The methods for splitting weights across multiple chips are well established. Groq/Cerebras can't hold a model on one chip either.

Umm I have an extra 35, do you have layer 6?