AI, Tokens, and the Gathering Storm

There is a possible future in which the AI boom does not end with a robot butler bringing us tea.

It ends because somebody asks:

“Wait. Why is this worth a trillion dollars?”

OpenAI and Anthropic are now discussed less like companies and more like weather systems. Their valuations roll across the financial sky in terrifying shapes. Hundreds of billions. Maybe a trillion. Possibly, by next Tuesday, one bajillion, pending a secondary-market trade by a guy named Brad.

The trouble is that private valuations are not reality. They are vibes wearing a Patagonia vest.

A company can be “worth” an astonishing amount on paper because a small number of investors bought a tiny slice at a silly price. This does not mean the entire company could be sold for that amount. It means somebody, somewhere, paid the financial equivalent of £19 for airport Pringles.

Meanwhile, the actual business question remains awkwardly alive:

Can these companies make money without subsidising the product?

That is where tokens enter the story.

Tokens are the little bits of language AI systems process. Every prompt uses them. Every answer uses more. The longer the conversation, the more the machine chews through compute, electricity, servers, cooling, and what economists call “oh no.”

For years, users have been trained on the all inclusive buffet model: pay a monthly fee, consume wildly, and assume the robot is basically free. But AI is not free. It is a meter running in a basement full of expensive chips.

If OpenAI and Anthropic ever have to charge users closer to the real cost of usage, the magic may start to look like a taxi ride from Heathrow during a surge-pricing event.

This is the gathering storm.

The valuations assume heroic growth, massive adoption, and eventually wonderful margins. But the products still appear expensive to run, unpredictable to price, and hard to evaluate. A model might solve something in one shot. Or it might spend five minutes confidently rearranging the furniture in your codebase while billing you for the privilege.

So one possible outcome is simple: token billing arrives, users notice the true cost, usage falls, and the valuations start looking less like genius and more like group hypnosis with a term sheet.

Not because AI is useless. It clearly is not.

But useful and worth-a-trillion are different things.

A kettle is useful. Nobody is raising Series G funding for the kettle at a $900 billion valuation, although give Silicon Valley time.

The question is not whether OpenAI or Anthropic can produce impressive demos. They can. The question is whether they can build businesses large enough, profitable enough, and durable enough to justify the numbers now being stapled to them by investors who appear to have spent too long inhaling GPU exhaust.

Maybe they can.

But one possible outcome is that the great AI reckoning does not come from regulation, rebellion, or a sentient chatbot demanding dental insurance.

It comes from the bill.

And when the bill arrives, the industry may discover that the storm was not intelligence.

It was arithmetic.