[zamadatix]

Access to the NPU is done through DirectML which is open to anyone to use. The hard part is making useful AI people will pay for that runs on consumer devices themselves without a race to the bottom on the same few use cases (e.g. local translation). There may eventually be some bundling type concerns with the service being pushed by the OS over time though.

[wtallis]

Forget about "useful AI people will pay for"; the limitations of these NPUs and the immaturity of the software stacks necessary to target them means it's hard to get anything running well enough to be a cool tech demo. Good enough that people would use it for free is still a ways off, and good enough to be worth paying extra for may not even be possible on the NPU hardware Intel and AMD are currently shipping.