Can anyone with good insight explain why AMD hasn't traversed the CUDA moat yet? Surely they have been investing in building a similar layer and getting it ups

Can anyone with good insight explain why AMD hasn't traversed the CUDA moat yet? Surely they have been investing in building a similar layer and getting it upstream into the ML frameworks and intermediary compilation systems?