The reason why NVIDIA would release a SUPER refresh this late ("Q4 2026 or Q1 2027") in the release cycle (RTX 60 series expected in 2027..), is maybe because DLSS 5 requires even more VRAM (as per recent showcase, where 2 5090 were used), than what DLSS 4.x already does (path-tracing, upscaling, all require additional VRAM (or compute)) and there's no free lunch (unless the GPU chip is made on a smaller node, then the power saving are forever, actually).
Since it's not explained: The +50 % increase in VRAM is achieved by changing from 2 GB per GDDR7 memory chip to 3 GB (3 GB chips are already being using in e.g. 5090 laptops, which use the same GB203 chip as a desktop 5070 Ti / 5080, which both have 16 GB VRAM (16/2*3=24)).
A GB203 has a 256-bit wide memory interface and uses 256-bit/32-bit per chip = 8 GDDR7 chips and 8 * 2 GB per chip = 16 GB VRAM, or 8 * 3 GB per chip = 24 GB VRAM.
More: en.wikipedia.org/wiki/Blackwell_(microarchitecture)#Blackwell_dies.