716 workloads, one region, $2.77M a year between the best and worst migration models
The challenge
This was the largest estate in our assessment portfolio, and two constraints shaped the entire engagement.
Region was fixed. Regulatory and latency requirements tied the target to Africa (Cape Town). That removed any possibility of arbitraging cheaper regions and made right-sizing against measured utilisation the primary lever available on compute cost.
The estate was Windows-dense. With 557 Windows servers in scope, Microsoft licensing dominated. Windows Server licensing alone ran to $1,214,521 a year under a standard shared-tenancy model, 28% of the total bill and second only to compute.
There was a third complication. The estate was hybrid, spanning conventional virtual machines and a substantial Azure platform-services footprint. Counting each hosted object would have inflated the model badly: 238 App Services and 41 Function Apps do not each need their own instance, and 149 managed databases do not each need their own database server. Any credible cost model had to map platform constructs to comparable AWS targets on the basis of real consumption.
Our approach
We built a defensible mapping for the Azure footprint first. App Service plans were sized as EC2 instances rather than counting each hosted application, since the App Services, deployment slots and Function Apps all ride on the underlying plans. SQL Managed Instances were mapped to their host rather than to each of the 149 databases they contained. Azure VMs and scale-set members were mapped one to one, with scale-set definition rows excluded so members were not double-counted.
We right-sized every workload against measured utilisation rather than provisioned specification.
We modelled a mixed-tenancy design using the bank's own Microsoft entitlement: 208 instances packed onto 12 EC2 Dedicated Hosts under Windows Server and SQL Server BYOL, drawing on 596 Windows Datacenter cores purchased before October 2019, with the remaining 515 instances on shared tenancy and 60 SQL Server Standard cores carried under License Mobility.
We also priced the seven virtual desktops onto Amazon WorkSpaces and modelled AWS Elastic Disaster Recovery across all 716 in-scope servers.
The results
Right-sizing alone removed $502,394 a year against the committed direct-match figure, before any licensing decision was taken.
Applying the bank's own Windows entitlement cut the Windows Server licensing line from $1,214,521 to $772,877, a further $441,644 a year. SQL Server licensing fell to $0.
The recommended position sits 54% below the on-demand lift-and-shift baseline and 29% below the committed one. Disaster recovery across the full estate priced at $630,011 a year, and virtual desktop replacement at $6,216.
Why it matters
At this scale, percentage points carry six-figure consequences. The gap between the worst modelled position and the recommended one is $2.77M a year, which is the entire business case for doing the work properly.
The second lesson is about hybrid estates. Organisations that already run cloud platform services alongside on-premises infrastructure need both mapped in the same model, using the same assumptions, or the migration is priced against an estate that does not exist. Publishing the mapping table alongside the numbers let the bank's own architects challenge each assumption line by line rather than accept a headline figure on trust.
Collection for this engagement ran through Cloudamize. What that tool can and cannot see: read the tool page →
Talk to us about large-estate migration
An AWS Optimization and Licensing Assessment handles estates of any size, including hybrid footprints spanning on-premises virtualisation and cloud platform services, and prices every commercial model side by side.