“30MB, 7x smaller”
traces back to nothing.

Search “Kubernetes vs Nomad resource overhead” and you'll get the same handful of numbers on a rotating cast of SEO-farm domains: Nomad's single binary “consumes just 30MB of RAM,” is “7x smaller than kubelet,” and Kubernetes is “3.6x more expensive than Nomad” on identical hardware. None of these pages disclose what was actually measured, what hardware it ran on, what version of either tool, or under what load. They just repeat each other. That's not a benchmark — it's a number with no chain of custody, and I'm not going to be the next site to repeat it.

Instead, this post is built entirely from primary sources: the official kubeadm documentation, the official K3s requirements page, HashiCorp's own Nomad production installation requirements and architecture explainer, current published pricing pages for the four major managed Kubernetes services, and HashiCorp's published case study on Cloudflare's Nomad deployment.

Per HashiCorp's own architecture guide, Kubernetes splits control-plane duties across etcd (state), kube-apiserver (API), kube-scheduler, and kube-controller-manager, plus kubelet and kube-proxy on every worker node. Nomad collapses server-side responsibilities into a single Go binary that implements Raft internally instead of running a separate etcd cluster, and the same binary runs in either server or client mode. Nomad also deliberately does less: it focuses on scheduling and cluster management and leaves networking and secrets to Consul and Vault rather than building them in, whereas Kubernetes bundles more functionality (networking, service discovery, secrets) into the core project and its ecosystem of controllers.

Docker Compose isn't in the same category as either. Per Docker's own documentation, Compose is a single-host tool for defining and running multi-container applications — there is no control plane, no multi-node scheduling, and no built-in self-healing across machines. Docker's own guides position it as the on-ramp for small deployments and local development, with Swarm or Kubernetes as the next step once you need more than one host.

Rather than an anonymous benchmark, HashiCorp's published case study on Cloudflare gives a named, attributable production reason to prefer Nomad for a specific shape of workload: Cloudflare needed to reliably run a single instance of a scheduling binary with resource isolation inside each of its 200+ distributed edge data centers, and wanted something lightweight and simple to provision repeatedly at that scale rather than a full control plane per site. The case study also notes Nomad has been proven to scale to 10,000+ node clusters with a comparatively small number of server nodes managing them — a genuinely different scaling profile than most teams will ever need, but real evidence the architecture holds up at the extreme end.

This is a narrower case than “Nomad is better than Kubernetes” — it's evidence for a specific pattern (many small, geographically distributed sites, each needing simple resource isolation) rather than a general verdict.

Docker Composeis the right default for a single server, a side project, or local development — per Docker's own docs, that's exactly the use case it's designed for, and reaching for an orchestrator before you have more than one host to coordinate is solving a problem you don't have yet.

Nomadis the strongest fit once you need multi-machine scheduling but want to keep the operational surface small — one binary, no separate etcd cluster to run, and (per the Cloudflare case study) a track record at genuinely large, geographically distributed scale. It's also a good fit if you need to schedule non-containerized workloads (batch jobs, VMs, raw binaries) alongside containers, which is a first-class Nomad feature and not a natural Kubernetes one.

Kubernetes(managed, not self-hosted, for almost everyone) wins once you need its ecosystem: Helm charts for nearly every piece of software you'd want to run, a much larger hiring pool of people who already know it, and GitOps/operator tooling that assumes Kubernetes as the target. The $40-73/month managed control-plane fee is a small price for not operating that complexity yourself.