AI companies are already profitable on marginal serving costs

The common narrative that AI companies are losing money overlooks the distinction between capital expenditures (data centers, training runs) and operational serving costs. OpenAI spent $25B in the first half of 2025 on only $4B revenue, but most of that was building out infrastructure for future training and inference. The marginal cost to respond to an API call is significantly lower than the price charged, meaning inference is already profitable.

Evidence comes from comparing prices of open-source models served by neutral providers — DeepSeek-V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens. Leading labs charge $2-5 for input and $12-25 for output, with only minor quality differences. This gap indicates either poor inference optimization at frontier labs or healthy profit margins. If funding dried up, current-gen models would remain viable and profitable indefinitely, albeit at slower improvement rates. The industry as a whole is not at risk of collapse.