Dec 1, live webinar: Agents in production, live: who gives the agent production access?

How to Modernize a Hybrid AWS and VMware Setup to Cloud-Native Infrastructure Without Disrupting Production

A phased plan for teams that run VMware on-premise next to AWS: group workloads, wire Direct Connect or VPN, pick rehost, VMware Cloud on AWS, Amazon EVS or containers on EKS per workload, move the data with CDC, and cut over with a rollback you have already rehearsed.

Guillaume Da Silva
Staff Solutions Architect
NOV 22, 2026 · 17 MIN
How to Modernize a Hybrid AWS and VMware Setup to Cloud-Native Infrastructure Without Disrupting Production

Key points:

  • Modernize a hybrid AWS and VMware estate in waves, never in one cutover. Group VMs by dependency, connect the data center to AWS with private networking, choose rehost, VMware-on-AWS or containers per workload, replicate data with change data capture, then shift traffic gradually while the VMware side stays warm as your rollback.
  • The licensing clock sets the pace. Broadcom ended perpetual VMware licenses and SnS renewals in favor of VMware Cloud Foundation and vSphere Foundation subscriptions, and 86% of North American enterprises surveyed by CloudBolt are now reducing their VMware footprint. Your next renewal date is the deadline for the rehost waves.
  • There are two VMware bridges on AWS in 2026. VMware Cloud on AWS is still available but sold only by Broadcom since May 2024. Amazon Elastic VMware Service (EVS) went GA on August 5, 2025 and supports VCF 9.0 and 9.1 since July 2026, with your own VCF subscription. Both move VMs out of your data center while keeping them on VMware licensing.
  • Containerize the slice that ships often, onto Amazon EKS. Stateless services, APIs and jobs go to Kubernetes first. Databases go to managed services with a full load plus CDC. Vendor appliances stay on VMs.
  • Qovery, the Agentic DevOps platform, gives you one deployment workflow on both sides of the migration. It deploys on EKS clusters in your own AWS account and connects on-premise Kubernetes with qovery cluster install, so moving a service from the data center to AWS becomes a change of target cluster, and you can start on the Free plan.

Qovery · Infrastructure control plane
Kubernetes, operated through one governed API
Learn more

Most of the VMware estates I work with are already hybrid. Somebody moved the website and a few APIs to AWS three years ago, the ERP and the databases stayed on vSphere, and a VPN tunnel between the two has quietly become the most important piece of infrastructure in the company. Then the Broadcom renewal quote lands, and "modernize to cloud-native" stops being a slide and becomes a date.

I am a Staff Solutions Architect at Qovery, and I spend a lot of my time on calls with teams in exactly this position. The ones who get hurt rarely fail on the technology. They fail on sequencing: they move a database before its consumers, they order the Direct Connect circuit too late, or they cut over without a way back. This is the plan I walk them through, written for a hybrid AWS and VMware starting point.

If your question is the narrower "should this VM go to EC2 or to containers", start with our rehost versus replatform decision framework. If you want a provider-neutral version of the wave plan, read the phased hybrid cloud migration plan. This article is the AWS plus VMware version, with the bridges and tools that only exist for that combination.

How do you modernize a hybrid AWS and VMware setup without disrupting production?

You modernize a hybrid AWS and VMware setup in seven phases: assess and group workloads, harden the network link between the data center and AWS, decide the target per workload group, containerize the services that ship often onto Amazon EKS, migrate data with replication, cut over gradually with a tested rollback, and decommission VMware once each group has been stable for a full business cycle. Production keeps running on VMware until the AWS side has proven itself.

Every phase needs a written exit criterion and one person who can call a rollback. Here is the plan on one page.

