Nov 28, Paris: AI Health Hack #2 with Alan, Qovery and AWS. Applications close Nov 21

9 ways to stop hand-managing EC2 fleets in 2026, ranked for small teams

A fair ranking of what abstracts away raw EC2 fleet management for a small team: Elastic Beanstalk, ECS on Fargate, ECS Managed Instances, ECS Express Mode, App Runner, EKS Auto Mode, Karpenter, Lightsail and Qovery on your own EKS, judged on who patches, scales and replaces the instances, and what each one costs.

Alessandro Carrano
Head of Product
NOV 19, 2026 · 16 MIN
9 ways to stop hand-managing EC2 fleets in 2026, ranked for small teams

Key points:

  • "Managed EC2" now means three different things on AWS. Some services hide instances completely (Fargate, ECS Express Mode), some run instances AWS patches and replaces for you (ECS Managed Instances, EKS Auto Mode, Elastic Beanstalk managed updates), and some only automate scaling while you keep the AMIs (Karpenter, plain ECS on EC2).
  • For one or two web services, ECS Express Mode is the shortest path in 2026. One API call provisions Fargate tasks, an Application Load Balancer, autoscaling and networking, at no extra charge, and AWS points former App Runner users to it.
  • App Runner is closed to new customers since 30 April 2026. Existing customers keep using it, but AWS plans no new features, so new projects should not start there.
  • For a team running many services, Kubernetes with managed nodes wins on flexibility. EKS Auto Mode patches nodes and caps their lifetime at 21 days for a per-instance fee; a platform like Qovery creates EKS with Karpenter in your own AWS account and adds the deployment layer Auto Mode leaves to you.
  • Lightsail and Elastic Beanstalk still have a place. Lightsail starts at $5 a month for a single box; Beanstalk costs nothing extra and patches its fleet weekly, but major platform moves stay manual.

Qovery · Infrastructure control plane
A control plane for platform teams and their coding agents
Learn more

The question I hear from small teams is rarely "should we use Kubernetes?" It is "we have eight EC2 instances, two of them are pets, nobody remembers which AMI the autoscaling group uses, and patch day is a Friday-night ritual. What takes this off our hands?"

AWS has shipped a lot in the last 14 months that answers exactly that: ECS Managed Instances in September 2025, ECS Express Mode in November 2025, and a retirement notice for App Runner in 2026. I went through each option against the same chores, and ranked them for a team of roughly 3 to 30 engineers with no dedicated DevOps person.

How did we pick and rank the options?

I ranked each option on six concrete criteria, all about the EC2 work a small team wants to stop doing by hand. The ranking favours options that remove the most chores for the least lock-in and cost, for a team without a DevOps engineer.

  1. Who picks and launches instances. Instance types, Availability Zones, capacity when demand spikes.
  2. Who patches the OS and replaces old nodes. The chore that most often slips.
  3. Whether deployment is included. Turning a commit into a running, load-balanced service with rollbacks.
  4. What it costs on top of EC2 or compute. The management fee, if any, from the vendor's own pricing page.
  5. How far it scales with you. Workers, cron jobs, GPUs, many services, several environments.
  6. How portable it is. Whether you can leave without rewriting your deployment model.

All nine options run in your own AWS account. I left out PaaS providers that host apps in their own cloud, because the question is about replacing your EC2 fleet, not moving off AWS.

What does hand-rolling an EC2 fleet involve?

Hand-rolling a fleet means owning six jobs: baking and rotating AMIs, patching the OS, sizing and scaling the Auto Scaling group, replacing unhealthy or retiring instances, wiring load balancers and certificates, and building the deploy pipeline. A managed option is worth adopting in proportion to how many of these it removes.

In practice that looks like a launch template pointing to an AMI someone baked with Packer a year ago, an Auto Scaling group with a CPU target, Systems Manager patch baselines that half the instances follow, and a deploy script that SSHes in or calls CodeDeploy. It works until the person who built it leaves.

