Tokenization is not strictly speaking necessary (you can train on bytes). What it is is really really efficient. Scaling is a challenge as is, bytes would just blow that up.