PhaseWhat you doExit criterionRollback
1. Assess and groupInventory VMs, map dependencies, sort into move groupsEvery VM has a group, an owner and a targetNothing moved yet
2. ConnectivitySite-to-Site VPN now, Direct Connect ordered, DNS and identity sharedBoth sides resolve each other's names, latency measuredVPN stays as backup path
3. Target per groupRetire, rehost to EC2, bridge on VMware Cloud on AWS or Amazon EVS, or containerizeSigned-off target list per waveDecision only
4. Containerize to EKSBuild images, deploy the pilot services next to the VMware versionsPilot serves real traffic for two weeksTraffic back to VMware
5. Data migrationFull load plus CDC into the target databaseReplication lag stable, row counts and checksums matchSource stays primary
6. CutoverShift traffic 1%, 10%, 50%, 100% per groupError rate and latency within agreed limitsWeights back to VMware
7. DecommissionPower off, keep snapshots, then deleteOne full business cycle with no rollbackRestore from snapshot

Hybrid is the normal operating state during this, often for a year or more. Plan for it with an end date.

Why is now the moment to modernize, and what changed in VMware licensing?

The trigger for most teams is commercial. In December 2023, Broadcom announced the end of sale of perpetual VMware licenses and of Support and Subscription renewals for perpetual offerings, and consolidated the portfolio into two subscriptions, VMware Cloud Foundation (VCF) and VMware vSphere Foundation (VVF). Existing perpetual licenses keep working, but the support contracts around them do not renew.

The cost effect depends on your contract, so quote it only from named sources. CISPE, the European cloud infrastructure trade association, says in its March 2026 competition complaint that Broadcom's licensing actions have increased some customers' costs by more than 1,000 percent. Broadcom disputes that framing. Your own renewal quote is the only number that matters for your business case.

Teams are moving, at a measured pace. CloudBolt's February 2026 survey of 302 IT decision-makers found 86% actively reducing their VMware footprint and 72% of migrating workloads heading to public cloud IaaS. Gartner expects VMware to lose 35% of its current workloads by 2028, as reported by The Register.

To be fair, staying is a legitimate answer for some workloads. Heavy stateful systems with predictable load, strict residency rules or certified vendor stacks can be cheaper on a renewed VMware contract, on Nutanix, or on Proxmox than on a hyperscaler bill. The plan below assumes you will keep some VMs for a while, and that is fine.

How do you assess and group VMware workloads before moving anything?

Group VMware workloads by dependency, not by application name or business unit. A move group is a set of VMs that talk to each other more than they talk to anything else, so they can move together without stretching a chatty connection across the WAN. Then give each group one of four labels: retire, rehost, bridge or containerize.

Start from data, not interviews. vCenter gives you the VM list, sizes and power state. Network flow data (NetFlow, NSX flow monitoring, or an agent-based discovery tool) gives you who talks to whom. The AWS side matters too: list which services already in AWS still call back into the data center, because those calls are your hidden dependencies.

AWS now automates part of this. AWS Transform for VMware, generally available since May 15, 2025, is an agentic service that runs discovery, dependency mapping, wave planning and network translation for migrations to EC2. AWS says it generated migration wave plans for 500 VMs in 15 minutes and performed network translations up to 80x faster than traditional methods. I treat its output as a strong first draft that a human who knows the estate still has to review.

The four labels I use:

  • Retire. Powered-off VMs, duplicate test environments, and servers nobody claims after two weeks of asking. Every estate I have seen has a meaningful share of these, and deleting them is the cheapest win in the project.
  • Rehost. Stateful or vendor-packaged workloads, Windows apps with licensing ties, anything nobody can rebuild. They go to EC2 with block-level replication.
  • Bridge. Workloads that must stay on VMware for a while (certified stacks, appliances, a team that cannot change tools this year). They move to VMware running in AWS, or stay on-premise.
  • Containerize. Stateless services, APIs, workers and scheduled jobs that deploy at least monthly. They go to Kubernetes.

Write the owner, the target and the rollback plan next to every group in a spreadsheet. If a group has no owner, it is not ready to move.

How should you connect on-premise VMware and AWS during the transition?

