“Pre-training currently constitutes the majority of the cost of developing a large language model.”