It seems to me that secure inference and horizontal integration is their main product, and the complement is the LLM itself.