Use an AWS Site-to-Site VPN immediately and order AWS Direct Connect in parallel, because the VPN comes up the same day and the dedicated circuit depends on a carrier and a colocation cross-connect. The date your Direct Connect goes live is usually the earliest date for your first data-heavy wave.

The numbers that decide which link carries what come from AWS's own documentation:

LinkBandwidthGood forWatch out for
Site-to-Site VPN, standard tunnelUp to 1.25 Gbps per tunnelDay-one connectivity, control traffic, backup pathInternet path, MTU of 1446 bytes, no jumbo frames
Site-to-Site VPN, large bandwidth tunnelUp to 5 Gbps per tunnelBigger replication streams without a circuitStill over the internet, latency varies
Direct Connect, hosted connection50 Mbps to 25 Gbps via a partnerMid-size estates, faster to provision than dedicatedPartner lead time
Direct Connect, dedicated connection1, 10, 100 or 400 Gbps portsBulk replication, chatty app-to-database trafficPhysical cross-connect lead time

Aggregate VPN tunnels with ECMP on a Transit Gateway if you need more than one tunnel can carry, and keep the VPN as the failover path after Direct Connect is live.

Three things break hybrid estates more often than bandwidth:

  • Overlapping IP ranges. If your AWS VPC CIDR overlaps a data center subnet, fix it before the first wave. NAT between them works and then haunts you for years.
  • DNS. Use Route 53 Resolver inbound and outbound endpoints so both sides resolve each other's private names. Cutover in phase 6 is mostly a DNS and load balancer change, so this must work first.
  • Identity. One identity source for humans (your IdP federated to AWS IAM Identity Center) and a clear plan for service credentials. Do not copy service account passwords into a second vault by hand.

Measure the round-trip latency between the data center and your AWS region before you split any application from its database. If the app makes many queries per request, move the app and its database in the same wave.

Should each workload be rehosted, kept on VMware in AWS, or replatformed to containers?

Pick per workload group. Rehost to EC2 when you need to clear a renewal date with minimum change, use VMware Cloud on AWS or Amazon EVS when a workload must stay on VMware tooling while it leaves the data center, and containerize onto EKS when the service is stateless and ships often. Most estates use three of these at once.

The two VMware bridges are easy to confuse, so here is what each one is in late 2026, from AWS and Broadcom pages:

Both bridges keep you on a Broadcom subscription. They get you out of the data center, which can matter a lot if a lease or hardware refresh is the real deadline, but they do not reduce your VMware dependency. I recommend them only for workload groups with a written end date on VMware.

OptionWhat movesVMware license afterEffortBest forNot for
Rehost to EC2 (AWS MGN, AWS Transform)VM as-is, block-level replicationNoneLow per VMStateful and vendor-packaged apps under a deadlineTeams expecting faster delivery afterwards
VMware Cloud on AWSVMs via HCX, same toolingBroadcom subscriptionLowLeaving the data center while keeping vCenter operationsCutting VMware cost
Amazon EVSVMs via HCX into VCF in your VPCYour own VCF subscriptionMedium (you run VCF)Teams with VCF skills who want VMware next to native AWS servicesTeams without VCF operators
VMs on Kubernetes (KubeVirt with Forklift)VMs into Kubernetes as VMsNoneMedium to highConsolidating a few VMs onto a cluster you already runLarge estates without Kubernetes skills
Containers on Amazon EKSRebuilt as container imagesNoneMedium per serviceStateless services, APIs, workers that ship oftenDatabases and appliances
Containers on EKS with QoverySame images, deployed through one workflow on AWS and on-premise clustersNoneMedium per service, less platform workTeams without a platform group that need EKS and on-prem Kubernetes during the transitionPure VM rehosting with no plan to containerize

The rehost path deserves its due. AWS Application Migration Service performs continuous block-level replication from VMware vSphere and converts servers to native EC2 instances, with the source still running until you cut over. For the long tail of a VMware estate, that is the fastest safe move. It does not fix deployment speed, idle capacity or configuration drift, which is why the next phase exists.

