They are saying that a hardware AI would be able to replace your GPU’s functionality, so yes, it would be the thing that game developers would use to output pixels to your screen.
It’s a bold statement in a way, but there’s solid reason behind wanting to make this happen.
Generative AI is currently capable of creating images that contain extremely complex rendering techniques. Global illumination, with an infinite number of area light sources, is just kind of free. Subsurface Scattering is just how light works in this model of the world, and doesn’t require pre-processing, multiple render passes, a g-buffer, or costly ray casts to get there. Reflections don’t require screen-space calculations, irradiance probes, or any other weird tricks. Transparency is order independent because it doesn’t make sense for it not to be, and light diffuses or diffracts because of course it does.
Modern high-end GPUs are hacks we have settled on for pushing information into pixels on the screen. They use a combination of a typical raster-based graphics pipeline, hardware accelerated ray tracing, and ai upscaling and denoising techniques to approximate solutions to a lot of these problems.
There are definitely things that developers would need in order for AI hardware to replace GPUs, such as any kind of temporal consistency, and significantly more control over the resulting pixels.
And to get both game developers and consumers to transition over you’ll need it to first be part of even low end GPUs for quite a while.
It’s not a terrible idea at its core. It’s definitely not 5-10 years out, though.