The other possibility is that we'll get cards that are very specifically designed just for the LLMs, basically ditching everything that is not strictly necessary for the sake of squeezing more compute / VRAM, and perhaps optimizing around int4/int8 (the latter is apparently "good enough" for training?).