The financial fragility of the AI industry

Translated from the original Spanish article.

At the AI party, the music is still playing. Through the ballroom windows, exuberant fireworks paint the sky with joy. Everyone laughs and dances: some thrilled about the future; others somewhere between curious, surprised and afraid.

Tokyo.

At ten o’clock on a Monday morning, Charles Prince walked into the room and shook hands with the Financial Times journalist Michiyo Nakamoto. The greeting was cordial, though Prince knew that didn’t mean the interview would be just as comfortable. He didn’t mind; he knew exactly what he thought.

It was July 9, 2007. That day, Charles “Chuck” Prince, CEO of Citigroup, brushed aside fears that the music of the leveraged buyout boom, fueled by cheap credit, was about to stop. He denied that Citi was pulling back from its positions and played down the first cracks in subprime mortgages, because liquidity was still putting out every fire.

“Aren’t you going to pull back?” Nakamoto pressed him.

That question produced his famous line: “When the music stops, in terms of liquidity, things will be complicated. But as long as the music is playing, you've got to get up and dance. We're still dancing.”

They kept dancing.

Weeks later, in August 2007, the credit crunch broke out and, with it, the road to the 2008 financial crisis opened up.

Prince resigned a few months after that interview, in November 2007.

This past June 28, 2026, just over a week ago, the BIS—the Bank for International Settlements, the central banks’ bank—published its Annual Economic Report. Its first chapter, titled Progress and peril, sets out a reading of the financial risks surrounding the artificial intelligence industry. The BIS is one of the most prudent and respected institutions in global finance. Its General Manager, incidentally, is the Spaniard Pablo Hernández de Cos, former Governor of the Bank of Spain.

The report offers a prudent, responsible reading. But a necessary one. And it’s worth discussing, because it doesn’t talk about one threat, but about four pressure points that have been building up. Let’s pause for a moment on each.

i. The risk of persistent inflation

Persistent inflation is the risk that prices, across the board, keep rising. All it takes is a scare, an oil price spike or a conflict, for the warning signs to flare up again and for central banks to have to raise rates once more, making credit more expensive worldwide.

ii. The sustainability of AI investment

On the sustainability of AI investment, the report asks whether the flood of money now being poured into this technology can be kept up. An enormous amount is being spent, much of it on credit, betting on profits that have yet to arrive, and if they’re slow to come or disappoint, that investment could grind to a sudden halt. And, by our very nature, this investment cycle is sustained by a narrative that breathes life into certain expectations. Well-founded, perhaps, but a narrative all the same.

iii. Ever-growing financial vulnerabilities

Financial vulnerabilities are the cracks in the market itself. Valuations are very high and stock markets are expensive, tech stocks above all, and investors have grown used to everything going up, so if fear returns, the fall could be steeper than usual.

iv. Weakening fiscal positions

Weakening fiscal positions refer to the state of public finances. Many governments carry such high debt that they have barely any room to react if something goes wrong. This speaks to the strength of the structural public safety net.

These are four distinct risks that, on top of that, are interrelated. When one gives way, the others are left more exposed.

The last three, for my purposes here, paint the same picture.

A great deal of euphoria resting on foundations that deserve a closer look. First, let’s put a number on that euphoria: the five largest hyperscalers (Alphabet (Google), Amazon, Meta, Microsoft and Oracle) are set to spend more than a trillion dollars on AI infrastructure between 2025 and 2026, more than they earn, and they’re already issuing debt to finance it. The BIS calls it, bluntly, “exuberance,” and warns that disappointing returns could turn the investment boom into a prolonged bust. As if the echo weren’t loud enough, the report itself compares the moment to the nineteenth-century railway mania and to the dot-com bubble. In other words, a real technological advance that draws in more capital than the returns can justify.

Because, of course, all this money moves on a promise. Sam Altman summed it up, half-jokingly, in front of a room full of investors, when he explained that he wasn’t promising profitability. What he promised was that, once they had AGI, “they would ask it how to get it.” The promise, at bottom, is that of a hyperproductive economy. As Elon Musk repeats whenever he gets the chance, an economy in which, within a few years, “work will be optional.”

But hyperproductivity, for now, is nowhere to be seen. And this is the interesting part, because several studies point in the same direction, and this time it’s the BIS itself saying so. At the level of a specific task, AI seems to save time clearly—though even that could be questioned—somewhere between 20 and 50%. But when you look at the aggregate effect on productivity, it stays below 1%. Why? The report attributes it, in no uncertain terms, to the difficulty of adopting the technology at scale and integrating it into production processes. The key factor, in short, isn’t the model’s capability. It’s putting the model to work.

We’ve experienced this firsthand, working with autonomous agents since January 2026. Here’s my take on one of the big promises: AGI could turn out to be, in practice, irrelevant.

Simply because it can’t be reliably integrated into a production system. At Kocoro, while building a hybrid team, what we discovered wasn’t that we needed more capable models, but that the bottleneck is ourselves; in other words, governance.

There can even be a curious paradox: a less capable model can prove more effective, simply because it’s easier to integrate and to govern.

The real challenge, then, isn’t technological. It’s about integration. And we need to understand the implications, because the gap lies in going from having an AI assistant in your work to living in a hybrid team of agents and people, where all of them are independent entities with their own roles and functions. This implies deep changes, which take time. To begin with, the skills required of workers change, and so do all of a company’s processes.

Mind you, the transformation needed isn’t only about production; it’s also social—in consumer habits—and regulatory. In fact, I maintain that sooner rather than later agents will have to be identified with some kind of ID or equivalent protocol linking them to the person, natural or legal, responsible for them.

But that’s a subject for another article.

In any case, if the challenge is what it is, the market should start looking toward companies devoted to integration, and correct the excess of expectations placed solely on those who develop the models. That, and not only the source of the technology, is where value will materialize.

For those of us without committed positions in the market, all this leaves a more practical lesson: avoid creating a strict operational dependence on AI.

Dependence, as I see it, can come in two forms. One, from growing comfortable relying on AI to do most of an office worker’s daily tasks. The second, in a better position but no less risky, from redefining processes in which AI is critical.

Because the thing is, if the market corrects, prices could rise sharply, and the quality, speed or availability of the models could change overnight.

It’s wise to have a contingency plan.

At Kocoro, this is how we’re setting it up: playbooks and workflows defined through hybrid processes, where we humans handle certain points and AI handles others. These AI “touch-points” are documented and detailed, so that local or less capable (cheaper) models can also enter the equation, since a process is a sequence of tasks, and not all of them require a particularly intelligent AI.

In any case, the situation isn’t the same as in 2007. But it’s a call to keep our eyes wide open.

For now, the music is still playing.