During my time at PDI/DreamWorks, all idle desktop machines would be incorporated into the render farm as well. I don't know if this is the case anymore with so many features of the renderer and simulations software depending on the GPU.
Pixar’s own Renderman is still primarily a CPU renderer for them even though they have XPU coming, and DreamWorks’ soon to be open sourced MoonRay is also primarily CPU only.
Most feature film scenes are still too large to fit in GPU memory comfortably. While some GPU renderers can handle out of core rendering (Redshift etc), there’s a lot of concessions to be made in terms of shader types (subsurface is/was much better on CPU renderers) and performance issues when shuttling data back and forth.
When we did some evaluation of using Redshift for some projects, our CPU renderer still handedly beat it for mid-large scenes, so it was rarely worth the effort to use it in our cases.
There’s also the issue of retrofitting infrastructure. The studios have hundreds of farm machines without GPUs. Most studios run on leases so would have a delay before GPUs would be next widely available in their farms.
Ironically, one of the best devices for XPU rendering would be the new Apple Silicon macs. Not having to shuttle data back and forth, plus having a relatively massive memory pool for the GPU is a really advantageous setup. In theory, they could actually out pace much higher end hardware for specific scenes if the renderers could take advantage of both CPU/GPU and unified memory.