Even "managed" container platforms can leave most of this with you. Plain ECS on EC2 with an Auto Scaling group capacity provider adds no ECS fee (AWS ECS pricing: "There is no additional charge for Amazon EC2 compute"), but you still own the ECS-optimized AMI updates, the Auto Scaling group and instance draining. It is the baseline the options below improve on.

Which options abstract away EC2 fleet management, compared side by side?

The table below is the shortlist, ranked for a small team. Each row stands alone: who manages the instances, who patches them, whether deployment is included, what it costs on top, and who it fits best.

RankOptionWho manages instancesOS patchingDeploy pipeline includedExtra fee on top of computeBest for
1ECS Express ModeNone visible (Fargate)AWSPartly: one API call or GitHub Action deploys an imageNoneOne to a few HTTP services, fastest start
2Qovery on your own EKSKarpenter on an EKS cluster Qovery creates and upgrades in your accountNode replacement via Karpenter; cluster upgrades managed by QoveryYes: git push, previews, rollbacksEKS control plane ($0.10/h) plus Qovery planTeams with many services and environments, no DevOps hire
3ECS on FargateNone visibleAWSNo, bring your own CI/CDFargate per-vCPU and per-GB pricingWorkers, jobs and APIs without instances
4ECS Managed InstancesAWS provisions, scales and replacesAWS, every 14 daysNoPer-instance management feeECS shops needing GPUs or specific instance types
5EKS Auto ModeAWS (Karpenter-based)AWS, 21-day max node lifetimeNoPer-instance management fee plus control planeTeams that already know Kubernetes
6Elastic BeanstalkBeanstalk-managed Auto Scaling groupWeekly managed updates (patch and minor)Yes, for supported platformsNoneClassic web apps on a supported runtime
7Karpenter on standard EKSKarpenter, which you install and runYou pick AMIs; Karpenter replaces drifted nodesNoNone (open source)Teams with Kubernetes skills who want full control
8Amazon LightsailYou, one fixed-size box at a timeYouContainer service only, basicFixed bundles from $5/monthSingle small app, predictable bill
9AWS App RunnerNone visibleAWSYesPer-vCPU and per-GB pricingExisting App Runner customers only

I ranked for a 10-person team that wants the most fleet chores gone for the least new tooling. Express Mode comes first because most small teams start with a few HTTP services; Qovery comes second because it only pays off once you run many services and environments.

1. ECS Express Mode: is it the new default for a simple web service?

For one to a handful of HTTP services, yes. ECS Express Mode turns a container image into a load-balanced, autoscaled Fargate service in one API call, at no extra charge. Skip it if you need workers without HTTP, many services sharing complex routing, or GPUs.

AWS's App Runner migration guide describes it well: "With a single API call, you provide a container image and two IAM roles, and Amazon ECS provisions a complete application stack in your AWS account, including an ECS service on Fargate, an Application Load Balancer, auto scaling, and networking." Provisioning takes three to five minutes. It "automatically consolidates up to 25 Express Mode services behind a single ALB when appropriate" (AWS), which keeps load balancer costs down.

What you still own:

  • The build. Express Mode deploys an image. For git-push behaviour, AWS's own guide adds a GitHub Actions workflow with OIDC, ECR and the amazon-ecs-deploy-express-service action.
  • Everything around the app. Databases, queues, secrets, preview environments.
  • Growth. Express Mode creates normal ECS resources, so you can later manage them directly, but at that point you are back to running ECS yourself.

2. Qovery on your own EKS: when do you want a platform on top of Kubernetes?

Qovery fits a small team that runs several services and environments and wants managed provisioning plus deployment in its own AWS account, without hiring someone to run Kubernetes. It is the wrong pick for a single service with steady traffic, where Express Mode does the job with fewer moving parts.

Qovery is an Agentic DevOps platform. It does not host your apps. When you connect your AWS account, Qovery creates the cluster in that account: per the Qovery EKS docs, that means an EKS cluster, a VPC with public and private subnets, NAT gateways, IAM roles, Karpenter, the AWS Load Balancer Controller and the EBS CSI driver, with spot and Graviton support. Setup takes 15 to 30 minutes. If you already run EKS, Qovery installs on it instead (bring your own Kubernetes, 1.31 or later), and the same applies to GCP, Azure, Scaleway or any other cluster.

