The same cluster,
priced on every cloud
One production cluster, 24 dedicated worker vCPUs, costs about €463 a month on Hetzner and about €1,051 on Azure. Identical capacity, a 2.3 times spread. Change the shape below and watch which line moves.
Price one cluster on 8 clouds
Describe the cluster you run. Every column is the same capacity priced from that provider's own published list, with the control plane, load balancers, storage and traffic included rather than assumed away.
The same cluster, a month, at list price
€463 on Hetzner to €1,051 on Azure
A 2.3x spread for identical vCPUs, memory, storage and traffic.
- Worker nodes
- Control plane
- Load balancers
- Block storage
- Network
Hetzner k3s on your own servers, Germany and Finland
€463 a month- Worker nodes3 × CCX33 (8 vCPU, 32 GB, dedicated)
- €415
- Control plane3 × CX33 control plane servers
- €25
- Load balancers1 × LB11 load balancer
- €7
- Block storage250 GB of block storage volumes
- €14
- Network20 TB of traffic included per server, then €1 per TB
- €0
- Hetzner sells no managed Kubernetes, so a highly available control plane is three servers you pay for and somebody operates.
- Arm CAX nodes are cheaper again and are left out here, because moving to them means every image in the cluster has to be multi-architecture.
Prices read 2026-09-14 from Hetzner Cloud pricing, the 15 June 2026 price adjustment.
On-demand list prices checked September 2026, before VAT, discounts, commitments or spot. Always-on month, European regions. USD and EUR convert at the ECB rate of 2026-09-15 (1 EUR = 1.1539 USD). Ankra charges the same per vCPU on every provider, so it never changes the order.
Six costs that survive every rightsizing exercise
None of them are about how big your pods are, which is why trimming resource requests never closes the gap.
A fee for having a cluster at all
A managed control plane is charged per cluster per hour whether or not anything runs on it, so the team that split one cluster into four for isolation quadrupled a line item that has nothing to do with workloads. Some providers charge nothing for the same thing.
Two products wearing one label
Shared, burstable and dedicated cores all get sold as vCPUs. On the same benchmark the dedicated tier averaged 7 times the work of the shared one, so a cheap vCPU can be a bargain or a trap depending on which one you bought.
Capacity you paid for and cannot use
On an entry-class GKE node, system pods reserve most of the CPU and 187 millicores of 2,000 are left for your workloads. Sizing a node pool off the spec sheet buys about a fifth of the node you think you are buying.
Storage that is not the same storage
Network volumes at a fixed IOPS cap and local NVMe are both priced per gigabyte. In our tests the gap in sequential write reached 27 times inside one price class, which decides whether a database needs three nodes or nine.
The network bill nobody models
Egress, gateways and cross-zone traffic are invisible when you plan and obvious when you reconcile. Some providers bundle tens of terabytes with every node, others bill each gigabyte leaving the building.
Platform time, which has no invoice
Upgrades, glue code and being on call for the delivery pipeline are a real recurring cost. They just arrive as engineer months rather than as a line on a bill, which is why they survive every cost review.
A vCPU is not a vCPU
Every machine below is sold as a 2 vCPU node. We ran the same benchmark on all of them.
Dedicated cores in blue, shared and burstable in orange. One day per provider, two workers each, one size class, from our six cloud benchmark and the round before it. There is no middle tier. There are two products wearing one label.
This is why the calculator asks whether you want shared or dedicated cores, and never quotes a fleet on dedicated cores as a saving on shared ones.
The levers that move the number
Ordered by what they return. The first two are worth more than everything a dashboard will ever recommend.
Put each workload where it actually belongs
Keep the user-facing API on premium infrastructure. Move logging, backups, build runners and batch work to cheaper capacity. Same definition, different cost profile, and none of your users can tell.
Make moving cheap enough that you do it
A Stack is one object you deploy to many clusters, so a second provider is minutes of work rather than a quarter long project. The saving only counts if somebody takes it.
Stop paying for a delivery control plane
Ankra drives Helm from the agent already running in your cluster, so the usual pile of controller pods stops existing in every cluster, and stops needing upgrades and on call.
