And doesn't account for cuTitle, NVidia's new API infrastructure that supports writing CUDA directly in Python via a JIT that is based on MLIR.