The internet's supply of high-quality text data for training large language models has been largely consumed.