AI is subsidized by 70x. Prepare for price increases.

OpenAI subscribers get to spend $14,000 of OpenAI tokens each month for $200. For Anthropic, it’s $8,000 of token value for $200. That’s a 70x and 40x discount on those tokens, respectively. AI companies are losing billions of dollars and developers are moving forward with confidence that those prices will remain constant.

In a new report released this month, David Rosenthal (a veteran tech investor and researcher) dives into the economics behind those discounts and how they cannot last. The numbers are wild, even by AI’s standards. In 2025, OpenAI’s revenue is projected to be $13.07 billion while their costs are $34 billion, giving them a nearly $21 billion net loss. They’re burning through 44% of their revenue ($5.73 billion) on sales and marketing expenses. The low cost of AI subscriptions is based entirely on burning through investor cash.

1When the pricing changes

That day of reckoning is coming. According to Rosenthal, one company moved from a fixed subscription model to a tokenized billing system (where you pay for actual usage at fair market value). Overnight, their costs went up 7x. Despite not making any changes to their product or their usage levels.

That change is happening by design. With IPOs on the horizon, AI companies need to prove that their business model is sustainable. To do so, they are moving from heavily subsidized subscriptions to tokenized pricing. That change exposes the true economics.

Rosenthal paints a picture of scarcity. Even with zero operating expenses, top hyperscalers have negative returns on investment in AI until 2030 (Microsoft: -9.2%, Alphabet: -15.7%, Meta: -28.8%). The industry took on $3 trillion of debt, which will cost an estimated $309 billion per year in interest alone (and that’s before factoring in any productivity gains).

And one final number: For some companies, using AI is now more expensive than employing someone full-time. Paying $28,000 or more each month in token costs versus an average employee salary. If automating something is more expensive than the work itself, it’s not really automation. It’s subsidization.

2What this means if you're building

If you’re developing a product or internal tool that uses AI, this is the elephant in the room that most “building in the AI era” discussions ignore. Your assumptions about the underlying economics might change.

There are two implications here. First, there’s a near term opportunity. These discounts are still available. You’re paying less today for AI than you should, based on sustainability. In fact, now is the perfect time to stress test your cost structure. Calculate what your product would cost if tokens were 5x more expensive. If you can still make it work, great. If not, you have time to adjust.

Second, there’s a design principle. When AI gets more expensive, the value of intentionally using it goes up. Products that are judicious in their use of AI and only invoke it when there’s a tangible benefit (measured in some way) will fare much better than those that assume tokens will always be cheap. The opportunity is still there. But it’s fleeting.