Scaleway Kapsule

What Scaleway Kapsule actually is

Kapsule is Scaleway's managed Kubernetes engine: you create a cluster, Scaleway runs the control plane, installs the node agents, wires the cloud controller manager, cluster autoscaler and CSI drivers, and keeps Kubernetes patched. Worker nodes are ordinary Scaleway Instances grouped into pools; the sibling product, Kosmos, uses the same managed control plane but accepts nodes from any cloud or on-premises (external nodes are billed at zero), with Kilo as CNI. In this review we focus on Kapsule, the pure-Scaleway variant.

The offer sits inside the wider Scaleway platform we review separately — same account, same regions, same per-hour billing — and the appeal for European teams is the combination: managed Kubernetes whose control plane, nodes, block storage and object storage all live under French and EU jurisdiction. It is CNCF-certified conformant Kubernetes (vanilla, not a fork), available in Paris, Amsterdam, Warsaw and — since 2026 — Milan.

Version handling is one of the better-documented parts. Scaleway publishes a release calendar with exact deprecation and end-of-support dates, ships new minor versions within days of upstream certification (1.37 landed in August 2026), and guarantees each minor version at least twelve months of support (fourteen months since 1.29). Patch auto-upgrade can be enabled per cluster; when a version hits end of support, Scaleway upgrades the cluster to the next supported minor within 30 days and notifies you by ticket. Like it or not, that is a clear policy — we prefer it to silent drift.

Pricing: what you really pay

The economics are simple and in our view the strongest argument for Kapsule: the shared control plane is free. You pay for nodes at standard Instance rates and nothing for the management layer. A three-node cluster on the cheapest eligible node (2 vCPU, 2GB) works out around €19.65 per month total — roughly €6.55 per node — versus ~€74 for the same shape on Azure AKS and ~€33 self-managed on Hetzner once you account for the control-plane nodes you must run yourself. Against AWS EKS or GKE at $0.10 per cluster per hour (about $73/month just for the control plane), a small Kapsule cluster is dramatically cheaper.

For heavy or business-critical workloads there is a dedicated control plane tier, priced from about €0.11 per cluster per hour — roughly €80/month — which raises the maximum cluster size from 150 nodes to as many as 500. Node prices follow the normal instance ladder (DEV1-S at €0.0088/hour, about €6.42/month, upward through PLAY2, PRO2 and the rest), billed per hour with no cluster premium. GPU node pools are available and billed from creation, including the setup window before the NVIDIA GPU Operator is ready — a small billing detail the docs state plainly.

Where costs hide: Block Storage volumes and snapshots bill separately (snapshots at €0.000049/GB/hour), public Load Balancers and flexible IPs add recurring charges, and egress is metered per GB on some classes (one comparison site records "per-GB, verify per zone" where OVHcloud bundles unlimited instance traffic). Node pools also bill while idle — a forgotten four-node pool costs the same at 3am as at noon, so enable the cluster autoscaler and use it.

Performance & infrastructure

No hands-on claims here: our dated benchmark suite covers Hetzner, DigitalOcean, UpCloud, Vultr and Linode instances, not Scaleway or Kapsule, so this is assessment from documentation and third-party analysis rather than our measurements.

What is solid on paper: all the components are managed — CCM, autoscaler, CoreDNS, metrics server, CSI drivers for Block and File Storage — with health checks that restart or replace an unresponsive node after 15 minutes. VPC integration keeps pod and node traffic private, there is a Helm-based application library in the console, and CNI choice is cilium or calico (Kosmos uses Kilo), with containerd as the only runtime. API, CLI and Terraform coverage is complete, and the changelog is public and specific down to component versions (Calico v3.32.x, Cilium v1.19/1.20, runc, containerd) — the kind of detail that makes incident forensics possible.

The documented limits deserve as much attention as the features. Maximum 110 pods per node (configurable) and 15 PVCs per node, per upstream recommendations. Dual-stack IPv4/IPv6 is not yet available — a real gap in 2026. ReadWriteMany and ReadOnlyMany access modes are not yet supported, so multi-attach filesystem workloads need another answer. Most important: the control plane's network access sits behind a Load Balancer in one availability zone, and Scaleway states plainly that a global failure of that zone makes the control plane unreachable — you cannot fix this yourself, and it is exactly the kind of single-point dependency that a managed service is supposed to remove.

