I’ve seen comments saying that many foundational model providers like DeepSeek haven’t done a full pretraining in a long time. Does that mean this use of chips is in reference to the past?

Whilst there aren't many papers on the matter, I would guess that pretraining from scratch is a bit of a waste of money when you could simply expand the depth/width of the 'old' model and retrain only the 'new' bit.

Even if they're not doing full-from-scratch training every cycle, any serious model updates still soak up GPU hours