Likely because they don't make inference accelerators, they make GPUs. It's entirely plausible that someone could optimize the hardware pipeline better by incorporating less complex logic; I wouldn't expect Nvidia to try that though, not with their main products.
