I can see a few reasons:

• automated tests on single frames
• batch renders on a server (e.g. for stills or cutscenes)
• comparisons across GPU archs - it could essentially be the "standard" for how a scene should be rendered

And of course, maybe some CPU manufacturer will build in an accelerator so lower end GPUs (say, APUs) could have reasonable raytracing in otherwise GPU limited games (i don't know enough about modern game pipelines to know if that's a possibility).

Or the final reason, which may be the most important of all: why not?

If the CPUs get strong enough, they could run old raytracing games at some point … especially on hardware platforms that don’t have ray tracing GPUs available for them.