One deployment workflow for both sides of the migration.
Qovery, the Agentic DevOps platform, deploys your apps on Kubernetes in your own AWS account and on the clusters you already run on-premise. Start on the Free plan, no credit card.

How do you containerize VMware workloads onto Amazon EKS without a big-bang?

Containerize one move group at a time and run the container version next to the VMware version until it has served real traffic for at least two weeks. Start with two or three stateless services that deploy often and have a working health check, prove the pipeline end to end, and only then open the floodgates.

The order I use with customers:

  1. Pick the pilot. One internal API, one public stateless service, one scheduled job. Avoid anything that writes to a database shared with ten other apps.
  2. Externalize state. Sessions go to Redis or the database, files go to S3, config goes to environment variables and a secrets manager. This step is most of the work on old VMware apps.
  3. Build the image in CI and push it to Amazon ECR. One image, promoted from staging to production by tag.
  4. Deploy next to the VMware version. Same DNS name behind a weighted record or a load balancer with two target groups, 0% of traffic at first.
  5. Shift traffic and watch. More on that in the cutover section.

You also have to decide where the Kubernetes clusters live during the transition. Three patterns work:

  • EKS in AWS only. Containers run in AWS and reach back to the data center over Direct Connect for anything not yet moved. Simplest, and the default I recommend.
  • EKS Hybrid Nodes. Your on-premise physical or virtual machines join an EKS cluster as nodes, with AWS running the control plane. They need a reliable connection to AWS and are billed per vCPU-hour, starting at $0.020 per vCPU-hour for the first tier. Useful when some containers must stay on-premise for latency or data reasons.
  • A separate on-premise Kubernetes cluster (k3s, RKE2, Tanzu, OpenShift, or upstream) next to EKS. More independence from the WAN, but two clusters to operate.

Whatever you choose, budget for Kubernetes' upgrade rhythm, which is much faster than vSphere's. Each EKS version gets 14 months of standard support at $0.10 per cluster-hour, then 12 months of extended support at $0.60. My colleague Romaric wrote the 8 things to set up after moving VMware workloads to EKS, and I wrote the recurring day 2 calendar that follows it. Read both before you size the team.

How do you migrate the data without a long outage?

Move databases with a full load followed by continuous change data capture (CDC), keep the source as the primary until the target has caught up and been verified, and switch writes in a short, rehearsed freeze. Never plan a database move around a weekend maintenance window you have not timed.

AWS Database Migration Service supports exactly this pattern with a "full load plus CDC" task. It reads changes from the source engine's own logs: LogMiner or binary reader for Oracle, MS-CDC for SQL Server, row-based binlogs for MySQL, logical replication slots for PostgreSQL. Two caveats from the same page belong in your runbook. DMS CDC "does not provide real-time replication" and has no latency SLA, so you measure lag yourself and set a threshold. And the source engines need preparation, for example supplemental logging on Oracle and row-level binary logging on MySQL.

The sequence I recommend:

  • Target first. Provision Amazon RDS or Aurora (or Cloud SQL, Azure Database or Scaleway Managed Database if AWS is not your only target), sized from real usage, not from the VMware VM size.
  • Full load plus CDC. Start the task, wait for full load, then watch CDC lag for at least a week of normal traffic, including your batch peaks.
  • Verify. Row counts and checksums on every table, plus a read-only shadow of the app pointed at the target.
  • Freeze and switch. Stop writes, let lag reach zero, switch the connection string, start writes. Time this in a rehearsal first.
  • Keep a way back. DMS supports bidirectional replication on Oracle, SQL Server, MySQL and PostgreSQL with loopback prevention, but without conflict resolution. Use it only if each table is written on one side at a time. Otherwise your rollback is "restore the source and replay," and you should know how long that takes.

