Right-sizing engineering and GPU workstations: when the next one should be smaller

Why like-for-like refresh is the expensive default

A workstation can stay on the asset register after its user's workload changes, and procurement then replaces it with the same class at refresh. The cost of that habit is the gap between the workstation and the device the user now needs, multiplied by the number of desks where it has happened, and at 2026 prices the gap has widened. Gartner noted in October 2026 that businesses focused spending on targeted investments such as workstations and AI PCs rather than broad deployments, which makes each workstation decision count for more.

There is no reliable public statistic for how many engineering workstations run office work; vendor case studies exist and we do not quote them. The honest method is to measure your own fleet, and the measurement is cheap.

The evidence that says a workstation is over-specified

Three sources, all already in the tenant. Installed and detected applications, from the Intune device inventory or the Configuration Manager installed programs view, show whether the engineering application is even present. Software metering in Configuration Manager, or recently used application data where you collect it, shows whether it has run in the last 180 days. And the resource performance report in Intune Advanced Analytics shows whether the processor and memory have been stressed: Microsoft defines CPU usage above 50 percent and RAM usage above 75 percent as spikes, averages the percentage of usage time spent in spikes over 14 days, and lists processor name, cores, base speed, RAM and disk type per device.

The video controller inventory completes the picture. The Intune properties catalog collects it under its own category, and Configuration Manager exposes it in the video controller view, one row per adapter. A device whose discrete or professional GPU shows no application that would use it is the clearest right-size signal in the fleet. In our fictional example, WS-4471 runs office and browser applications only, with no engineering application seen in 180 days. It is 4.3 years old against a 3.5-year class life and is also due under the four-year calendar. Its Right-size verdict replaces it with a Knowledge Worker device; no redeploy match is assumed.

Classes, lives and the price difference

Right-sizing only works when the classes have prices. The defaults we use are on the verdicts page: Knowledge Worker $1,250 with a four-year life, Developer $2,400 at three years, Engineer $3,200 at three and a half, 3D Engineer $5,500 at three. Your procurement prices replace them. The saving on a right-size verdict is the difference between the class bought and the class needed, counted once at the refresh: an Engineer desk that gets a Knowledge Worker device records $1,950 at default prices, and a 3D Engineer desk that steps down to Engineer records $2,300.

Count it on its own line, separate from deferred purchases and from the redeploy pool. The device being replaced may also redeploy to a due seat, and that is upside counted separately, as the redeploy guide explains. Keeping the three lines apart is what lets the headline equal their sum on the page.

The cases where the bigger class is right

Not every idle GPU is a mistake. A structural engineer who runs a finite element solver twice a quarter needs the workstation on those two days, and the 14-day resource average will not show it. Simulation and rendering workloads are bursty by nature. Ask the application owner for the list of users licensed for the heavy applications, and treat a licensed user with an idle GPU as a question for the manager rather than a verdict. The report flags it; a person decides.

The other direction matters too. A developer on a Knowledge Worker laptop whose resource performance shows RAM spikes every day is under-sized, and a refresh that gives that desk a Developer class device removes tickets as well as waiting time. Right-sizing is a fit test in both directions, and the class table is what you fit against.

AI PCs and the next class question

Copilot+ PCs add a new tier to the class table. Microsoft's specification requires a neural processing unit capable of 40 or more trillion operations per second, 16 GB of memory and a 256 GB solid-state drive, and it notes that some PCs with a qualifying NPU still do not count because of smaller memory or storage. Gartner wrote in February 2026 that rising AI PC prices will delay the projected 50 percent market penetration of AI PCs until 2028. For a refresh decision that means the AI PC is a class to assign on evidence of need, like the Engineer class, rather than a default for every Knowledge Worker seat.

The sample report shows how the right-size cohort appears beside the other three verdicts on a fictional fleet, with 2,330 due devices stepped down a class and the price difference totaled. For your own fleet, the initial report lists the right-size count by class from an inventory export and the Endpoint analytics performance report, and the readout is where the application owners get asked about the bursty cases.

A right-sizing pass in one afternoon

If you want to run the pass yourself before asking anyone for a report, the sequence is short. Pull the installed applications by device from Intune or from the Configuration Manager installed programs view, and pull the video controller inventory from the properties catalog or the video controller view. Join the two on the device name and keep every device with a discrete or professional adapter. Then open the resource performance report by model, or the metering data if you are on Configuration Manager, and mark the workstations with no engineering application run in 180 days and no CPU or RAM spike in 14 days.

Send that list to the application owners with one question per device: does this person need this class of machine for work the data cannot see? Agree a response deadline with the application owners and record their answers against each device. Confirmed office workers go to the right-size cohort at the next refresh and their workstation goes to the redeploy pool now. Confirmed bursty users stay in class, with a note on the record so the next pass does not ask again. The undecided ones get the question put to their manager, with the class prices in front of them, which is usually where the decision gets made quickly.

Common questions