I wonder why they haven't invested a lot more in the inference stack? Is it really that different from Google, OpenAI and other open weight models?