They already do. Local produces maybe >5% of the volume needed. It's NOT prefer or not. It just doesn't exist.
Some people believe that the key moat for AI is the model, and under this model (hah) torpedoing NVidia would be bad for China and good for the U.S. because they won't have access to cutting-edge NVidia hardware for model-training. Other people believe the key moat for AI is the hardware, and under this model torpedoing NVidia would be good for China because it'd allow Chinese hardware makers to catch up and bad for the U.S. because it destroys our key competitive advantage.
Interesting that a policy decision depends so heavily on a question of fact, but that isn't unique to this issue. (I suspect that left and right would agree on abortion if they could agree on whether personhood begins at conception or birth, but they don't, so). There's also an element of time preferences here too: it's likely that model training is dominant now, but hardware will be dominant when you project out a few years, and so both factions could be right but on different timescales.
They would prefer their companies buy local, as it will advance their local semiconductor infrastructure.
However, these companies would prefer Nvidia if they are allowed. All the advances in the past few years are due to advanced chips being restricted, so they had to step up.