Agree with you Anonymousgg regarding the write-up vs what was said. As for the analysis, Strix Point will have lower performance than Granite Ridge due to hitting a price point on size (yields) knowing it cannot be as 'fast' as a separate CPU and dGPU due to limited space/cost for memory channels and I/O since it is to be monolithic.
Strix Halo on the other hand, is rumored to be an essentially toned-down consumer version (only 256 bit memory interface) of the multi-chip MI300. It will be much larger in total Si area than Strix Point but aggregated from higher yielding chiplets. As a result, AMD could price it at say $550 since it would perform close to a $270 RX7600 plus a $300 7700X (forgot how many CPU cores it is rumored to have but guessing 8). Slightly less revenue than consumer buying those individually, but that is offset by a lower BOM since fewer PCBs, less total packaging & shipping costs, and maybe less supporting electrical components (VRMs, etc) since GPU is on CPU package that already has a 12 V, 5 V, rail, etc. That is, instead of 170W for a 7950X/7900X on AM5, imagine 70 W allocated to CPU with it's standard 128-bit DDR5 interface and 100 W to GPU with its own dedicated 128 bit memory bus.
Very similar design to what Tom mentions for the AI chips and server parts at the end of the video and to what others have mentioned before regarding only a single chiplet die on AM4. That is, use the other chiplet spot on the package for a on-package GPU chiplet.