Because everything is converging on a backlog of huge efficiency gains established in research, waiting to be combined. Looped transformers, a whole host of diffusion techniques and new quantization techniques, maturation of ternary distillation and new ways to separate logic from stuff that can be looked up. It would surprise me if most frontier models were actually even that big at that point in terms of active params. I highly doubt it.

You asserted that it was never going to happen first