Cheaper Than Hetzner: The Cheapest On-Demand Kubernetes on the Internet
On this page
Plays from YouTube. Open on YouTube instead
A Kubernetes environment with 2 vCPU and 4 GB of memory on Ankra costs €5.21 a month. The same shape at Hetzner, the provider everyone in Europe points to when they say “cheap”, is a CX23 at €5.49, plus €0.50 a month for its IPv4 address. And a CX23 isn’t a Kubernetes cluster. It’s a Linux server you still have to turn into one.
That gap holds at every size we sell. Each paid Playground is priced below the bare Hetzner server of the same shape, comes with 25 percent more storage, and arrives with the control plane, ingress, TLS and a public hostname already running. We went looking for anything cheaper that you can order on demand and keep, and didn’t find it. So we’re saying it plainly: Ankra Playgrounds are the cheapest on-demand Kubernetes on the internet. Here’s the math behind that, including the cases where a Playground is the wrong tool.
The numbers
Hetzner’s figures are its current list prices for Germany and Finland, excluding VAT, from the price adjustment that took effect on June 15, 2026. We covered what that repricing did to production clusters in July. Playground prices are our published catalogue, also in EUR excluding VAT.
| Shape | Ankra Playground | Hetzner server + IPv4 | Storage vs disk |
|---|---|---|---|
| 2 vCPU / 4 GB | Small: €5.21 | CX23: €5.49 + €0.50 = €5.99 | 50 GB vs 40 GB |
| 4 vCPU / 8 GB | Medium: €8.06 | CX33: €8.49 + €0.50 = €8.99 | 100 GB vs 80 GB |
| 8 vCPU / 16 GB | Large: €15.19 | CX43: €15.99 + €0.50 = €16.49 | 200 GB vs 160 GB |
| 16 vCPU / 32 GB | Extra large: €28.01 | CX53: €29.49 + €0.50 = €29.99 | 400 GB vs 320 GB |
On sticker price alone, every Playground is 5 percent under the matching CX server. Count the IPv4 address a Hetzner server needs to be reachable the way a Playground hostname is, and the gap becomes 13 percent on Small, 10 on Medium, 8 on Large and 7 on Extra large.
We picked the CX line on purpose. It’s Hetzner’s cheapest x86 family, it runs on shared vCPU like ours, and it came through June’s repricing up by about a third rather than doubling like the CPX and CCX lines. Against the CPX line, where a 2 vCPU / 4 GB CPX22 now lists at €19.49, a Small Playground costs just over a quarter as much.
Then there’s the adjustment most comparisons skip. On a CX23, everything you install to make it a cluster comes out of the 2 vCPU and 4 GB you paid for: the Kubernetes control plane, CoreDNS, an ingress controller, cert-manager. In a Playground those components run on top of the size you order, not inside it, so the full 2 vCPU and 4 GB are for your own pods. If you want that much room free for your own workloads on Hetzner, the next size up is a CX33, €8.99 a month with its IPv4, which is 73 percent more than a Small Playground.
What €5.21 actually buys
“Kubernetes” covers everything from a kind cluster on a laptop to a multi-region fleet, so it’s worth being specific.
A Playground is a virtual cluster with its own API server, running on dedicated servers Ankra operates. It’s single-tenant and fully writable: deploy things, delete things, break things. It comes up in one to three minutes, and the film at the top of this post clocks a real one at 82 seconds. By the time it reports ready, the Ankra agent is installed and connected and a security baseline is already scanning your images.
Anything you publish through an Ingress gets a public HTTPS hostname under your Playground’s own domain, with a certificate that actually validates. On a fresh server that’s usually the step that eats an afternoon: install an ingress controller, install cert-manager, write a ClusterIssuer, point DNS somewhere, wait on Let’s Encrypt. Here it’s done before you arrive.
Small includes 50 GB of storage for PersistentVolumeClaims, room for 40 pods and 10 volumes, and the larger sizes scale up from there. There’s no per-GB charge, no egress charge, no IP charge and no load balancer to rent, and a Playground never counts against the 30 free worker vCPUs on your Ankra plan. The monthly price is the whole bill.
You drive it like any other Ankra cluster: the portal, the CLI, the API, or Ankra AI, which can list the sizes, order one, resize it and destroy it, and always names the size and the price before it spends your money.
ankra cluster playground plansankra cluster playground create --size smallankra cluster playground status <cluster-id>On demand means by the second
Hetzner bills cloud servers by the hour and rounds up, so a server that exists for ten minutes is charged a full hour. It also bills for as long as the server exists, whether it’s running or powered off. That’s a reasonable model for servers. It fits less well with how development environments actually get used: spin one up to review a branch, test a chart upgrade, show a customer something, tear it down.
Playgrounds bill pro-rata by the second, from the moment the environment reports ready until you destroy it. A Small works out to roughly €0.007 an hour, against €0.0088 for a CX23 before its IP address. A full eight-hour day of experimenting costs about six cents. Destroying one takes under a minute, and the next one is ready in about two.
Two more details matter if you keep one around. Resizing happens in place: only the quota and the price change, prorated across the change, and nothing you deployed is touched. And the price you order at is your price of record. If we change the catalogue later, the new prices apply to new orders only, never to an environment you’re already running. After the year cloud pricing has had, we think that promise belongs in writing.
