Stress-testing AI inference profitability

I built a small simulator to stress-test the unit economics of AI inference. The question I wanted to isolate is simple: under what assumptions does frontier AI inference become profitable enough to justify the current capex cycle? My current read is that AI inference can become very profitable, but not just because inference gets cheaper. The profitable case needs several assumptions to line up at the same time: - paid adoption scales quickly - GPU capacity does not outrun demand by too much -