If you've been reading me from the start, you've probably noticed that many of the predictions and intuitions we developed have come true. Between the explosion in the price of RAM and of online services (techflation), and the fact that the topic of 2026 would not be digital sovereignty but cognitive autonomy, I didn't expect recent events to take such an explicit turn.

Two weeks ago, I told you about the embargo on intelligence. Washington had just blocked Fable 5 and Mythos.

Since then, the launch of GPT 5.6, which OpenAI's teams had announced to us in Paris, is only being rolled out in dribs and drabs. One thing is certain: the two most powerful models on the market are no longer available in Europe. The cognitive splinternet is a reality.

This ban comes at the worst possible time.

First of all because these players have IPOs planned and this complicates their revenue lines. OpenAI, for that matter, has just pushed its IPO back to 2027, citing complicated conditions.

But, in my view, that's not the real issue: this decision comes just as the question of token usage, the AI companies' golden goose, is starting to surface. After token maxxing, we're entering a new phase of efficiency, even budget cuts, which could only have been offset by the arrival of models that are no longer available.

Let's recap.

This week, the site 404 Media published the recording of an internal meeting at Accenture. In it, you hear the firm's executives worrying about an AI bill that is getting away from them, and pointing to a culprit nobody expected: it's not the engineers who are blowing up the bill, it's the non-technical employees. The ones who ask AI to turn a PDF into a presentation or reformat a document, tasks that, only yesterday, cost nothing but time.

Suddenly, we discover two things: that AI has accelerated a form of intellectual laziness, which is normal, it's one of the platforms' hopes for creating addiction; but, and we'll see this later, the existing work architecture, such as PDFs, isn't made for the agentic world and remains expensive in tokens.

The sticker shock of enterprise AI

The Financial Times recently explained that Uber had burned through its entire AI budget for the year by April, and now caps each employee at $1,500 per month per tool (which, in my view, is idiotic).

Axios, for its part, reported that an unnamed company had spent $500 million in a single month on Claude, having failed to set the slightest limit. Amazon, Walmart, Cisco and even Meta, the king of token maxxing, have slammed on the brakes.

This is AI's sticker shock: the moment when companies discover that what they thought was a magic tool, sold at a flat rate or absorbed into innovation budgets, becomes a variable, measurable and sometimes uncontrollable bill.

Accenture, which now sells the remedy with its Token IQ offering, detailed the audit of its own internal usage: $110 million a year on Claude, 79% of it generated by just 20% of users, and 61% concentrated on Opus, the most expensive model. Nobody inside saw that meter running.

We explained in a previous newsletter that, for two years, AI was sold at a flat rate, a fixed price for unlimited use.

Then Anthropic and OpenAI switched to per-token billing, by the quantity of data processed. Workato's CIO puts it best: on the day of the switch, his spending was multiplied by seven.

The token subsidy was becoming unsustainable for several reasons: the first, obviously, is that data center capacity was no longer sustainable, and the bad subsidy practices also had to be “cleaned up” before the IPO.

Why AI is getting cheaper and more expensive at the same time

On the face of it, the price of intelligence has kept falling. Opus went from $15 and $75 per million tokens to $5 and $25, a two-thirds drop in a single generation. But the new models and the agentic tools ask the models for tasks that are more and more complex, longer and longer, more and more chained together. So, in the end, more expensive.

Now the question on the table is to spend more, but get better. Goldman Sachs, quoted by the Financial Times, is betting on a 24-fold increase in token consumption by 2030. Nobody is betting on less AI for now, but everyone is starting to wonder whether “this money is buying value.” For now, that remains hard to explain.

We weren't paying for intelligence but for inefficiency

My intuition is that a large share of the current bill buys neither intelligence nor value. It buys friction.

Take the PDF again. A 90-page document sent as plain text costs about 56,000 tokens. The same one as a PDF costs four times as much, because the model doesn't read the text, it photographs every page. The PDF was invented to freeze a rendering on a printed sheet, at the expense of the text's structure.

Perfect for a human reader. Absurd when the reader is a machine.

And the format is only a symptom. Don't forget the pointless back-and-forth, the clumsy way we address the models, the agents we let run without guardrails: all that inefficiency is billed to us at full price.

I'm working on a new project on these questions, which I'll announce soon.

The answer isn't to cut AI, but to build what's missing between the human and the machine. That's what we're going to talk about here for subscribers. For free readers, perhaps an annual subscription will pay for itself with the tips and information we give you a little further down.

Go paid!

I learned again last week that a reader who started running a local version of AI had the good idea of buying his Mac just before the price increase we had discussed several times. Result: €2,000 saved, not bad!

Since June 2026, this newsletter's goal is changing a little.

From now on, we're going to talk about cognitive autonomy, and learn to make better use of the tools at our disposal, because we won't necessarily have access to everything. We'll keep observing the world with realism while most observers (and, sadly, many professionals) seem lost, even dazed, by a situation that, if you haven't thought it through with enough information, seems inextricable.

Thank you to everyone who supports us and who still wants to understand how the digital world (or the world, full stop) is going to evolve over 2026. There will be a lot of twists and a lot of breakage, but we'll keep exploring the future together.