Fair bit of our computers are BrandA - VersionB and run SoftwareX1 - VersionY1, SoftwareX2 - VersionY2 and we run it on Windows 7 - upgradeZ
We do that because the client requires it.
That specific software combination has been verified and validated on that specific family of chipsets and while I'm sure running these programs on win10 won't burn down the house down or even change the outcome in a statistically relevant way, it may just mean that none of our models/results can be integrated with the work of any of the other teams. Or worse, it would integrate with a very slight anomaly and influence the surrounding systems.
That would be a bad thing.
Apparently running SoftwareX1 - VersionY1, SoftwareX2 - VersionY2 on the old Windows XP - upgradeZ but on a certain other family of CPUs causes divide by zero errors in specific modules of SoftwareX1 due to the way . . . . certain floating doubles are handled?, although I've never tested this.