The Expensive Intelligence Bubble

Revision, October 3, 2026 — This essay has been shortened and reorganized to remove repetition and tangential passages. Its central argument is unchanged.

I am currently developing Équilibre, an accounting application that has become much more ambitious than I initially expected.

Its core must keep accounting entries consistent, preserve their history and allow corrections without destroying the audit trail. An architectural error can remain invisible for a long time: the software runs, the numbers look plausible, but something is wrong.

To build its foundations, I used GPT-6 Astra, which OpenAI describes as its most capable model. Then I began working with GPT-6.1 Sol, using a high reasoning setting. The results remained satisfactory for the tasks I was assigning it, and I was using considerably less of my quota.¹

I was glad to have Astra to design the core of the system. But once the foundations were in place, using the most powerful model for every routine change sometimes felt like hiring a chief engineer to tighten every bolt.

API prices give a sense of the difference: for identical text volumes, Sol’s rates are one-fifth of Astra’s.² That ratio does not measure my quota savings, but it helps frame the choice I face: what capability is worth paying for to do the work at hand?

The most powerful model is not necessarily the most useful.

This may seem like a modest observation. Yet it raises a much broader question: what happens to investment in AI if a growing share of intellectual work becomes inexpensive?

What we are really buying

The race to develop AI is usually told in terms of capabilities: more powerful models, more computation, new data centres.

For the user, though, the decisive measure lies elsewhere. How much does a useful result cost?

Finding a bug, programming a feature, reconciling a bank transaction: that is what I am actually buying. The cost includes checking and correcting the result. A plausible answer is not yet a completed task.

A capability that requires an expensive model today could become available tomorrow through a more efficient one. What was exceptional would become commonplace, then inexpensive.

Pushing the frontier forward and making its capabilities accessible are two forms of progress. The first makes headlines. The second could reshape the economics of AI just as profoundly.

Useful infrastructure can be a poor investment

The financial risk then takes a different form.

Data centres could be heavily used, applications could multiply and AI could genuinely transform the economy, while delivering less than some investors expected.

We can be right about the importance of a technology and wrong about the price worth paying to invest in it.

If the cost of useful work falls quickly, suppliers may have to sell much more to generate the revenues they anticipated. Their facilities can run at full capacity without earning enough to cover operating costs, financing and replacement.

That is not enough to establish that a bubble exists. If costs fall faster than prices, or demand rises sufficiently, the investment can remain highly profitable. Everything depends on what was paid today for the revenues expected tomorrow.

The fragile assumption would be to confuse the value of the service with the profits its suppliers can retain. A technology can become indispensable precisely because it becomes affordable.

Cheaper, so used more often

There is still an objection: falling costs could send demand soaring.

As early as 1865, William Stanley Jevons argued that using coal more efficiently could increase its consumption by opening up new uses.³ That is the rebound mechanism: additional uses absorb some of the savings, and sometimes more than offset them.⁴

If a task goes from a hundred units of computation to twenty, but we perform it ten times as often, total consumption doubles.

I already see this mechanism on my own scale. Less expensive AI does not simply encourage me to save my quota. It makes tasks reasonable that I would have hesitated to assign to the most expensive model: more tests, more checks and more small iterations.

Unit cost falls; usage rises.

But that increase does not settle the question of returns on its own. We still need to know what price these new uses will command and who will capture the benefits.

The technology and its price

Railways offer a useful reminder. In Britain in the 1840s, Railway Mania sent share prices soaring before a long decline wiped out fortunes.⁵ Yet railways continued to transform the economy.

The reality of the technological revolution did not protect investors from getting the price wrong.

AI could follow a similar path. Its success could come from capabilities becoming affordable enough to enter a multitude of activities. Some of the value would then accrue to those who use them, without generating the returns expected by those who had financed their development.

We pay a great deal of attention to the power of the next generation. We should pay just as much attention to the price at which existing capabilities become accessible.

The AI bubble, if there is one, could therefore rest on a paradoxical conviction: having understood that intelligence would become ubiquitous, while betting that it would remain expensive.