AI Is 70× Subsidized. The Price Hikes Are Already Here.

For $200 a month, OpenAI subscribers can consume $14,000 worth of compute. Anthropic's equivalent: $8,000 in token value for that same $200. A 70-times and 40-times subsidy, respectively — while AI companies lose billions and builders confidently plan on these prices staying put.

A detailed analysis published this month by David Rosenthal — a longtime tech investor and researcher — traces how these economics became possible and why they can't continue. The numbers are striking even for an industry accustomed to eye-watering figures. OpenAI's 2025 financials: $13.07 billion in revenue, $34 billion in costs, a net loss approaching $21 billion. The company spent roughly 44% of its revenue — about $5.73 billion — on sales and marketing alone. The entire value proposition of low AI subscription prices has been built on burning investor dollars at scale.

When the pricing changes

That reckoning is starting to arrive. Rosenthal cites a case where a company switched from a flat monthly subscription to token-based billing — where you pay for exactly what you use at real rates. Their costs increased 7 times overnight. Without changing what they were building or how much they were using.

This shift isn't accidental. As AI companies approach IPOs and need to show a path to sustainable unit economics, token-based billing replaces the heavily subsidized subscription model. The transition reveals the gap that subscriptions were masking.

The broader picture Rosenthal constructs is one of structural constraint. Even assuming zero operating costs, leading hyperscalers show negative returns on their AI investments through 2030: Microsoft at -9.2%, Alphabet at -15.7%, Meta at -28.8%. The industry accumulated roughly $3 trillion in debt, requiring an estimated $309 billion annually just in interest payments — a sum that no amount of productivity gains has yet offset.

One more data point: at some companies, AI usage now costs more per month than hiring a full-time person — $28,000 or more in monthly token spend, against an average employee salary. Automation that costs more than the labor it replaces isn't automation — it's a subsidy-funded experiment.

What this means if you're building

For anyone building products or internal tools on top of AI, this is the background context that most "building in the AI era" conversations skip over. The cost structure you're designing around may not persist.

Two things follow from this. The first is a short-term opportunity: the subsidies still exist. The prices you pay today are lower than they should be by any sustainable measure. This is actually the moment to pressure-test your cost basis — run the numbers on what your product would cost at 5× current token prices. If it still works, you're well-positioned. If it doesn't, you now have time to fix it.

The second is a design signal: when AI costs rise, the premium on using it deliberately jumps. Products that call AI only where it adds clear, measurable value will hold up much better than those built assuming tokens stay cheap forever. The window is still open. But it's narrowing.