You don't need clean data to train, just "reasonable" data. General bulk data from the wider world isn't the difficult part. AFAIU task specific RL environments, custom annotated datasets, and access to the raw compute necessary to actually perform the training are the primary limiting factors. The latter in particular is no joke. Most companies won't be able to afford the pretraining for a frontier sized model. Even "small" models can run into the millions USD IIUC.