AI feels cheap right now. You pay about twenty bucks a month for Microsoft (MSFT)-backed ChatGPT or Anthropic's Claude and get access to some of the most sophisticated—and expensive—software ever built. It's a great deal. Maybe too great.

Newly surfaced financials, reported ahead of potential IPOs, suggest that pricing might be an illusion. The math, at least for now, doesn't really work.

Take OpenAI. It's projected to spend a staggering $121 billion on compute by 2028, and losses remain steep even as revenue grows. Anthropic is on a similar, if smaller, path. The killer detail? For both companies, the cost of "inference"—the actual computing power needed to answer your prompts—already eats up more than half of their revenue. The more people use these tools, the more that cost scales. It's a growth model with a built-in cost escalator.

So why are we only paying $20? Because right now, profitability isn't the point. Growth is. Only a small fraction of users actually pay for these services. The vast majority of usage is effectively subsidized as companies execute a classic tech land-grab: get users hooked, dominate the enterprise market, build an ecosystem. Monetization comes later. It's the scale-first, profit-later playbook.

But that playbook works until it doesn't. When does the music stop? Probably when investors stop cheering for user growth and start asking for cash flow.

If serving answers already costs more than 50 cents of every revenue dollar, and you add tens of billions in upfront training costs on top of that, the current $20 price tag almost certainly doesn't reflect the true cost. When investor pressure shifts, pricing becomes the obvious lever to pull.

Don't expect an overnight doubling. The shift will likely be gradual. Think higher premium tiers ($30, then $50), pay-per-use models for power users, or maybe ads creeping into the experience for free users. The trigger won't be a lack of demand; it will be the simple financial reality that you can't sell a Rolls-Royce at a Toyota price forever.

The core issue is a business model mismatch. AI is being sold like a software-as-a-service product with predictable margins. But it's being built like massively expensive infrastructure—more like building and maintaining a global power grid than developing an app. At some point, those two realities have to meet. And when they do, your monthly bill might start to look a lot more like what it actually costs to run.