Reliability signals otherwise are decent: multi-AZ node pools are supported (a Terraform tutorial walks through a 3-AZ cluster), Milan is now a full three-AZ region, and the parent platform carries ISO 27001 and HDS certifications. Scaleway does not publish a Kapsule-specific SLA figure that we could confirm; platform SLAs range from 99.5% (GPU instances) to 99.99% (PRO2). If you need a contractual number, ask before buying — do not infer one.

Where it falls short

The single-AZ control plane ingress is the limitation we would weigh hardest for production workloads. It is honestly documented, which we respect, but "honestly documented single point of failure" is still a single point of failure. Compare with GKE's regional control planes before committing anything that cannot tolerate an unreachable API server for the duration of an incident.

The feature gaps add up for advanced users: no dual-stack, no RWX storage, Kubernetes-native policy is fine but there is no equivalent of GKE Autopilot or EKS Fargate for serverless pods, and multi-cluster or fleet tooling is absent. The ecosystem is smaller — fewer third-party integrations, fewer community guides, and support is Scaleway's standard ticket queue rather than a Kubernetes-specialized desk. If your team lives in GKE or EKS tooling, expect friction on the margins.

Forced upgrades at end-of-support are the policy trade-off: clusters move to the next minor within 30 days of EOL whether you have tested that version or not. The release calendar lets you plan, but a team that pins versions for a reason (a webhook that breaks on admission changes, say) needs to stay ahead of it. And while node pricing is published, the full monthly number depends on storage, IPs, load balancers and egress — the cluster price is never just the cluster.

Day-to-day operations are smoother than the price suggests. Node pools support labels, taints and startup taints, isolation levels are configurable per pool (fully isolated or controlled isolation), and security groups can be attached at pool level. The console exposes logs and metrics through the Observability Cockpit, and the community Slack plus a fairly complete tutorial library cover the common paths — multi-AZ clusters with Terraform, log shipping, ingress setup — without needing a support ticket.

Who should pick this

Our pick for European teams that want managed Kubernetes without paying a control-plane tax: startups standardizing on Scaleway Instances, GDPR-bound workloads that must stay in the EU, and platform teams that would rather spend effort on their applications than on etcd upgrades. The free shared control plane makes small clusters genuinely cheap, Terraform support makes multi-environment setups tractable, and Kosmos is an interesting escape hatch if you later want to attach nodes from elsewhere without rebuilding.

Also a reasonable landing spot for teams leaving AKS or EKS on cost: the per-cluster overhead is zero versus roughly $73/month at AWS, and the managed experience covers the parts that actually break (node healing, patching, CSI). Bring your own observability — Cockpit exists but is basic.

We would not pick it for workloads needing RWX volumes or IPv6 today, for organizations that require a published SLA with a support organization behind it, or for anyone whose architecture assumes a highly available API server during a zonal failure.

Key takeaways

• Managed CNCF-conformant Kubernetes in Paris, Amsterdam, Warsaw and Milan, run by Scaleway (iliad group, France) under EU law.
• Free shared control plane — you pay only for nodes at standard Instance rates; a 3-node DEV1-S cluster runs about €19.65/month.
• Dedicated control plane from roughly €0.11/hour raises cluster limits from 150 to 500 nodes; GPU node pools are supported.
• Clear version policy: 14-month support per minor, published deprecation calendar, optional patch auto-upgrade, forced upgrades within 30 days of end of support.
• Documented gaps: no dual-stack IPv4/IPv6, no RWX storage, and a control plane behind a single-zone Load Balancer.
• Strong value for EU teams that can live with those gaps; SLA-critical or multi-cloud-native workloads should compare with GKE, AKS and EKS first.

Sources

• https://www.scaleway.com/en/kubernetes-kapsule/
• https://www.scaleway.com/en/docs/kubernetes/reference-content/version-support-policy
• https://www.scaleway.com/en/developers/api/kubernetes
• https://www.scaleway.com/en/docs/kubernetes/
• https://www.qovery.com/blog/eu-cloud-providers-managed-kubernetes-2026-sovereignty-checklist
• https://www.eucloudcost.com/compare/hetzner-vs-scaleway/
• https://www.eucloudcost.com/compare/azure-vs-scaleway/
• https://www.scaleway.com/en/docs/instances/faq/
• https://www.scaleway.com/en/pricing/storage/
• https://scaleway.com/en/docs/account/reference-content/products-availability
• http://en.wikipedia.org/wiki/Scaleway