How Cloud Regions and Availability Zones Actually Work
Regions and zones decide your latency floor, your legal exposure and the blast radius of your next outage. Here is what the words mean and what they cost you.
Every cloud provider sells you a map before it sells you a server. You pick a region from a dropdown, perhaps a zone inside it, and the console moves on to the interesting part. That one choice quietly fixes your latency floor, a slice of your monthly bill, your legal exposure and the blast radius of your next outage.
The vocabulary is close enough between providers to be genuinely confusing. At one provider a region is a metropolitan area containing several independent data centre campuses. At another it is a single building with a redundant power feed and a second fibre route. Knowing which one you actually bought matters far more than knowing how many of them a provider advertises.
What a region actually is
A region is a geographic grouping of capacity that a provider treats as one unit for pricing, for service availability and usually for data residency. When you launch a virtual machine, you launch it into a region. When you create a storage bucket, the bucket belongs to a region. The practical consequence is that most resources cannot see each other across a region boundary without you explicitly building a path between them.
Regions are not uniform. The oldest regions at any provider tend to have the widest service catalogue, the deepest capacity and the lowest prices, because the hardware has been amortised and the demand justified building more of it. Newer regions often launch with a subset of services and a higher rate card. If you choose a region because it is close to you and then discover the managed database you wanted is not offered there, you have learned this the expensive way.
There is also a quieter difference in how much redundancy a region contains. Large hyperscale regions are built from multiple independent facilities. Many excellent regional and value providers run a region that is one well engineered building. Neither is wrong. The architecture you put on top has to match. Providers at both ends of that spectrum are represented in our cloud account catalogue, and the difference is worth checking before you design around it.
Availability zones are a failure boundary, not a marketing label
An availability zone is the provider’s claim about correlated failure. Two resources in different zones of the same region should not go down together because of a single power event, a single cooling failure, a single network device or a single maintenance mistake. That is the entire promise. It is a statement about shared dependencies, nothing more.
The claim is not identical everywhere. Some providers back zones with physically separate buildings, separate utility feeds and separate network spines. Others describe logically isolated sections of one facility. Both can be honest descriptions, and both can be useful, but only the first protects you from a fire or a flood. When you are planning anything that needs to survive a bad day, read the provider’s own description of what a zone is rather than assuming the industry has standardised the word.
It follows that spreading three identical servers across three zones buys you something real, while spreading them across three racks in one zone mostly buys you comfort. The question to ask of any zone claim is simple: name the single component whose failure takes out both halves. If you cannot name one, the separation is doing its job.
Why latency between zones is small but not zero
Zones inside a region are connected by dedicated, high capacity links, usually measured in a small number of milliseconds round trip, often well under one. That is close enough that synchronous database replication across zones is normal practice, and far enough that it is not free. Chatty application code that makes dozens of sequential calls per request will feel the difference between a same zone and a cross zone hop.
The useful mental model is that within a zone you can be careless about round trips, within a region you should be deliberate about them, and between regions you must design around them. A single user request that fans out to twenty sequential backend calls is fine at sub millisecond latency and painful at thirty. Most cross zone latency complaints are really application design complaints that a faster network simply postponed.
Regional services, zonal services, and global services
Every provider’s catalogue splits into three scopes, and reading which scope a service has tells you most of what you need to know about its failure behaviour.
| Scope | Typical examples | What fails with it |
|---|---|---|
| Zonal | Virtual machines, attached block volumes, zone bound subnets | Loss of one zone takes the resource with it |
| Regional | Object storage, managed load balancers, many managed databases | Survives a zone loss, fails with the region |
| Global | DNS, identity and access management, billing, the console itself | Rare, but affects everything at once |
The trap is mixing scopes without noticing. A regional load balancer in front of virtual machines that all happen to sit in one zone is a regional component guarding a zonal single point of failure. A global DNS record pointing at one regional endpoint is a global component with a regional weak link. Draw the scope of each piece before you call a design redundant, and read our guide to redundancy that is worth the cost before you pay for more of it than you need.
How region choice affects your bill
The same instance specification does not cost the same everywhere. Power prices, property costs, local taxes, network transit costs and simple competitive pressure all feed into regional rate cards, and the spread between the cheapest and most expensive region at a single provider can be substantial. Nobody tells you this at signup, because the console defaults to a region near you.
Data transfer pricing also varies by region and by boundary. Traffic between zones often carries a charge, traffic between regions almost always does, and traffic out to the public internet is its own category. A multi region design can therefore add a recurring cost that has nothing to do with compute. We cover that mechanism in detail in our guide to egress fees and how they break budgets.
If your workload has no latency or residency constraint, which is true of a surprising amount of batch processing, build and test infrastructure, and internal tooling, region choice is a lever you can pull with no architectural cost at all.
Data residency and the legal side of region choice
For many teams the region question is settled before the engineering starts. Contracts, sector regulation and customer expectations can all require that specific categories of data stay inside a named jurisdiction. The region is the control that enforces that, so it stops being a performance decision and becomes a compliance one.
Two details catch people out. First, residency applies to copies as well as originals: backups, logs, monitoring data, support attachments and analytics exports are all data, and they often default to somewhere convenient rather than somewhere compliant. Second, where the data sits and who can be compelled to produce it are different questions, which is why the jurisdiction of the operating company matters alongside the location of the hardware. Teams for whom this is the deciding factor usually start with operators headquartered in their own jurisdiction, which is why accounts such as Buy Hetzner Account and Buy Scaleway Account come up so often in European procurement conversations, and why the Scaleway and OVHcloud comparison tends to be read alongside them.
Picking a first region without overthinking it
For a new project, work through four questions in order. Where are the users who will notice latency? Where is the data legally allowed to live? Does the region offer every managed service in your design? What does the rate card look like compared to its neighbours? The first constraint that binds wins, and if none of them bind, take the cheaper region.
Then write the region down somewhere that is not the console. Put it in your infrastructure code, in your runbook and in your onboarding notes. Regions are one of the hardest things to change later, not because the provider makes it difficult but because every hardcoded endpoint, every stored object reference and every replication relationship has to move with you. The cost of that move is the real reason to spend an hour on the decision now, and if you are still weighing providers rather than regions, the guides index and the frequently asked questions are both reasonable places to carry on.