What it adds on top of the fleet:

  • Deployments. Git push to deploy, rollbacks, environment cloning and a preview environment per pull request.
  • Scaling. Karpenter picks instance types and consolidates nodes, with spot and on-demand node pools you configure in the Console.
  • Day 2. Cluster upgrades handled by Qovery, cost visibility, RBAC and audit logs.
  • Agents. Coding agents deploy through the Qovery API with scoped tokens instead of AWS keys.

The cost is the EKS control plane, $0.10 per cluster per hour (about $73 a month), your EC2 nodes, and a Qovery plan. That control plane fee is why I would not put one tiny app on Kubernetes.

Didn't Qovery have an EC2 option? It did. In 2022 we announced Qovery EC2, a single-instance mode we only recommended for development and hobby projects, because one instance in one Availability Zone goes down with that zone. It is not part of the current Qovery documentation, which covers EKS with Karpenter, so treat that post as history.

3. ECS on Fargate: when should you remove instances entirely?

Fargate fits teams that want zero instances to think about for APIs, workers and scheduled jobs, and accept a per-vCPU premium for it. It is a poor fit for GPU workloads and for steady, high-utilisation services where reserved EC2 capacity is much cheaper.

In US East (N. Virginia), Linux/x86 Fargate costs $0.000011244 per vCPU-second and $0.000001235 per GB-second (AWS Fargate pricing). That is about $0.0405 per vCPU-hour, so a 1 vCPU, 2 GB task running all month costs roughly $36. ARM is about 20% cheaper per vCPU, and Fargate Spot offers "up to a 70% discount" for interruptible work.

Fargate removes every fleet chore. It removes none of the deployment chores: task definitions, services, load balancer rules, IAM roles and the CI pipeline are yours, usually in Terraform. That is the gap Express Mode closes for simple services.

Managed provisioning and deploys, in your own AWS account.
Qovery creates and runs EKS with Karpenter in your AWS account, then gives your team git-push deploys and preview environments, with no DevOps hire needed. Also on GCP, Azure, Scaleway or your existing cluster.

4. ECS Managed Instances: what if you need EC2 but not the chores?

ECS Managed Instances fit teams already on ECS that need specific instance types (GPUs, network-optimised, burstable) but want AWS to provision, scale and patch them. Avoid it if you don't need EC2 features, since Fargate removes more work.

Announced on 30 September 2025, it lets "Amazon ECS automatically provision, configure and operate" instances, with "regular security patching initiated every 14 days". The patching docs describe a 14 to 21 day lifecycle: ECS drains an instance from day 14 and terminates it by day 21 at the latest, respecting your service's maximumPercent and minimumHealthyPercent, and you can align maintenance with EC2 event windows.

The cost is a per-instance management fee on top of EC2. AWS's ECS pricing example for US West (Oregon) lists $0.037 per hour for a c6a.2xlarge (about $27 a month) and $0.021 for an m5a.xlarge. As with Fargate, the deployment layer stays with you.

5. EKS Auto Mode: does it make Kubernetes nodes hands-off?

EKS Auto Mode makes nodes close to hands-off: AWS picks, patches and replaces them, using Karpenter underneath. It suits teams that already write Kubernetes manifests. It does not give developers a deployment workflow, and the per-instance fee adds up on large fleets.

From the EKS Auto Mode docs: nodes run Bottlerocket-based AMIs with read-only root file systems and no SSH or SSM access, and "nodes launched by EKS Auto Mode have a maximum lifetime of 21 days (which you can reduce), after which they are automatically replaced". AWS also manages load balancer integration, pod networking, network policy enforcement, DNS and EBS storage. You keep the application containers, the VPC and cluster configuration.

The EKS pricing page bills Auto Mode per instance "in addition to the Amazon EC2 instance price". Its example fee for a c6a.2xlarge is $0.03672 an hour, on top of the $0.10 hourly control plane.

The catch for a small team is everything above the nodes: Deployments, Ingress, Helm charts, secrets, CI/CD, preview environments and upgrades of your own add-ons. We compared the options around it in the best alternatives to EKS Auto Mode.

