transit contracts are final and the routing has stopped changing shape week to week.
Pre-order sites
Johannesburg is currently the only site in pre-order. The hardware is in the rack and powered; what is still moving is the transit, and until that has stopped moving we are not willing to quote a round-trip time we would have to correct later.
A pre-order is a queue position, not a deposit we are sitting on.
What it buys
A dated position in the queue and the launch price locked for the first twelve months. Instances are provisioned in queue order on the day the site opens, and you are told the date as soon as we know it rather than when marketing would prefer.
What it costs now
Nothing until provisioning. No card exists to charge and no invoice is raised, because there is no server yet. The first payment request arrives when the machine does, and the seven-day refund window starts from that point rather than from the pre-order.
If we cancel
Sites do occasionally fail their thirty days of quiet. If that happens you get an email explaining precisely which numbers were unacceptable, and your queue position transfers to any other site you like at the same locked price.
Sites marked low are a different thing entirely. They are selling normally, the popular configurations are simply thin on the ground, and a large order there may need a few days rather than forty-seven seconds. Ask before you order and we will tell you what is on the floor right now.
Questions about placement
Yes, and it is a migration rather than a resize: a new instance is built at the destination, we help you move the data, and the old one is destroyed once you confirm. The IP address changes, because addresses belong to the site. There is no fee, but the two instances overlap for a day or two and both are billed for that period.
Amsterdam if your users are in Europe, New York if they are in the Americas, Singapore if they are in Asia. Those three carry the deepest stock, every product line and the best-understood routing in their regions. You can always move later, and plenty of people do.
Bucharest, then Sofia and Warsaw. Transit in south-eastern Europe costs a fraction of what it costs in Zurich or Sydney, and our storage node pricing reflects that. If you are moving serious volume, tell support roughly how much before you order and we will point you at the right room.
That is the normal case. One account, one balance, one API token, and instances anywhere in the fleet. Off-node backups deliberately land in a different site to the instance they protect, and you can pick which one if you care about the country it lands in.
The hardware is ours in all of them, bought outright and installed by people we employ. The buildings are not ours anywhere, which is true of almost everyone who tells you otherwise. What we do not do is resell somebody else’s virtual machines and call it a location.
Thirty-four rooms. Pick one.
Configuration takes about a minute, payment takes as long as your chosen chain does, and root credentials follow a median of forty-seven seconds after that. If the city you want is not on the list, say so — that tally is how the list gets longer.