Five Things About VDI That Only Become Obvious After Year Two

We have deployed and operated virtual desktop infrastructure at real scale for a long time. Between our customer base and our own delivery platform, we've stood up over a million VDI sessions across Citrix, VMware Horizon, AVD, and more recently Windows 365 and various private-cloud-hosted flavors. What follows is not the stuff you learn in the vendor training class. It's what we've learned from operating the thing in production for year after year, through user migrations, pandemic spikes, hardware refreshes, and the slow grind of reality meeting slide decks.

If you're evaluating VDI, or you're a year into a deployment and wondering why things feel harder than you expected, this is the post I wish somebody had handed me in 2015.

1. The Hardest Part of VDI Is Not the Desktops

Everyone sizing a VDI deployment starts with user counts and concurrency and session density per host. Fine. Those numbers matter. But they are not the hard part.

The hard part is the profile, the user data, the application delivery, and the printing. That is where 80 percent of your operational pain lives for the lifetime of the deployment.

Profiles are an ongoing negotiation with Windows

FSLogix, UPM, roaming profiles — whatever you use, the profile container is the thing that will corrupt, bloat, desync, and get stuck. Plan for it. Build tooling to detect runaway profiles (over 20 GB is usually a sign something is broken). Build a workflow for the help desk to reset profiles without destroying user data. Back them up separately from the base image. Expect profile problems to be 30 percent of your VDI-related tickets, forever.

Application delivery is its own engineering problem

The dream is a golden image with all applications installed. The reality is that most organizations have 200 to 600 distinct applications, many of which conflict, many of which need per-user licensing, and many of which the vendor has never heard of VDI. App attach, MSIX app attach, FlexApp, App Volumes, ThinApp — pick a strategy and commit to it, because the "just install everything in the golden image" approach stops working somewhere around application number 40.

Printing will always be broken

Universal print drivers, session printers, print redirection, cloud print — every solution is partially broken in a different way. Budget for print engineering as a continuous line item. Organizations that try to solve printing once and move on are the ones with the worst support experiences. Accept that this is a standing concern and resource it.

2. Storage Performance Is the Secret Determinant of User Experience

The number one user complaint about VDI is "it's slow." The number one cause of "it's slow" is storage latency, not CPU and not network.

VDI is a pathological workload for storage. Boot storms, login storms, antivirus scans that kick off across a thousand sessions simultaneously, profile reads and writes, temporary file churn — all of it hits the storage layer hard, and almost all of it is small random I/O, which is the exact thing most storage systems are worst at. If your average session experience feels sluggish, the first place to look is IOPS per session, storage latency at the 99th percentile, and queue depth during login windows.

Local NVMe is almost always the right answer

For persistent data you use shared storage, obviously. But for the running desktop OS, for profile cache, for Windows paging, for temp — local NVMe on the host is dramatically faster than any SAN, any NAS, and any cloud-managed disk. We deliver our own hosted VDI on local NVMe with a separate profile container tier on shared storage, and user experience metrics are meaningfully better than the same workloads on all-shared-storage designs.

Antivirus exclusions are not optional

If you do nothing else, go read your AV vendor's VDI exclusions list and apply it. Defender, CrowdStrike, SentinelOne, Trend — they all publish recommended exclusion lists for VDI base images and profile paths. Failing to apply them is the single most common cause of mysteriously slow VDI we see.

3. The TCO Math Changes Completely at Year Three

The cost case for VDI on day one is almost always worse than the cost case for endpoints. Hardware, licensing, deployment labor, initial training — a laptop is just cheaper up front. The TCO case comes from year two onward, and it depends on operational maturity you might not have yet.

Year one: more expensive than laptops

Assume this. Plan for it. Do not pitch VDI as a cost-savings project in year one, because you will be wrong and it will damage the program's credibility.

Year two: roughly break-even, depending on your refresh cycle

Year two is where the math starts turning. If you would otherwise be refreshing endpoints on a three-year cycle, avoiding that refresh (by replacing laptops with thin clients or just keeping the old hardware longer) starts moving the numbers. Image management becomes a single operation instead of 500. Onboarding a new user drops from a half-day desk visit to a click.

Year three onward: the real case

Year three is where you see the operational savings compound. Security patching is centralized. Help desk calls drop because rebuilding a broken desktop takes two minutes instead of two hours. Contractor access becomes a matter of provisioning a session instead of shipping a laptop. Regulatory audits become dramatically easier because the data never left the datacenter. This is the real case for VDI, and it takes years to accrue.

If leadership is looking for 12-month ROI, don't do VDI. Pick a different project.

4. Cloud VDI Is Not Cheaper Than Private Cloud VDI For Steady-State Workloads

Windows 365, AVD, WorkSpaces, and Horizon Cloud are all fine products. They solve real problems — burst capacity, rapid provisioning, geographic distribution, minimal infrastructure commitment. They are not cheaper for steady-state workloads, and most VDI is steady-state.

The public cloud VDI premium is real

A permanent desktop in Windows 365 or AVD costs somewhere between 2x and 4x the per-seat cost of the same workload on a well-run private cloud in a colo, once you include storage, backup, networking, and the licensing math. We've run the numbers dozens of times. The answer is almost always the same: if you know your concurrency, if your users are on for 8 hours a day 5 days a week, if you're going to run this deployment for three years or more — private cloud wins by a wide margin.

Cloud VDI wins for specific use cases

It wins for: burst workloads (tax season, back-to-school, disaster recovery), small populations (under 50 users), geographically distributed users without a good regional datacenter, and pilot deployments. For those use cases the cloud premium is worth it, because the alternative is standing up capital infrastructure you may not need.

Most deployments should be hybrid

The honest answer — and the one we end up deploying most often — is a private-cloud base footprint for the 80 percent of users who are steady-state, with a cloud VDI overflow tier for bursts, seasonal workers, and business continuity. The management pane is unified. The user experience is consistent. The TCO beats either pure option.

5. GPU-Accelerated Sessions Are No Longer a Niche

For a long time, GPU-accelerated VDI was something you did for engineering, CAD, medical imaging, and a small number of other niches. That's changing.

Windows 11 assumes a GPU

Windows 11 and the modern Edge/Chrome browsing experience are meaningfully worse without GPU acceleration. Microsoft Teams, in particular, does not degrade gracefully on GPU-less VDI — video calls will eat a CPU for lunch and degrade both the call and every other session on the host. If you're deploying VDI today for general knowledge workers, plan for at least a fractional GPU per session. The NVIDIA vGPU and AMD MxGPU options both exist, and the cost delta is smaller than the user experience delta.

AI tooling in the desktop will push this further

Copilot, local AI assistants, image generation in creative apps, and the next generation of productivity tools are going to assume GPU resources on the endpoint. A VDI deployment designed without GPU headroom in 2022 will feel old in 2026. Budget for it.

What We Actually Deliver

When a customer asks us to scope a VDI deployment, the delivery is never just "desktops on a platform." It's a profile strategy, an application delivery strategy, a print strategy, a storage design that accounts for IOPS per session at the 99th percentile, a GPU sizing decision, an antivirus tuning pass, a disaster recovery plan, and a year-two operational handoff. The desktops are the easy part. Everything around them is the project.

Three Takeaways

1. VDI is an operations project, not a product project. The platform you pick matters less than the operational maturity you build around it. Budget accordingly.
2. Storage IOPS per session at the p99, not CPU, is the user experience metric that matters most. Measure it, alert on it, engineer for it.
3. Private cloud beats public cloud for steady-state VDI, but hybrid beats both. Keep the predictable load on capital-efficient infrastructure, and use cloud VDI for the burst cases where elasticity actually matters.