There is the full power of CUDA kernels, for starters. Then there's a lot of potential low level optimizations that browsers don't enable that can easily make a 2x to 10x performance difference. Also, there is no good way to give hard bounds on memory usage.
I think the use of CUDA is niche in the CAD world, probably constrained to a sub-set of FEA. We dropped 6 figures on an Abaqus system -> new hardware, new software within the last couple of years, and a lot of the analysis still runs on the CPU, with a bit of off load to the GPU.
On the performance end, the is a wide range of CAD that was done on systems more than 10x slower, so that also may not be a deal-breaker. For the simpler end of professional CAD operation, I haven't noticed any performance gain from a 2nd get Core i7 to a 10th gen. On the high-complexity end, there is some benefit for some operations, but certainly there is a wide scope of application that could be fine with only 10% of peak performance on newer work stations.
CUDA kernels are a non starter because they require Nvidia GPUs. Things like CAD software will mostly use some kind of lower common denominator graphics layer like openGL. Webgpu is actually a decent alternative to that.
There certainly are optimisations that aren’t possible in web browsers today. Arbitrary wasm memory constraints and difficulties around cpu multithreading or simd for example. But CUDA kernels aren’t a realistic option for most cad software.