Does not seem that complicated. OpenAI basically had to do a lock-in deal with Microsoft/Azure at the time, and they pioneered this circular funding hyperscaler deal structure so there were some rough edges.

Anthropic (all ex Open AI) knew the negatives of the deal, so they made a slightly better deal with AWS, not a full lock in. They also grounded it in hardware from the start, ie. being the flagship customer for Trainium, the flagship customer for external usage of TPU's.