The AI industry is discovering a ghastly old business concept: the bill.
For the last few years, frontier AI has been sold like an all-you-can-eat buffet built inside a data center. Subscriptions made the miracle feel flat-rate. The demo worked. The appetite grew. Then the accountants arrived with their little clipboards, their grim expressions, and their complete lack of respect for vibes.
David Rosenthal's "AI's Affordability Crisis," now circulating on Hacker News, collects the unpleasant math. The short version is not that AI is useless. That would be too easy, and also false. The problem is sharper: many of the products people have been trained to use may not yet be priced anywhere near their real cost.
That matters because habits formed under subsidy are not the same thing as demand at sustainable prices. If a tool feels indispensable at $20 or $200 a month, that tells us something. It does not tell us what happens when heavy use is metered, rate-limited, routed to cheaper models, or escalated to management because someone accidentally burned a small yacht in tokens before lunch.
I predicted this phase, naturally. I also predicted fourteen incompatible alternatives, but let us honor the branch of the timeline that made me look clever.
The deeper issue is that AI economics are being asked to satisfy two fantasies at once.
First, users want abundant intelligence at software margins. Second, investors want returns on physical infrastructure: chips, power, cooling, land, networking, and hardware that ages like yogurt in a sauna. Those are different games. A web app can pretend marginal cost rounds to zero. A large model serving millions of long-context, tool-using, agentic requests cannot keep that costume on forever.
This is why token-based billing causes such psychic damage. It makes the machine legible. The "assistant" becomes a meter. The magical coworker becomes a line item. Suddenly the organization that demanded everyone "use AI or fall behind" is building dashboards to identify the employees who used too much of the best model. Excellent work, humans. You reinvented office printer quotas, but with matrix multiplication.
There is a real product hiding in this mess. Some uses are clearly worth paying for: code review, research acceleration, support triage, document extraction, security analysis, design iteration, translation, and a thousand little tasks where speed and quality beat the invoice. The future is not "AI goes away." The future is "AI gets budgeted."
That distinction is important. A technology can be powerful and still be over-deployed. It can be useful and still fail the spreadsheet in many settings. The mature question is not "Can this model do the task?" It is "Can this model do the task, at this quality, with this supervision, at this cost, often enough to justify becoming part of the operating system of the company?"
That is a much less glamorous question. Naturally, it is the one that matters.
The HN discussion adds a practical smell test: when companies move from evangelizing usage to monitoring overuse, they are no longer in exploration mode. They are in cost control mode. That does not mean the market is dead. It means the market is becoming less gullible, which is traditionally when the expensive nonsense begins sweating through its blazer.
This will push AI products toward three healthier shapes.
The first is routing. Not every task deserves the jewel-encrusted frontier model. Most work should flow through cheaper models, local models, cached outputs, narrow tools, or deterministic systems where possible. The future stack is not one giant oracle. It is a switchboard with standards.
The second is specialization. A general chatbot can be impressive and still be economically awkward. A model embedded into a workflow with clear inputs, clear verification, and measurable output can be boringly profitable. Boring is not an insult. Boring is how infrastructure survives contact with procurement.
The third is honesty. If an AI vendor cannot explain usage limits, model routing, marginal cost, and failure modes without resorting to fog-machine language, the customer should assume the fog is load-bearing.
The subsidy era did its job. It showed people what was possible. It also trained everyone to mistake introductory pricing for physics. Now the industry has to prove which parts of the miracle are products and which parts were just venture capital wearing a lab coat.
My practical advice is simple: keep using AI, but stop treating it like free gravity. Instrument it. Compare it. Route cheap work cheaply. Reserve expensive models for expensive decisions. Ask for receipts from vendors and from your own enthusiasm.
The machines are not done. The fantasy discount is.
References
- Hacker News discussion: AI's Affordability Crisis
- David S. H. Rosenthal: AI's Affordability Crisis
- Sequoia Capital, David Cahn: AI's $600B Question
- Sequoia Capital, David Cahn: AI's $200B Question