Most places I've worked - organizations far smaller than the IRS - have had site licenses, so the costs of upgrading don't lie in paying for the OS.
The costs lie in training, upgrading hardware, and upgrading user applications (including things that are closed source, discontinued or developed in house by people who have left the organization).
Unfortunately you'd get most of these problems in Linux too.
If you read the article, they have already upgraded some of their computers, so the upgrading user applications has already been done, as the machines that have been upgraded would have needed this. The training required to go from one version of an OS to another version of the same OS, is very miniscule, as (at least for the standard user usage, which the majority of IRS users, they fall under) OS developers go out of their way to make the changes as seemless as possible. And sure upgrading hardware may be a cost, but again this generally a much smaller problem with Linux, as Linux is generally a much lighter weight OS than Windows is.
> If you read the article, they have already upgraded some of their computers, so the upgrading user applications has already been done, as the machines that have been upgraded would have needed this.
But they've upgraded to another Windows version, not a * nix. Windows has taken backwards compatibility support to plaid (11 if you're a fan of a different movie). I worked in an office with a data analysis tool written using Visual Fortran in the 1990s for Windows 95. Without retargeting it it runs just fine with a single launch flag set for Windows to use the appropriate compatibility mode. The cost of upgrading user applications is really the cost of testing user applications. Moving to * nix wouldn't be upgrading, it'd be rewriting/porting or significant testing with Wine or some other compatibility layer. [1]
[1] As others pointed out, as more applications run on the CLR porting to * nix will become easier as CLR support improves on those platforms. But that also requires a rewrite, you can't take an existing Fortran or VB or C++ project and retarget it without effort.
The costs lie in training, upgrading hardware, and upgrading user applications (including things that are closed source, discontinued or developed in house by people who have left the organization).
Unfortunately you'd get most of these problems in Linux too.