EKS

EKS is a Series B+ decision. The Kubernetes tax is real:

• EKS control plane: $73/month (before any workload runs)
• Minimum viable production cluster: ~$300-500/month (control plane + 2-3 nodes + ALB + NAT)
• Platform engineering time: 20-40% of one engineer’s bandwidth for maintenance, upgrades, RBAC, networking
• Time to first deploy: days/weeks vs hours for ECS Fargate

Choose EKS only when:

• Team > 10-15 engineers with multiple teams needing namespace isolation
• You’re hiring primarily from a Kubernetes-experienced talent pool
• You need advanced scheduling (GPU sharing, bin-packing, multi-tenancy)
• Your compute spend exceeds $20K/month and Karpenter’s optimization justifies the platform cost
• You’re already running K8s and migrating to AWS (don’t re-learn ECS just for AWS)

Startup Cost Traps

1. The $73/month control plane for dev/staging: Many startups create 3 clusters (dev/staging/prod) = $219/month in control planes alone before a single pod runs. At pre-Series A, use namespaces on a single cluster for isolation. Add dedicated clusters at Series B.

2. Managed node groups with oversized instances: Default tutorials use m5.large nodes. A startup running 3-5 microservices needs maybe 2-4 vCPUs total. Use t3.medium nodes or Karpenter with tight resource limits.

3. Add-on sprawl: Each add-on (cert-manager, external-dns, ArgoCD, Datadog, Istio) runs pods that consume node resources. A “production-ready” cluster with common add-ons needs 2-4 vCPUs just for platform components. Budget this separately.

4. Karpenter consolidation disabled: Karpenter defaults to not consolidating nodes. Enable consolidationPolicy: WhenEmptyOrUnderutilized immediately — it’s the primary mechanism for right-sizing your fleet and avoiding paying for idle capacity.

5. Load Balancer per service: Each Kubernetes Service type: LoadBalancer creates an NLB (~$16/month). Use an Ingress controller (ALB Controller or nginx) to share one load balancer across services.

Stage-Specific Recommendations (If You’re Committed to EKS)

Early (1 cluster, <$5K/month compute)

• Single cluster, namespaces for env separation (dev/staging/prod namespaces)
• Karpenter from day one — never use Cluster Autoscaler for new clusters
• Fargate profiles for low-traffic namespaces (dev) to avoid idle node costs
• Skip service mesh (Istio/Linkerd) — use simple K8s Services and Network Policies
• ArgoCD for GitOps from the start — it’s free and prevents kubectl apply drift

Growth ($5K-50K/month compute)

• Separate prod cluster from non-prod
• Karpenter with Spot for stateless workloads (70% savings)
• Pod Identity for IAM (not IRSA) — simpler, fewer moving parts
• Enable control plane logging for security/audit
• Implement PodDisruptionBudgets for all production workloads

Scale ($50K+/month compute)

• Dedicated node groups for isolation (GPU, high-memory, batch)
• Multi-cluster with fleet management (if multi-region needed)
• Full observability stack (Prometheus/Grafana or Datadog)
• Consider EKS Anywhere for hybrid if needed

Counterintuitive Startup Advice

• ECS to EKS migration is easier than EKS to ECS. If unsure, start with ECS. If you later need K8s, your containers and Dockerfiles transfer unchanged — only the orchestration layer changes. The reverse (EKS → ECS) requires removing all K8s-specific config (Helm charts, CRDs, operators).

• A well-run ECS setup is operationally simpler AND cheaper than EKS until $20K/month compute. The K8s ecosystem flexibility only pays off when you have enough services and team size to leverage it.

• Don’t use Helm for everything. Early startups over-invest in templating 3-5 services with complex Helm charts. Plain Kubernetes manifests + Kustomize overlays are more readable and debuggable for small teams. Switch to Helm when you have 10+ services with shared patterns.

• Fargate on EKS is worse than Fargate on ECS. EKS Fargate has more limitations (no DaemonSets, no persistent volumes, no GPUs, higher per-pod overhead, slower scheduling). If you want Fargate simplicity, use ECS. If you want EKS, use managed node groups with Karpenter.

When to Graduate TO EKS

Credits-Specific Guidance

• EKS control plane ($73/month) is covered by credits but is a fixed cost — don’t create clusters you won’t use.
• During credits: experiment with EKS to validate if your team can operate it. This is the time to learn without cost pressure.
• If your team struggles with EKS during the credits period, that’s your answer — switch to ECS before credits expire and you’re paying $500+/month for a platform you can’t operate well.