AWS Compute Compared: EC2, ECS, EKS, Fargate, Lambda, App Runner

AWS gives you six main compute options. The choice depends on workload shape, your appetite for operational burden, and cost model. The spectrum runs from most-managed (Lambda) to most-control (EC2).

Lambda — event-driven, spiky, short

• You provide a function; AWS provides everything else — servers, scaling, patching
• Pay per invocation plus execution time; cheapest at low usage; gets expensive past sustained 1000+ RPS
• Limits: 15-minute max execution, 10GB memory, cold starts
• Java cold starts are the main pain — 1 to 3 seconds; mitigate with SnapStart (drops to ~200ms), provisioned concurrency (always warm but you pay for idle), or GraalVM native image
• Best for API Gateway endpoints with bursty traffic, S3 event triggers, scheduled tasks; not great for sustained high-RPS production APIs

App Runner — stateless web services, simplest container option

• Push a Dockerfile or image, AWS gives you an HTTPS URL with auto-scaling
• Lacks advanced features like VPC complexity, sidecars; outgrow it = migrate to Fargate

Fargate — serverless containers

• Containers running on ECS or EKS without managing the underlying EC2 instances
• Slight premium over EC2 but no node management; perfect when you want containers without K8s overhead
• Fargate Spot exists but is less cost-effective than EC2 Spot — if you need real Spot pricing, ECS on EC2 or EKS on EC2

ECS on EC2 — cost-optimized container fleet

• ECS orchestrator with EC2 instances you manage; capacity providers let you mix On-Demand and Spot
• Spot saves 50 to 90 percent for interruption-tolerant workloads; Graviton (ARM) instances 20 to 40 percent cheaper than x86
• Best when scale is high enough that the Fargate premium adds up

EKS — Kubernetes ecosystem, multi-cloud portability

• AWS-managed control plane at $0.10 per hour per cluster ($73 per month); you (or Fargate) run the data plane
• Standard K8s manifests run on GKE, AKS, on-prem — the portability story
• Best when you need K8s features (CRDs, Operators, service mesh) and have the ops to manage the complexity
• Anti-pattern: 5-engineer startup adopting EKS for 3 services; over-engineering

EC2 — raw VMs, full control

• The escape hatch — GPU workloads for ML, HPC, specialized hardware (FPGA, Inferentia), Windows GUI, software that doesn't containerize cleanly
• You manage everything — OS patches, AMIs, scaling, monitoring

The cost crossovers everyone misses

• Lambda is cheaper than Fargate at low utilization; the crossover is around 30 to 50 percent sustained utilization
• Fargate is easier than EC2 at low scale; EC2 wins on cost at high scale with Spot, Reserved, and Graviton
• EKS control plane at $73 per month per cluster — negligible at scale, significant for many small clusters

Vendor lock-in is real

• Lambda is the most locked-in — code calls AWS SDK directly, uses AWS-specific event payloads, integrates tightly with API Gateway; migration means rewriting dispatch
• EKS is the most portable — same K8s manifests run anywhere; the open-source standard
• Factor lock-in into long-term strategy alongside cost and ops burden

Default recommendation by stage

• Startup with bursty workloads: Lambda + App Runner; minimal ops, pay only when used
• Mid-size with steady workloads: Fargate or ECS on EC2; balance cost and ops
• Mature org with K8s expertise and multi-cloud goals: EKS
• Specialized needs: EC2 with the right instance type