Cost Optimization

Hetzner has no managed Kubernetes. Eight ways to run it in 2026, from kube-hetzner and Talos to Syself, Cloudfleet and Ankra, and what each leaves to you.

Ankra's own platform now runs in Helsinki. The timeline, the bill, the benchmark numbers, and the places UpCloud let us down, from the team that did the move.

Every Ankra Playground size costs less than the Hetzner server of the same shape, with Kubernetes, TLS and storage included. The math, and the fine print.

What 10 heavy Claude developers cost, why one 8xB200 server cannot run Kimi K3, and the UpCloud cluster that could.

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.

We ran the same k6, iperf3 and fio suite on EKS, GKE and Hetzner and lined the results up against UpCloud, OVH and DigitalOcean. The cheapest node in the test won more than it had any right to.

Two identical cloud VMs measured 2.8x apart on CPU; one failed etcd's fsync bar. A 15-minute in-cluster benchmark suite catches this before production does.

UpCloud vs DigitalOcean vs OVH on identical Kubernetes clusters. UpCloud led on storage and network, OVH on latency, DigitalOcean had the slowest tails.

Three releases a year, 14 months of patches, and a 6x cluster fee when you lapse. How to turn Kubernetes upgrades from a dreaded project into a routine.

On June 15, Hetzner raised cloud prices, with some instance families more than doubling. A worked-numbers look at what a production Kubernetes cluster costs now, and why the real lesson is portability, not provider loyalty.

'Don't use Kubernetes until you have a platform team' was good advice in 2019. The premises changed. What a two-to-six person company should set up in week one, what to deliberately skip, and what it actually costs: about zero.