← All articles

What it would cost to move 1,300 VMs to AWS: numbers from a real estimate

This year we built a cost model for a company that asked a simple question: what would our on-prem estate cost if we moved it to AWS? Management expected the answer "cheaper over ten years". The numbers said otherwise.

The company and the figures are anonymised: amounts are rounded and some details are changed. The method and the proportions are real.

How the estimate was built: the on-prem estate on the left is sized VM by VM into an AWS bill on the right; the migration effort is shown below.ON-PREM TODAY1,300 VMs12,000 vCPU · 27 TB RAM950 TB of disk570 TB on six file servers~100 SQL · ~150 Windows VMslicenses bought years ago3 sites, owned hardwarecolocation, links, supportTeam of 63 network · 3 systems$2.6M / year+ $0.4M teamPER-VM SIZINGRAM : vCPU→ instance familyAWS BILL, PER MONTHComputeVMs, Kubernetes, GPUBlock storagevolumes and snapshots$0.31-0.72MFile storage570 TB moved off block disksMicrosoft licensing35-41% of the bill$0.20-0.58MTransfer, supportegress measured, support tiers$0.06-0.10M$0.57-1.40M / monthGetting there: 86-154 person-months, 5-9 people for about 18 months570 TB of files is about 53 days at 1 Gb/s. SQL and queues need a coordinated cutover.
How the estimate was built. Numbers are rounded; ranges run from the optimistic to the worst scenario. AWS icons © Amazon Web Services.

The estate

  • Three sites: two main data centers and a smaller third one.
  • About 1,300 virtual machines in scope, 12,000 vCPU, 27 TB of RAM and 950 TB of disk.
  • VMware for most workloads, a Kubernetes cluster, a few GPU servers and several generations of storage arrays.
  • Around 100 SQL Server VMs and 150 Windows Server VMs.
  • A team of six engineers: three network, three systems.

One scope decision mattered early. The second data center mostly holds standby copies of services from the first one. In the cloud, resilience is built with availability zones, not with a second copy of everything, so those copies were left out. Only the services that exist on the second site alone were counted.

How the cloud side was priced

Every VM was mapped to an instance by its own size, not by host. The instance family came from the RAM-to-vCPU ratio (compute-, general- or memory-optimised), and the size was rounded up to the next AWS step. Then we added what quick comparisons usually forget:

  • block storage, snapshots and a separate design for very large file volumes;
  • Microsoft licensing for SQL Server and Windows;
  • data transfer out, measured from real traffic over two months;
  • the support plan, which scales with the bill.

We gave three scenarios. The optimistic one uses committed pricing and reuses existing licenses where the rules allow it. The middle one is a realistic mix. The worst case is on-demand pricing with license-included instances.

Per month Optimistic Middle Worst
Compute, storage, network $0.31M $0.52M $0.72M
Microsoft licensing $0.20M $0.32M $0.58M
Transfer, support, other $0.06M $0.08M $0.10M
Total per month $0.57M $0.92M $1.40M
Total per year $6.8M $11.0M $16.8M

How the on-prem side was priced

We took the approved annual budget for the infrastructure and summed only the leaf lines, never the group totals, to avoid double counting: colocation and remote hands, internet and links between the sites, new hardware for the year, vendor support and licenses. Services that stay the same in both worlds (CDN, SaaS, existing cloud accounts) were excluded from both sides.

Result: about $2.6M a year for the sites in scope, plus about $0.4M a year for the team.

The comparison

Optimistic Middle Worst
Cloud vs on-prem, infrastructure only x2.6 x4.2 x6.4
Cloud + target team vs on-prem + current team x2.3 x3.7 x5.6
Ten years, including two years of running both x2.4 x3.7 x5.5
Yearly infrastructure cost to scale: on-prem 2.6 million dollars; AWS 6.8 million optimistic, 11.0 million middle, 16.8 million worst.YEARLY INFRASTRUCTURE COST, TO SCALE$0M$3M$6M$9M$12M$15M$18MOn-prem today$2.6MAWS optimistic$6.8Mx2.6AWS middle$11.0Mx4.2AWS worst$16.8Mx6.4Infrastructure only. With teams on both sides: x2.3 / x3.7 / x5.6. Over ten years: x2.4 / x3.7 / x5.5.
The cloud bill against today's on-prem budget. Warmer colour means a more expensive scenario.

Even with no Microsoft licenses at all, the cloud stayed about 1.7 times more expensive than on-prem.

What drove the result

Microsoft licensing was 35-41% of the cloud bill. On-prem, those licenses were bought years ago and barely show up in the budget. In the cloud, Windows can only use your own licenses on dedicated hosts, and SQL Server needs active Software Assurance to move. Without proof of Software Assurance, the difference between scenarios was about $360k a month. This was the first thing we asked the company to check.

Owned hardware is cheap when it runs for years. Most servers were well past their purchase cost and still under affordable support. The cloud charges for that capacity every month.

Fewer engineers does not close the gap. The cloud needs different skills (FinOps, identity and security, cloud networking), and a 24/7 rotation still needs at least four people. Even cutting the team from six to two would cover only about 6% of the yearly difference in the optimistic scenario.

The cost of getting there

The bill is only half of the decision. The migration itself was estimated at 86-154 person-months, which means 5-9 people working on it full time for about 18 months:

  • 570 TB sits on six file server VMs. At 1 Gb/s that is about 53 days of continuous transfer, before any re-sync.
  • Older storage arrays cannot run replication agents, so those workloads need a different path.
  • Around 200 message-queue VMs and 100 SQL Server VMs need a coordinated cutover, not a VM-by-VM move.
  • A single cloud instance has a 99.5% SLA. Getting to 99.99% means running across two availability zones, and each 10% of load in a second zone adds $16-44k a month.
  • The team's skills are built on VMware. Training and hiring take time while the same people keep running production.

Not "move" and not "stay". Our recommendation was to:

  1. Confirm the Microsoft licensing position first, because it moves the answer more than anything else.
  2. Run a pilot with one isolated, spiky service where the cloud should win, and measure real costs for three months.
  3. Plan the next hardware refresh and the VMware renewal as the real decision points, with this model as the baseline.

The pattern we see again and again: stable, heavy workloads on hardware you already own are hard to beat on price. The cloud wins on elasticity, speed and managed services, and that is where it should be used.

If you are facing the same question, write to us. We will tell you which data we need and how long the model takes to build.

Have an infrastructure decision coming up?

Tell us what you are planning and we will come back with how we would approach it.

Email us[email protected]