Running a Real Data Center on Hyper-V, No VMware Required
A hoster built a GPU-accelerated CAD desktop business on 100% Microsoft infrastructure, no VMware. The real rules for migrating an entire data center off it.
Bronson Magnan runs a hosting business on 100% Microsoft infrastructure. No VMware, no ESXi, no third party hypervisor licensing anywhere in the stack. His company, Coastal IT Consulting, hosts GPU-accelerated CAD desktops for engineering firms, with entire physical GPUs assigned straight into virtual machines. On this episode of VMware to Microsoft, he walks through how that business is built and what it takes to run a full data center on Hyper-V.
Why Hyper-V, not VMware
Bronson gave two reasons, and only one of them is about the sticker price.
First, his engineering goal is 100% virtualization and 100% automation. A stack that’s entirely Microsoft is entirely scriptable in PowerShell. Mixed hypervisors mean mixed APIs and mixed automation, and that friction adds up.
Second, and more important for the business model, it’s a licensing-economics argument, not a feature comparison: hosting isn’t a cost center for him, it’s the product. Windows Server Datacenter licensing already covers Hyper-V. No separate hypervisor license stacked on top. For a hoster trying to build a service people will pay for, that’s the entire economics of the business.
GPUs, passed straight through
Coastal’s hosted desktops run Windows Remote Desktop Services with Nvidia GPUs for AutoCAD and other GPU-dependent engineering apps. Bronson’s team was one of the first three hosting partners in North America running Nvidia GPUs on Hyper-V.
The feature that makes it work is Direct Device Assignment: an entire physical GPU handed to one VM, not shared, not virtualized. His nodes run two cards with four GPUs each, eight GPUs per node, one GPU per RDS host VM. The GPU isn’t just rendering CAD models. RDS also uses it for hardware-accelerated encoding, which is what makes the remote session feel responsive on the client side.
Being first meant doing the certification work nobody else had done. Bronson worked with the OEM’s labs in Germany, Mexico, and China, and had to sign off on running GPU-equipped chassis at higher temperatures than the hardware was originally rated for. That triggered a chassis redesign on the OEM’s end.
Direct Device Assignment has a real limitation: you can’t live migrate a VM with a GPU attached. Bronson’s workaround is scripted. Shut the VM down, dismount the GPU, quick-migrate the VM to another node, reattach a GPU there, boot it back up. By hand, at his scale, that’s not viable. As a PowerShell runbook, it’s routine maintenance.
Storage: together or apart
Storage runs on Storage Spaces Direct: local disks, no RAID controller, no SAN. Bronson uses it two ways. Hyperconverged, where compute and storage share the same nodes, and disaggregated, where a dedicated storage cluster serves capacity to separate compute nodes over SMB.
The disaggregated setup solves a real hyperconverged problem: uneven fill. If every VM template uses the same ratio of CPU, RAM, and disk, a hyperconverged cluster runs out of one resource before the other, and whatever’s left over sits wasted. Splitting storage onto its own cluster lets compute and storage scale independently.
It’s not free. Storage-only nodes still need Windows Server Datacenter licensing for the Storage Spaces Direct feature, and compute nodes need it too for VM density, so you’re licensing both. Bronson’s offset is core count. Licensing is priced per core, and storage nodes don’t need many, just fast ones. Fewer cores on the storage side, lower licensing bill, and you never bought a SAN to begin with.
The migration is the hard part
Bronson has run full VMware to Hyper-V data center migrations and breaks the process into eight rules. A few are worth calling out directly.
Isolate the fabric completely. Your management network and its Active Directory forest should have no trust relationship with the tenant or business network. Hyper-V hosts run Windows, unlike ESXi, which means they can be crypto-locked. A compromised workload should never be able to reach the infrastructure running it.
Don’t assume your hardware carries over. Storage Spaces Direct has specific NIC and disk backplane requirements, and software defined networking has its own NIC and switch feature requirements. Some of what you own today survives the move. Some doesn’t, and a consultant during the architecture phase will tell you which is which before you buy anything.
Automate everything, on principle, not just for convenience. Bronson uses the free Microsoft Deployment Toolkit with network boot to image hosts automatically, down to static IP assignment, with fresh images generated monthly.
Budget real time for VM conversion. At real scale that means multiple licenses of a third-party conversion tool running jobs in parallel, not someone clicking through conversions one at a time.
His closing advice for any enterprise weighing the move: don’t rule it out purely on hardware cost. If your current servers are more than a few years old, newer hardware plus zero VMware licensing can land cheaper than keeping the old fleet and renewing VMware support.
This is part of my VMware to Microsoft podcast series.
Subscribe for more of these.