those gross profit margins aren't that useful since training at fixed capacity is continually getting cheaper, so there's a treadmill effect where staying in business requires training new models constantly to not fall behind. If the big companies stop training models, they only have a year before someone else catches up with way less debt and puts them out of business.

Only if training new models leads to better models. If the newly trained models are just a bit cheaper but not better most users wont switch. Then the entrenched labs can stop training so much and focus on profitable inference