If you only want to look around, the free Trial gives you 1 vCPU, 2 GB and 20 GB of storage with no card. It lasts as long as someone in your organisation is around, and expires after three hours with everyone away (a week if you’re on a paid plan), with a warning by email and in the app before it goes.
How a cluster can cost less than the server under it
Fair question. If Hetzner sells a 2 vCPU server for €5.49, how can a 2 vCPU Kubernetes cluster cost €5.21?
Because a Playground isn’t a virtual machine. When you rent a CX23 you rent a whole slice of a host: a kernel, an operating system, a disk image, a public IP, and then, once you install it, a Kubernetes control plane that exists only for you. Most of that is overhead, and you pay for all of it whether your pods use it or not.
A Playground is a quota-enforced share of dedicated servers that are already running Kubernetes. Your cluster gets its own API server and its own isolation, while the host kernel, the ingress layer, the certificate machinery and the network are shared by every Playground on the host. It’s the same trick that makes shared vCPU cheaper than dedicated vCPU everywhere, applied one layer up the stack. Keeping that density sane is our job, and it’s why a size can show as unavailable in the picker when its pool is full.
The platform components still cost something per environment, roughly two vCPU and two gigabytes each by our own accounting, and we absorb that rather than bill it. For comparison, Amazon charges USD 0.10 an hour, about USD 73 a month, for an EKS control plane before you’ve added a single node.
Is it really the cheapest on the internet?
We checked the usual alternatives on September 11, against each provider’s own pricing page. Most managed Kubernetes services give the control plane away now, so what you actually pay for is the nodes, plus a load balancer if you want traffic in from the internet.
| Provider | Control plane | 2 vCPU / 4 GB node | Load balancer |
|---|---|---|---|
| Scaleway Kapsule | free | about €20.10 (PLAY2-NANO) | extra |
| Civo | free | $21.73 | $10.86 |
| DigitalOcean DOKS | free | $24.00 | $12.00 |
At the same shape, those come to roughly three to four times a Small Playground before you’ve paid for a way in from the internet. You can get a smaller bill if you shrink the machine: Civo’s extra-small node is $5.43 a month, but that’s 1 vCPU and 1 GB, a quarter of the memory a Small Playground gives you for about the same money. Scaleway lists instances for under a euro a month, but Kapsule won’t take them as nodes because they don’t have enough memory.
Then there’s free, and it would be silly to pretend it doesn’t exist. Our own Trial is free. Oracle’s Always Free tier gives every account the equivalent of two Arm cores and 12 GB of memory, and OKE basic clusters carry no control plane fee. That’s a genuinely good deal if you can get it, but free capacity in a region can run out, and Oracle reclaims instances that sit idle for a week. Killercoda’s free Kubernetes playgrounds are great for learning and last an hour a session. Those are allowances and sandboxes, not something you order in the size you need and keep for as long as you like.
So here’s the claim, stated precisely enough to check: for a given amount of vCPU and memory, a Playground is the cheapest Kubernetes environment we could find that you can order on demand, in the size you need, and keep. If you know of a cheaper one, tell us at hello@ankra.ai. We’ll look, and if you’re right, we’ll update this post.
What a Playground is not
A cheap environment is only cheap if nothing about it surprises you later, so here is the fine print.
Playgrounds are for development and evaluation. They’re not for production workloads or personal data, they’re excluded from the SLA, and their storage is local to the Playground host and not backed up. There’s no kubeconfig and no GitOps repository wiring, since those belong to clusters you own. Egress is restricted to what deploys need. Publish with an Ingress, not a LoadBalancer Service, because the host runs no load balancer implementation and the external IP would stay pending forever. And paid Playgrounds share an 8 GB memory budget per organisation by default, which is two Smalls or one Medium out of the box. You can request a higher limit from the size picker, and the Ankra team reviews it.
If you need root on a box, a production cluster or dedicated cores, Hetzner is still where we’d point you, and Ankra will build and run Kubernetes on Hetzner on your own account. Our six-cloud benchmark found the CX23 to be the best value node in the whole test, and nothing in this post changes that. The two aren’t really rivals. Build and try things in a Playground, and when something outgrows development, move it to a cluster of your own. The stacks come across unchanged, because they’re standard Helm charts and Kubernetes manifests and were never Playground dialect to begin with.
Try one
Sign up, open Clusters, choose Create cluster and pick Playground. The free Trial is preselected and the paid sizes sit right below it. From a terminal it’s one command, and ankra cluster playground plans shows live availability and how much of your memory budget is left. The Playground page and the Playground guide cover the rest, and the pricing page always carries the current catalogue.
Get started: Create a free account on Ankra and start a free Playground in about two minutes.
Join our community: Slack
Follow us on: LinkedIn | GitHub
Contact us: hello@ankra.ai
Get the next post in your inbox
Related Posts
Ten Claude Developers, a €350K Bill, and the Kimi K3 Cluster It Takes
What 10 heavy Claude developers cost, why one 8xB200 server cannot run Kimi K3, and the UpCloud cluster that could.
A Managed AI Stack for Developers: Private, Under Your Control, Costed to the Token
Public AI chat quietly leaks IP and breaks NDAs. Deploy a private chat and API on a GPU you own: your data stays home and every token has a known price.