Application delivery insights since 2011.
Cloud desktops, IaaS silicon, egress and hybrid multicloud economics.
Microsoft’s users-per-vCPU table is a starting point, and Microsoft says so. Doubling vCPUs returns 1.5 to 1.9 times the users, logon bursts need their own headroom, and whether autoscale saves money depends on how you set load balancing per phase.
The exit strategy works on paper until somebody runs the egress estimate. EU Data Act switching charges end on 12 January 2027, but the DORA exit test, the parallel run and the partial rebalance are all still metered. What a 5,000-seat EUC estate actually costs to move, and the three mitigations that work.
You are paying full price for hardware amortized years ago, and the VM you get is drawn from a lottery of CPU generations spanning a decade. Both share a root cause: you are consuming a managed VM layer you do not control. The case for bare metal with your own hypervisor, and how it turns the exit plan DORA requires into something you can test.
After the Intel-side breakdown, the obvious follow-up was: does the same lottery exist on the AMD SKUs? Short answer — yes, but it plays out very differently. A pragmatic walk through Azure Das/Eas, AWS T3a/M5a/M6a/M7a/M8a, and GCP N2D/T2D/C3D, with mitigation that actually works.
When you request a Standard_D4s_v3, Azure assigns you a physical host from whatever is available in the cluster. Microsoft publishes the per-host VM densities but not the fleet mix, so the odds of landing on 2019-or-older silicon are a model, not a measurement. Here is that model, and what it means for AVD and Windows 365.
Hyperscalers fill racks with CPUs at launch, depreciate them over five to seven years, then charge you the same price while pocketing an ever-growing margin. The EUC industry is along for the ride — and the bill is about to increase.