Buy on benchmarks, not on spec sheets
We publish the same Kubernetes benchmark suite across six clouds. Price per unit of delivered work is the number that matters, and it is rarely the number on the pricing page.
Pick a platform that does not tax cluster count
Ankra bills per worker vCPU with a free allowance, and no per-cluster fee, so splitting one cluster into three for isolation does not triple the platform bill.
Workload placement has its own page, with the tier by tier breakdown and the before and after bill. Multi-cloud and on-prem.
Portability is the hedge
On 15 June 2026 a Hetzner CCX33 went from 62.49 to 138.49 euros a month. The teams that shrugged were not on a cheaper provider, they were able to leave.
Import what you already run
One Helm command per cluster, on any provider, managed or self-run. Nothing needs rebuilding and no firewall changes, so the first useful day is the first day.
Turn each environment into a Stack
Standard Helm charts and manifests in your own Git repository, with the deployment order written down. This is the step that turns a migration into a clone.
Move the cheapest thing to be wrong about
Logging, backups or build runners first. Per-cluster variables absorb the differences, such as domains, storage classes and registry credentials.
Keep the option open
Once the stack is portable, the next repricing letter is a pricing decision instead of an engineering emergency. That option is worth more than this month's saving.
Your node groups, your volumes, your load balancers
The calculator above prices a shape you typed. Inside Ankra, Compare providers prices the cluster you actually run, node group by node group, against today's metered bill, and tells you what changes about each machine if you move.
Plays from YouTube. Open on YouTube instead
Questions people ask before they move
How much does a Kubernetes cluster cost per month?+
It depends far more on where you run it than on how you size it. A production cluster with 24 dedicated worker vCPUs, 96 GB of memory, a load balancer and 250 GB of storage ranges from about €463 a month on Hetzner to about €1,051 on Azure at list prices, a 2.3 times spread for the same capacity.
Which cloud is cheapest for Kubernetes?+
On published list prices for a like-for-like cluster, Hetzner is the cheapest of the 8 providers compared here. That is not the same as the best choice, because the hyperscalers include managed control planes, wider regional coverage and services this comparison does not price. The useful question is which workloads need what you are paying a premium for.
Why is EKS more expensive than it looks?+
Three costs arrive after the instance price. The control plane is billed per cluster per hour, a cluster that ages out of standard support moves to extended support at 6 times that fee, and private node pools usually run behind NAT gateways that bill per hour and per gigabyte processed.
Is a cheap vCPU the same as an expensive one?+
No. In our benchmark every shared or burstable node landed between about 1,070 and 1,350 CPU events per second, while every dedicated-core node landed between about 7,970 and 9,620. All of them were sold as 2 vCPU machines. Compare price against benchmarked work, not against the spec sheet.
Do I have to change provider to cut my Kubernetes bill?+
Usually not. Most of the saving comes from putting workloads on the right tier, picking the right instance family, and switching off environments that idle overnight. Being able to move is what keeps a provider honest at renewal, and it is worth having whether or not you use it.
What does Ankra add to the bill?+
Ankra bills per worker vCPU with a free allowance, control planes are never billed, and there is no per-cluster fee. The rate is the same whichever provider you run on, so it never changes which column is cheapest.
The numbers behind this page
Every claim above comes from something we published with the working shown. List prices were last checked in September 2026.
Hetzner Repriced. Here Is the Math for Your Kubernetes Cluster.
What a production cluster costs after the June 2026 increase, why the panic takes were wrong, and why the lesson is portability rather than provider loyalty.
Read itAWS vs GCP vs Hetzner: Six Clouds, One Kubernetes Benchmark
The same suite on six providers, so you can compare price against delivered performance instead of against a spec sheet.
Read itFalling Behind on Kubernetes Versions Now Has a List Price
Three releases a year, fourteen months of patches, and a six times cluster fee when you lapse. How to make upgrades routine before they become a bill.
Read itKubernetes Logging in 2026
Elasticsearch, OpenSearch, Loki, or renting it from your cloud, with the cost consequences of each spelled out.
Read itFind out what your fleet actually costs
Import your clusters in five minutes, see the whole fleet in one place, and price it on every provider Ankra can build on before you move anything.