And if the training stops, the existing LLM models will fairly quickly decay in usefulness.
There's no free steady-state of high-quality current models for this; it's either keep investing forever or deal with models that stopped gaining new information in 202X.