Could PicoPaaS run somewhere else?

We priced a port to Azure and AWS and looked at making apps Kubernetes CRDs. The verdict is that we are staying on DigitalOcean, with AWS as plan B.

Share

The question came up: could PicoPaaS run somewhere other than DigitalOcean? We wrote two research studies, one for Azure and one for AWS, plus a third on whether the platform should be a Kubernetes CRD. Nothing was provisioned; this is all desk research. Prices are as of October 2026 and are list prices.

The cost table

Today we pay about $75 a month on DigitalOcean. Like for like on a managed Kubernetes service costs more elsewhere. Running our own k3s on a VM brings it back down.

OptionMonthly (as of October 2026)
DigitalOcean (today)about $75
Azure, AKSabout $121
Azure, self-run k3sabout $68
AWS, EKSabout $155
AWS, self-run k3sabout $62

The managed variants lose mostly to fixed costs: AKS system pools cannot use cheap burstable VMs, EKS charges for the control plane, and both need a paid load balancer. The self-run k3s variants are cheaper, but then we run the control plane.

Where the numbers come from: most lines were pulled from the providers' public price lists and some are estimates that need a spike before anyone relies on them. Treat the table as the right order of magnitude, not a quote. We also compared region choices to our current compute location, and the closest equivalents were used on each side.

Azure: the object store is the catch

Azure Blob Storage has no S3 API. Our contract with ourselves is that blob storage is S3-compatible, so Azure means a second storage implementation and a decision about that contract. That is the biggest single item in an estimated 26 agent-days of work.

The rest of the Azure port is mostly infrastructure and workflow rewrites: Terraform, bootstrap, the load balancer and DNS integration, and the cluster autoscaling story. The recommendation from that study was to do only the cheap preparatory refactor, putting the storage and cluster providers behind clean interfaces, unless credits cover a full move.

AWS keeps S3

AWS is the better fallback. S3 is native, so the storage contract is untouched, and the estimate drops to about 20 agent-days. It is still more expensive on managed Kubernetes and egress is metered, so a hostile app could run up a bill unless we add limits.

There are further AWS costs to watch: extended-support fees for an un-upgraded EKS cluster, and node groups that do not upgrade themselves. Self-running k3s avoids those but puts the upgrade calendar in our hands.

A security win either way

Both clouds offer workload identity (OIDC): pods get scoped, short-lived credentials. Today the platform holds account-wide credentials. If we ever move, that is a real improvement. It also tells us to look at narrowing what we have now.

Should it be a CRD?

A third study asked whether apps should be Kubernetes custom resources. The finding is that the reconciler already behaves like an operator: it watches desired state and converges the cluster. Making that a true CRD would mean giving tenants Kubernetes credentials, which is the wrong trust boundary for a multi-tenant platform. GitOps is also not a goal here. The programmatic path is the API, and a Terraform provider later.

Verdict

We are staying on DigitalOcean. It is the cheapest managed option, the contract holds as written, and the numbers do not justify a port. If credits or circumstances change, AWS is the plan B. Meanwhile the cheap parts of the studies, like narrowing credentials, are worth doing anyway.