True, they have some interesting features. But at least if they use Windows ML, these apps can run on any hardware. The only difference is the performance that can be achieved.
They said something in the video you linked that the developers could use a larger model thanks to the performance of the Snapdragon Chip. And with Davinci, they said that the performance with NPU is better than with GPU.
What I’m trying to say is that these things are not a new invention of 2024. NPUs were introduced at least 7 years ago (for ARM CPUs). The Microsoft APIs have been there for three years. The only new thing is the marketing hype around AI.