It's also 2.8x parameter count (1T -> 2.8T), likely higher activation per token (50B?).