6. Elastic Beanstalk: is it still a good choice in 2026?

Elastic Beanstalk is still a reasonable choice for a classic web app on a supported runtime that wants patched instances and simple deploys at no extra cost. It ages badly once you run many services, containers with sidecars, or need to move across major platform versions.

Beanstalk pricing is blunt: "There is no additional charge for AWS Elastic Beanstalk." Its managed platform updates apply patch and minor platform versions in a two-hour weekly window, always as immutable updates that launch a fresh set of instances before terminating the old one, and you can force a weekly instance replacement even without a new version.

The limits are spelled out in the same page. Managed updates "don't support updates across platform branches", so moving to a new OS or runtime major version is your job. AWS also warns that more than 200 environments in one account with closely scheduled updates can hit throttling. If you are weighing the move from Beanstalk to Kubernetes, this five-option comparison has the migration estimates.

7. Karpenter on standard EKS: is autoscaling enough?

Karpenter is enough if you have Kubernetes skills in-house and want full control of node pools without a per-instance fee. It automates choosing, consolidating and replacing instances, but you still install and upgrade Karpenter itself, choose AMIs and run everything above the nodes.

Karpenter reached v1.0 in August 2024 (AWS). Its disruption controls cover consolidation, drift and expiration, with a default expireAfter of 720 hours (30 days). Karpenter never patches an AMI in place. When your amiSelectorTerms resolve to a new image, it marks nodes as drifted and replaces them. That is a good model, but someone has to own the selector and test the new AMI.

Karpenter is also the engine under EKS Auto Mode and under Qovery-managed EKS clusters. The difference between the three is who runs it.

8. Amazon Lightsail: when is a simple VPS the right answer?

Lightsail is the right answer for a single small app with a predictable bill and no scaling needs. Each instance is a box you size and patch yourself, so Lightsail stops fitting the moment you need several services or autoscaling.

Lightsail pricing starts at $5 a month for a Linux bundle with 0.5 GB of memory, 2 vCPUs, 20 GB of SSD and 1 TB of transfer. Its container service starts at $7 a month for a Nano node (0.25 shared vCPU, 512 MB), a managed database at $15 and a load balancer at $18. For a side service that never grows, nothing on this list is cheaper or simpler.

9. AWS App Runner: should you still start a new project on it?

No. AWS App Runner is "no longer open to new customers", effective 30 April 2026, and AWS states it does "not plan to introduce new features" (AWS App Runner availability change). Existing customers can keep running and creating services.

App Runner was the closest thing AWS had to "give me a URL for this container", and it worked well for that. AWS now recommends ECS Express Mode as the replacement and publishes a blue/green migration guide using Route 53 weighted records. If you are on App Runner today there is no emergency, but I would plan the move before you need a feature it will never get.

Which option should a small team pick?

Pick by the number of services you run and whether anyone on the team knows Kubernetes. Fewer than five HTTP services and no Kubernetes experience: ECS Express Mode. Many services, several environments and no DevOps hire: Kubernetes with managed nodes plus a deployment layer, such as Qovery on your own EKS.

Your situationPickWhy
One small app, fixed budgetLightsail$5 to $10 a month, nothing to learn
Classic web app on a supported runtime, already on BeanstalkStay on Elastic BeanstalkFree weekly patching; move when you hit its limits
One to five HTTP services, no Kubernetes skillsECS Express ModeOne API call, Fargate, shared ALB, no fee
APIs plus workers and cron jobs, no GPUsECS on FargateNo instances at all; you own the pipeline
ECS shop that needs GPUs or specific instance typesECS Managed InstancesAWS patches every 14 days
Kubernetes-fluent team, wants AWS to own nodesEKS Auto Mode21-day node rotation, managed add-ons
Kubernetes-fluent team, wants full control, no feeKarpenter on EKSOpen source, you run it
Many services and environments, no DevOps hire, own AWS accountQovery on your own EKSManaged cluster plus git-push deploys and previews

Two mistakes I see often. First, moving a single steady service to Kubernetes for the CV, where the $73 control plane and the learning curve buy you nothing there. Second, staying on hand-built EC2 because "it works": it works until the one person who knows the AMI pipeline goes on holiday during a CVE.

