The real problem with the AI industry is unit economics

Spent about two hours yesterday researching the economics of the AI industry.

I discovered that the problem with the industry is not innovation or producing new models.

The real problem, and one that literally no one seems to have solved, is unit economics.

SaaS businesses used to work like this: spend money, build a piece of software, license it or run a subscription model, and any new user comes at essentially zero marginal cost. That is why mature SaaS runs 80 to 90% gross margins.

AI breaks that model. Every new user, at least for now, comes with a real inference cost, so gross margins land closer to 50 to 60% instead. On the heaviest users it is worse. A customer paying $20 a month can generate $18 to $25 in inference costs during heavy reasoning tasks. At the extreme end, OpenAI reportedly generated about $3.7 billion in revenue in 2025 yet lost an estimated $5 billion, spending $1.35 for every dollar it earned.

Net contribution margins, in many cases, are still negative.

Think of the AI industry like a normal factory producing a product. You have cogs and raw materials, except here the materials, compute and inference, are extremely expensive, and you need them just to produce the final output for every single unit sold.

My analyst mind says it is broken... My engineering mind says it will be solved... and I don't know which I should listen to...