I've been speaking to them about it. It's super sophisticated and extends the work done by the Coalition on Gears 5.
With Gears they trained an AI to look at SDR and HDR output frames of various games and understand what the difference in colour space transform was:
So HDR became the ground truth and SDR was the image that was derived from that.
The AI could see how HDR to SDR had and once it had been trained with enough data would actually be able to do the opposite. To take the SDR image and make it look like the same HDR image.
Trained with enough data you can then take an SDR game and perform that same inverse tone mapping.
They are working with an even more sophisticated setup than the software only approach of Gears 5.
Now the advantage with doing it at a source level, is that it is possible to have access to the original 11bit or 16bit data including, rather than have to apply it as a post-process to an 8bit image - which doesn't contain enough data to avoid banding.
Also because of it is happening at source, it is far faster and latency-free compared to doing it in other ways (or having a TV do it)
The Solution for the Xbox Series X goes further and all of this happens on the display controller so there is 0 affect on memory and GPU performance , working in HDR normally has a minor hit on these things. Which is actually why the S is overlocked vs the XB1 OG.