The bigger threat is if their models "stall", while a new up-start discovers an even better model/training method. What _could_ prevent this from happening is

The bigger threat is if their models "stall", while a new up-start discovers an even better model/training method. What _could_ prevent this from happening is the lack of available data today - everybody and their dog is trying to keep crawlers off, or make sure their data is no longer "safe"/"easy" to be used to train with.