File shares and object data follow the same rule: copy, sync the delta, freeze, final sync, switch. Large one-time copies may justify a physical transfer device, which my colleagues compared in Snowball versus network transfer for EKS migrations.

How do you cut over to AWS and roll back if something goes wrong?

Cut over by shifting traffic in steps (1%, 10%, 50%, 100%) with Route 53 weighted records or load balancer target group weights, and hold each step until error rate and latency stay inside limits you wrote down before you started. Rollback is setting the weights back to VMware, which only works if the VMware side is still running and still receiving data.

Write the rollback triggers before the first percent moves. Mine usually look like this:

  • Error rate on the new path above the old path's baseline for more than ten minutes.
  • p95 latency above the agreed budget, measured from the client side, not inside the cluster.
  • Replication lag above the threshold you set in the data phase.
  • Any data correctness alarm, which means an immediate rollback and no debate.

Each trigger has a named person who can pull it without asking anyone. Rehearse the rollback once in staging and once in production at 1% traffic, and record how long it takes.

Route 53 weighted routing is the simplest tool when both sides have their own endpoints. Lower the TTL a day before the cutover so clients pick up weight changes quickly. For HTTP traffic that already enters through an Application Load Balancer, weighted target groups give you faster, more precise shifts than DNS.

How long should you keep both sides running, and how do you avoid running two platforms?

Keep the VMware side of each move group running for one full business cycle after it reaches 100% on AWS, usually a month-end close or a quarter, then power it off, keep the snapshots for an agreed period, and delete. The whole hybrid phase needs a written end date tied to your Broadcom renewal.

The expensive failure in a hybrid modernization is operational duplication: two deployment paths, two ways to define environments, two sets of credentials, two on-call rotations. Teams accept it as temporary, and in the estates I see it often outlives the migration project.

Three rules keep it under control:

  • One deployment workflow for every container, whether it runs on EKS or on an on-premise cluster. Moving a service then means changing its target cluster, not rewriting its pipeline.
  • One environment definition. Staging and production should be described once and cloned per target, so the on-prem and AWS copies cannot drift.
  • Budget the double run explicitly. You pay Broadcom and AWS at the same time for every month a group sits in both places. Put that line in the business case on day one.

The hybrid cloud migration providers comparison covers who can run the migration itself if your team cannot take it on.

How does Qovery handle a hybrid AWS and VMware modernization?

Qovery, the Agentic DevOps platform, gives you one deployment workflow for containers on both sides of the migration: Amazon EKS in your own AWS account, and the Kubernetes clusters you run next to VMware in the data center.

On the AWS side, Qovery can create and manage an EKS cluster in your AWS account (paid plans), or connect an EKS cluster you already own through the Qovery Operator. You install the Operator once with Helm, it opens an outbound connection to Qovery, runs deployments as Kubernetes Jobs inside your cluster, and builds images in the cluster before pushing them to Amazon ECR in your account. Qovery never stores your kubeconfig. Details are in the BYOK documentation. On the data center side, an on-premise cluster (k3s, RKE2 or any Kubernetes 1.31+) connects with qovery cluster install.

From there, every service gets git-push deployments, environment cloning, a preview environment per pull request, RBAC and audit logs, the same way on both clusters. Moving a service from the data center to AWS becomes a clone of its environment onto the EKS cluster, followed by the traffic shift above. Qovery's control plane is separate from your clusters and never hosts your apps, so the AWS bill and your Savings Plans stay in your name.

For the move itself, Qovery's agentic migration works from your repository: an agent reads it and deploys it into the AWS account you already own, after you approve the plan. It does not convert VMs or read vCenter, so the rehost and bridge groups still go through AWS MGN, AWS Transform or HCX.

Start on the Free plan (3 users, unlimited AI agents, one cluster of your own, no credit card), or try it locally with qovery demo up.

