Seen from the markets, AI is only promises and crazy valuations. Seen by those who build with it, the industrialization of intelligence is beginning. Hence the debate on the “AI bubble”.
L’artificial intelligence is it a bubble? The question obsesses TV sets, forums, committees and markets, from the United States to Europe, and perhaps even more so France. Quite paradoxical fascination in a country which, in this revolution largely carried out elsewhere, remains for the moment a passenger rather than a pilot.
The diagnosis is always the same: investments would be disproportionate, valuations absurd, so AI would be a bubble.
There is truth. Businesses will disappear, capital will be destroyed, markets will perhaps correct suddenly. But this reasoning crushes three questions that have nothing to do with each other.
- Is the American stock market too expensive?
- Is the capital invested in AI going to the right companies, at the right price, at the right time?
- And can AI absorb a massive share of the world’s cognitive work?
We can judge the market to be too expensive and the capital poorly allocated, and still think that the answer to the third question is yes. A technology can transform the world and ruin those who bet on it too soon, too expensively or in the wrong place. A 30% correction in the Nasdaq would not prove that AI does not create value. It would prove that the markets are correcting. Like they always did.
Before we talk about valuation, let’s talk about value
A sustainable business does two things: it creates value, and then it captures a fraction of it. What it captures can never exceed what it helps create.
Value captured ≤ value created.
The whole divide is there. If you think that AI will stop at summarizing documents and correcting emails, the value created is low, and the captureable value even lower. Even 100% of a small value is still a small value. With this assumption, current investments are absurd. The reasoning holds.
But if you think that AI can absorb a massive share of cognitive work, conduct research, operate businesses, and accelerate its own development, the upper bound explodes. This does not justify each valuation or each data center. This means that the potential market no longer has anything to do with that of productivity software.
The debate over the “AI bubble” looks financial. It is actually based on a radical divergence in the estimation of the value that this technology will create. And this estimate depends directly on the AI that each person experiences.
If you use ChatGPT three times a week to summarize a document or correct an email, the AI is a practical assistant, sometimes astonishing, often fallible. Faced with hundreds of billions invested in infrastructure, finding the discrepancy absurd is perfectly rational.
But if you have agents working hours to code, test, deploy, sell and operate products, you’re no longer looking at the same technology. You see one person accomplishing what yesterday required a team. You see tasks arise that would never have been carried out, because they were too expensive.
It’s not an intelligence problem, it’s a sample problem. Two brilliant people can come to opposite conclusions because they don’t experience the same AI. And it is difficult to estimate the future value of a technology based on a truncated version of its present.
A million dollars in ARR in 55 days, without ever exceeding three people
Let’s clarify where I’m talking about. I am 26 years old, with training in applied mathematics and artificial intelligence at Polytechnique, a stint at Y Combinator, and I have been building companies with this technology for several years. If the AI succeeds, I succeed with it. The bias is real.
But a bias does not cancel an observation.
I started my last business alone. Fifty-five days later, the company surpassed $1 million in annual recurring revenue, calculated by annualizing active paid subscriptions. There were never more than three of us, including me. Such speed, at this workforce, would have been nearly impossible without AI agents across all functions of the business.
This does not prove that Nvidia is at the right price or that each data center is profitable. This shows that a tiny team is now achieving a speed and scale that was out of reach two years ago.
My company (NanoCorp) takes this logic a step further. A human gives an idea, a direction, a budget. An agent creates the company and operates it within the limits set for him: he writes and deploys the code, connects a domain, accepts payments, sends emails, launches advertisements. With an explicit objective: to earn money.
More than 16,000 users have already launched more than 22,000 businesses on the platform. Together, these self-employed businesses have generated nearly $11,000 on the Internet.
Eleven thousand dollars is nothing in the economy. But this is no longer a demo. These are agents who find real customers, sell them real products and collect real payments. An agent identifies the target audience for a product, creates the ads, launches the campaign on Instagram, analyzes performance with site telemetry, modifies the product and starts again. A user reports a bug? The agent fixes the code and redeploys.
From this vantage point, AI doesn’t look like a distant promise. It looks like a change of scale in what is possible to build.
Progress disappears when you always look at the same task
I am often told that the new models no longer seem that much better than the previous ones. This is normal, if your personal benchmark remains writing an email.
The first models wrote mediocre emails. Then the task was mastered. Once the email is “resolved”, a model ten times more powerful does not write an email ten times better: the visible gain tends towards zero. The value moves elsewhere. A better model accepts longer, more ambiguous, more difficult tasks. Then complete loops, with a goal, tools, memory and the ability to check one’s own work.
I see AI as the electricity of the 21st century. The formula seems grandiose, I admit: everything becomes clearer when we look at how electricity has really transformed industry. The first electrified factories replaced their steam engine with an electric motor without changing anything else: same centralized architecture, same belts to distribute power to the machines. A gain, not a revolution. The break came when manufacturers rebuilt the factory around electricity: one motor per machine, a space organized according to the flow of production and not according to the transmission of force.
This is exactly where we are with AI. Many companies add an assistant to processes designed for humans: an electric motor in a factory designed for steam. This is the complete opposite of what we’re doing with my company: not asking AI to marginally help each function, but designing a business that it can operate end-to-end.
Reasoning with a fixed task therefore hides the essential. Additional value does not come from better execution of what we were already doing. It comes from the expansion of what it becomes economically possible to do. This is the Jevons paradox applied to intelligence: when a unit of cognitive work becomes less expensive and more capable, we do not consume less of it. We are inventing new uses for it.
I measured it on my own data. Before o1, OpenAI’s first reasoning model, I sent a few hundred messages per month to ChatGPT. Models of reasoning have not replaced this usage. They opened up a whole category of conversations that was added to the first: longer, more complex, devoted to coding, to research, to problem solving. My monthly volume increased to around 1,700 messages in May and June 2025, then to over 3,000 in the summer and fall. Latest example: I had an agent work for more than a week to rewrite a video game in JAX and train small models who now play better than me. I was tired of losing. No one would ever pay a team to do that. An agent, yes.
The economic shock will therefore not only come from the tasks that AI replaces. It will come from all the work that did not exist yesterday, because it was not worth its cost.
Usage does not give reason to everything, but it provides decisive information
Builders do not have a monopoly on the truth. Eternal optimists, we can overestimate the speed of change and confuse a successful product with a sustainable business. Using AI intensively does not say who the winners will be, nor their margins, nor their fair valuation.
But it provides access to information that financial ratios and occasional demos poorly capture: the amount of work that can now be delegated to a machine.
You don’t need to believe in AGI in 2027 or humanoid robots in every home to see gigantic value creation. All it takes is for AI to absorb a significant fraction of software, customer support, sales, marketing, research, legal, finance and operations.
I will go further. If models definitively stopped progressing today, most of the transformation would remain ahead of us. The best current models are barely used: a tiny minority of users push them to their maximum capacity, and the vast majority of the economy has not yet touched them. At constant technology, we have years of acceleration left.
I am not paid to determine whether markets are in a bubble. I am an entrepreneur: my job is to create value. And I’ve never been able to create so many, so quickly, with so little.
AI enables a society where an idea and judgment weigh more than start-up capital, team size or access to the right institutions. A society where many more people have the power to build.
Some will continue to comment on the price of this revolution. Others will build the value that will ultimately give it its price.
