Hetzner Kubernetes Math Still Works After the Price Hikes

A post that hit the Hacker News front page this week makes a familiar pitch: a highly available Kubernetes cluster — three control-plane nodes, three workers, two load balancers, spread across three Hetzner Cloud zones — for €47.95 a month before VAT. It's a good walkthrough of what has quietly become the community-standard stack for budget Kubernetes. It's also, through no fault of the author, already out of date. The post was written against 2025 prices, and Hetzner has raised cloud prices twice since April.

The interesting question isn't whether the tutorial works. It does. It's whether the math that made "Kubernetes on Hetzner" a genre of blog post still holds in mid-2026. Short answer: yes, but the margin is shrinking, and the invoice was never the real cost anyway.

The stack everyone converged on

The setup in question uses k3s, Rancher's lightweight CNCF-certified distribution, deployed via the kube-hetzner Terraform module onto openSUSE MicroOS. That combination isn't an idiosyncratic choice; it's the default answer. The module sits at around 3.9k GitHub stars and handles the parts that used to make DIY Kubernetes miserable: Hetzner's cloud controller manager for provisioning load balancers, the Hetzner CSI driver for block storage, kured plus the System Upgrade Controller for automated rolling OS and k3s upgrades, and an immutable base OS with atomic rollbacks so a bad update doesn't strand a node.

That last part matters more than it sounds. The historical objection to self-hosted Kubernetes was never day 1 — kubeadm init has been easy for years — it was the drumbeat of upgrades, reboots, and certificate rotations that a managed control plane absorbs for you. The MicroOS-plus-kured pattern automates most of that drumbeat. Not all of it, which we'll get to.

If you'd rather not run a general-purpose OS at all, Talos Linux on Hetzner is the other serious contender, and Vito Botta's hetzner-k3s CLI covers people allergic to Terraform. But kube-hetzner is where the community mass is, and for once the popular option is also the defensible one.

Redo the math with 2026 prices

Here's what changed. Hetzner announced a broad price increase in February, effective April 1, 2026 — roughly 30–37% on cloud plans in Germany and Finland, and it applied to existing resources, not just new orders. Then on June 15 it repriced again for new orders and rescales: the Intel and ARM shared lines rose another ~30–40%, and the AMD shared (CPX) and dedicated vCPU lines roughly doubled or worse.

Concretely: the CX23 the post uses for control-plane nodes was €3.49/month when it was written. New orders now pay €5.49. The CX33 workers went from €5.49 to €8.49. The compute portion of that reference cluster went from about €27 to about €42 a month, a 56% jump, and the full build — six nodes, two load balancers, public IPv4s — now lands somewhere near €57 depending on the small line items. There's also a nasty operational footnote: existing instances keep their old rate, but resizing one reprices it at the new rate. For a cluster whose scaling story is "rescale the node pool," that grandfathering is worth less than it looks.

So is the pitch dead? Not close. An EKS control plane costs $0.10/hour — about $73 a month — before you've attached a single node, and AWS's on-demand instances and $0.09/GB egress do the rest of the damage. Hetzner's EU plans include 20TB of traffic. For anything bandwidth-heavy, egress alone decides the argument.

The fairer comparison is the providers with free control planes: AKS, DigitalOcean, Scaleway, OVHcloud. Against those, Hetzner's edge is purely node price and included traffic, and after two hikes it's an edge, not a chasm. A DOKS cluster with comparable capacity costs maybe 2–3x, not 10x — and somebody else carries the pager for etcd.

What you're actually buying

That's the honest frame for this decision: you're not buying cheap Kubernetes, you're selling your own SRE hours at a discount. kube-hetzner automates upgrades, but when a MicroOS transactional update interacts badly with a k3s release, or Terraform state drifts from reality, or an etcd member gets wedged during a zone hiccup, there's no support ticket to file. The module's answer to disaster recovery — k3s's built-in etcd snapshots shipped to S3-compatible storage — is genuinely solid, but "solid" assumes someone on your team has restored from it at least once before doing it at 3 a.m.

My rule of thumb: this setup is a great fit if you have at least one person who's comfortable debugging Kubernetes itself (not just deploying to it), your SLOs tolerate an occasional hour of self-inflicted downtime, and your workload is EU-based — Hetzner's US locations cost ~20% more and include 1TB of traffic instead of 20, which quietly guts the value proposition. It's a bad fit for compliance-heavy shops and for teams where Kubernetes knowledge lives in exactly one head.

The uncomfortable trend line

The part of this story nobody on the HN thread wants to sit with: Hetzner raised prices multiple times in 2026, citing hardware and infrastructure costs, and applied at least one round to existing contracts. The entire genre of "cut my infra bill 75% with Hetzner" content assumes a stable arbitrage between hyperscaler margins and bare-cost EU hosting. That gap is real, but it's a market condition, not a law of physics, and 2026 is the year it visibly started to compress.

Which is, ironically, the best argument for this stack. A k3s cluster defined in Terraform against commodity VMs is about as portable as infrastructure gets — the same module patterns work anywhere with a cloud controller manager, and Talos-based variants are even less coupled. If you build on EKS, AWS's pricing is your pricing forever. If you build on kube-hetzner and the math stops working, you're a terraform apply and an etcd restore away from somewhere else. Do it for the €57 invoice if you like, but the exit option is the part that compounds.