that makes sense to me in a conceptual sense

however inference is very profitable and plummeting in cost for a given point on the intelligence curve, and nvidia gpus can serve different models so they are protected post-buildout

You are describing the justification that NVDA is using to explain their behavior; they see it as something of a 'bridge-loan' until the LLM business model reaches steady-state. The problem is that this explanation has been used for many bubbles, where companies mis-categorize ongoing costs as one-time expenses.

> however inference is very profitable

Is it? OpenAI and Anthropic are burning cash faster than anyone has ever shoveled cash into a furnace.

> however inference is very profitable

Is there any actual evidence of that?