If NTC is true, and is in the PS6 SDK (which, us unlikely being that Cerny is going with Universal compression, albeit nothing stops them from also using the neural arrays for their own NTC equivalent)
EVERYONE will use it. Immediately. In fact, it would even be a requirement. A compression factor of 6 - 10x, is impossible to ignore. The PS6 will likely use the inference-on-load variant, which amounts to a 6x compression factor. Being able to store 18GB of textures using only 3GB of space... is too much to ignore, and this approach almost comes at no cost at all.
However, what sony seems to be going with is UC, not NTC. they could do both though.
The benefits are certainly there, but even something less disruptive like Kraken took a while for everyone to implement on PS5 even though the benefits were massive right out the box for not much work. In this case, you literally have to redesign how textures work. You need to train individual MLPs for every single texture, identify the most optimal latent texture and not ship the original texture at all. That's a significant workflow change that most devs and artists wouldn't even know how to do without proper training and trial and error.
And NTC cannot be used as-is with traditional anisotropic filtering. So they either need to use stochastic filtering, which is even more bleeding edge with tons of drawbacks yet to be overcome in research, fall back to NTBC or do something creative like a pre-filtered inference (like Ubisoft did recently).
And if they instead use NTBC, then the benefits of lower VRAM usage are gone.
To truly reap the benefits of NTC, you need to implement something like this Ubisoft approach:
https://www.ubisoft.com/en-us/news/...exture-compression-in-assassin-s-creed-mirage
They released a paper on this approach a little while ago. A really good read and likely where all NTC implementations will go:
https://arxiv.org/pdf/2311.16121
But to do this across the board and actually ensure good performance and production readiness, you have to have Neural Arrays and Cooperative Vectors, both of which won't be available till devkits go out next year. Performance would suck without them. Until then, they would just be experimental in nature. It's certainly the future, but launch window is unrealistic imo. I can see a couple of Sony studios (like Insomniac) and a couple of 3rd parties jumping onboard right away, but the majority will likely take years.
On the topic of UC, I have no idea as there is no material out there. My understanding is it will be in the path of the CUs and VRAM. Not SSD and GPU (i.e. the I/O complex). My guess would be they would likely work hand in hand.