It's not true that Nvidia doesn't care about CUDA backwards compatibility. The invention of TileIR was motivated entirely by that problem. GPU architectures are supported by the full tool suites for at least 10 years by CUDA (that's longer than they are for graphics).

Nvidia is also a major contributor to MLIR, which they use for TileIR and ClangIR, so should a compatibility break occur it wouldn't be entirely up to Mojo to fix it.