For a wider look at the AWS container services, see App Runner vs ECS vs EKS.

How do I set this up with Qovery from Claude Code or Cursor?

Install the Qovery Agent Skills, then ask your coding agent to deploy your app to your AWS account with Qovery. The agent creates the account and cluster setup with you, writes a Dockerfile if one is missing, and deploys, asking for approval at each step.

BASH|1. Agent Skills for every compatible agent (Claude Code, Cursor, Codex, OpenCode, Gemini CLI)
curl -fsSL https://skill.qovery.com/install.sh | bash

# 2. Or scoped to the current project only
curl -fsSL https://skill.qovery.com/install.sh | bash -s -- --project

# 3. MCP Server for Claude Code over OAuth, read-only by default
claude mcp add --transport http qovery https://mcp.qovery.com/mcp --callback-port 4242

# 4. Write access, only when you explicitly want it (also needs write access enabled in the Qovery Console)
claude mcp add --transport http qovery "https://mcp.qovery.com/mcp?read_write=true" --callback-port 4242

Command 1 or 2 asks you to approve running the install script. Command 3 opens an OAuth login in your browser and gives the agent read-only access. Then ask in plain language: "deploy this app to my AWS account with Qovery". The skills (qovery-onboard, qovery-deploy and others) handle setup and deployment; the MCP Server operates what already exists. The AI agent quickstart and the Claude Code integration cover the details.

Frequently asked questions
What abstracts away raw EC2 fleet management so a small team gets managed provisioning and deployment?

For a few HTTP services, ECS Express Mode provisions Fargate tasks, a load balancer and autoscaling from one API call at no extra charge. For many services and environments, a platform like Qovery creates EKS with Karpenter in your own AWS account and adds git-push deploys and preview environments. Elastic Beanstalk, ECS Managed Instances and EKS Auto Mode also manage instances, but leave more of the deployment work to you.

Is AWS App Runner still available for new projects?

No. AWS closed App Runner to new customers on 30 April 2026. Existing customers can keep using it and create new services, but AWS plans no new features and recommends ECS Express Mode for new and migrating workloads.

What is the difference between ECS Managed Instances and EKS Auto Mode?

Both run EC2 instances that AWS provisions, scales and patches, and both charge a per-instance management fee on top of EC2. ECS Managed Instances drain and replace instances between day 14 and day 21 and run ECS tasks. EKS Auto Mode caps node lifetime at 21 days, uses Karpenter and Bottlerocket, and runs Kubernetes pods, so it also needs the $0.10 per hour EKS control plane.

Does Elastic Beanstalk patch my EC2 instances automatically?

Yes, for patch and minor platform versions. Managed platform updates run in a weekly two-hour window as immutable updates, launching new instances before terminating the old ones. Moving to a new platform branch, such as a new OS or runtime major version, is not automatic.

Does Qovery still offer an EC2-only option?

Qovery announced a single-instance EC2 mode in 2022 for development and hobby projects. It does not appear in the current Qovery documentation, which provisions EKS with Karpenter in your AWS account or installs on a Kubernetes cluster you already run. Treat the 2022 announcement as historical.

Is Kubernetes overkill for a small team that wants to leave hand-managed EC2?

For one or two steady services, usually yes: ECS Express Mode or Fargate removes the fleet with less to learn. Kubernetes pays off when you run many services, workers and environments, and a managed layer like EKS Auto Mode or Qovery keeps the team from having to operate the cluster itself.

Alessandro Carrano
About the author
Alessandro Carrano

Alessandro leads product at Qovery. He drives the changelog, roadmap, and product strategy - turning customer feedback into platform capabilities.

Next step

Managed provisioning and deploys, in your own AWS account.

Qovery creates and runs EKS with Karpenter in your AWS account, then gives your team git-push deploys and preview environments, with no DevOps hire needed. Also on GCP, Azure, Scaleway or your existing cluster.