If the cheapest 5070 Ti 24GB Supers hit our shores at around the R17 - 18k mark in Q1 of next year, then I suppose it's not going to be a wise purchase to get a 9070 XT 16GB anytime soon at R16k or 5070 Ti 16GB at R17k?
The performance of the cards themselves in the short to medium term is not so much the factor to consider, it's the extra longevity that the 24GB cards will provide?
The Team Red shills still say that there will be 9090 XTX or whatever coming. All that it is, in my opinion, is a rumour. From where I am sitting, it is only "Redstone" that AMD is holding in its deck to retain some mindshare. FSR 4 will also be rolling out to RDNA 3, and now with them having threaded queueing at a driver level, the GPUs might improve a bit. This will mitigate RDNA 3 users upgrading to RTX 50. I am being speculative, but the writing is within their own preview drivers.
All that said. I do expect RDNA 5 (that will not be UDNA per se) to be announced much sooner than anticipated.
The issue Nvidia has, not only with RTX 50, but also with RTX 40, is that its technology stack is being adopted too slowly. Though by and large, this also depends on where DXR 1.2 is... All in all, RDNA 3/4 and RTX 40/50, maybe 30/20, stand to win in 2026. The Vulkan stack is also integrating similar tech stacks to DXR 1.2. This is all good, since we will have very good texture compression, which will be better on newer architectures, better ray tracing/path tracing compute, as well as better shader compute. It is possible that some GPUs could improve by up to 15%. It is important to know that this is dependent on compilers and developers adopting these tech stacks.
Yes, 24GB GPUs are better than 16GB GPUs, but texture compression will help alleviate these pains. For the most part, it is more to sell and utilize localized AI.