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

Install the Qovery Agent Skills once, optionally add the Qovery MCP Server, then ask your coding agent in plain language to connect a cluster and deploy a service. The agent asks you to approve each action.

  1. Install the Agent Skills for every compatible agent (Claude Code, Cursor, Codex, OpenCode, Gemini CLI), globally or for one project:
BASH
curl -fsSL https://skill.qovery.com/install.sh | bash
curl -fsSL https://skill.qovery.com/install.sh | bash -s -- --project
  1. Add the Qovery MCP Server to Claude Code so the agent can read clusters, environments, deployment status and logs. It uses OAuth and is read-only by default:
BASH
claude mcp add --transport http qovery https://mcp.qovery.com/mcp --callback-port 4242
  1. Only if you want the agent to change things, add write access. This also requires enabling write access in the Qovery Console:
BASH
claude mcp add --transport http qovery "https://mcp.qovery.com/mcp?read_write=true" --callback-port 4242

The skills (qovery-onboard, qovery-deploy, qovery-troubleshoot, qovery-optimize, qovery-speedup, qovery-preview, qovery-terraform) deploy; the MCP Server operates what already exists. Then ask something like: "Onboard my EKS cluster to Qovery, deploy the API from this repo to a staging environment, and clone it from the on-premise cluster's environment." Start with the AI agent quickstart or the Claude Code integration. More Kubernetes guides live on our Kubernetes topic page.

Frequently asked questions
How do I modernize a hybrid AWS and VMware setup to cloud-native infrastructure without disrupting production?

Move in waves of dependency-grouped workloads: connect the data center to AWS with VPN and Direct Connect, choose rehost, a VMware bridge or containers on EKS per group, replicate data with full load plus CDC, and shift traffic in steps while the VMware side stays running as your rollback. Decommission each group only after a full business cycle without a rollback. Qovery helps with the container part by giving EKS and your on-premise Kubernetes one deployment workflow.

Is VMware Cloud on AWS still available in 2026?

Yes. Broadcom stated in May 2024 that VMware Cloud on AWS continues to serve customers, but it is no longer sold by AWS or AWS partners, only by Broadcom and its authorized resellers. Broadcom still publishes service listings for it in 2026.

What is the difference between VMware Cloud on AWS and Amazon Elastic VMware Service?

VMware Cloud on AWS is a managed service bought from Broadcom, where the VMware stack is operated for you. Amazon EVS runs VMware Cloud Foundation on bare-metal EC2 inside your own VPC, with your own VCF subscription, and you patch and upgrade VCF yourself. Both keep you on a Broadcom subscription.

Should I move VMware workloads to EC2 or to containers on EKS?

Rehost to EC2 for stateful, vendor-packaged or Windows workloads, and for anything you must move before a renewal date. Containerize stateless services that deploy at least monthly onto EKS. Most hybrid estates do both in parallel.

Can Qovery manage both my EKS clusters and my on-premise Kubernetes during the migration?

Yes. Qovery connects EKS clusters you own through the Qovery Operator and on-premise clusters through qovery cluster install, then deploys services the same way on both. Qovery does not migrate VMs, so rehosting still goes through AWS MGN, AWS Transform or HCX.

How long does a hybrid VMware to AWS modernization take?

Rehost waves take days to a few weeks per group, containerizing takes weeks to a few months per group, and the whole hybrid phase commonly runs a year or more for a mid-sized estate. Tie the end date to your next Broadcom renewal and budget the months where you pay for both sides.

Guillaume Da Silva
About the author
Guillaume Da Silva

Guillaume is a Staff Solutions Architect at Qovery. He helps customers troubleshoot production infrastructure and writes about Kubernetes operations, AI agents, and reliable cloud platforms.

Next step

One deployment workflow for both sides of the migration.

Qovery, the Agentic DevOps platform, deploys your apps on Kubernetes in your own AWS account and on the clusters you already run on-premise. Start on the Free plan, no credit card.