Le stesse promesse, gli stessi rischi. E lo stesso errore di fondo: confondere lo strumento con la soluzione.
Five American companies (Amazon, Alphabet, Meta, Microsoft and Oracle) will spend between 660 and 690 billion dollars on AI infrastructure in 2026. An increase of 69% on the previous year. Goldman Sachs estimates that annual AI Capex will reach 765 billion this year and 1.6 trillion by 2031. Last year, AI Capex had already accounted for more than a third of all US GDP growth in the first half of 2025.
And yet, over the same period, a report by the MIT Media Lab found that, despite 30-40 billion in enterprise investment in generative AI, 95% of organisations reported zero measurable return.
700 billion in spending.
95% with zero return.
The gap between these two figures is crucial, and it is the reason so many are already talking about an AI bubble.
Let’s work out why together, by analysing what is working in the AI market, and what is not.
A runaway train
The American railways of the 19th century connected continents, moved goods and built cities. Their economic impact was fundamental. Yet between 1870 and 1900 the railway companies burned through enormous amounts of investor capital. More was built than the market could absorb. Share prices soared far beyond their real valuation at the time. And so, when the correction came, it wiped out entire fortunes.
The technology was right. The only thing wrong was the timing of the investment, which was too far ahead of its time. And as a consequence, the narrative justifying the investment was built on enthusiasm for what the railway would do, not on evidence of what it was already doing.
The structure of the current AI cycle presents the same problems. Valuations are running ahead of revenues. OpenAI is worth 730 billion dollars and projects cumulative losses of 140 billion through 2029. Nvidia has reached a market capitalisation of 4.5 trillion. The top five companies in the S&P 500 now account for 30% of the entire index: the highest concentration in fifty years. Between 15 and 25% of the total value of the S&P 500 today is attributable to expectations about AI. Should those expectations prove wrong (at least as regards timing), those index points would evaporate, and the entire US market would collapse.
A distorted calculation
Everyone talks about the dot-com bubble, comparing it to the situation with AI. Yet one variable makes this bubble different from all the previous ones: geopolitical pressure. The Trump administration has framed AI as a race between nations, a field of competition with China in which falling behind would represent (as with the race to the Moon against the USSR) a strategic defeat. This is precisely why the Stargate project has committed 500 billion dollars to American AI infrastructure.
When investment becomes a matter of national security, the calculation of economic return changes in nature. Governments do not optimise for quarterly ROI. Companies operating in this climate feel implicit pressure to invest regardless of evidence of returns, because not investing signals strategic weakness. This amplifies the distortion: capex grows not only on economic expectations but out of fear of being shut out of a perceived competitive advantage.
The result is an investment cycle partly disconnected from market logic. And cycles disconnected from market logic always correct themselves, sooner or later, through market mechanisms.
The real cost of AI
In April 2026, Bryan Catanzaro, vice president at Nvidia, publicly stated what many had realised long ago: “For my team, the cost of compute is far beyond the costs of the employees”. A 2024 MIT study found that AI automation is economically viable in only 23% of jobs: for the remaining 77%, human workers cost less.
Uber’s CTO, Praveen Neppalli Naga, said he had burned through the entire 2026 AI budget in four months. AI software costs rose by between 20 and 37% in a single year. Gartner forecasts that, even though inference costs will fall by 90% by 2030, agentic models will consume so many more tokens per task that enterprise costs will keep rising.
The paradox is that the very company that makes the chips powering AI admits that, for its own team, AI costs more than employees. Yet that sector’s public narrative continues to describe AI as the most efficient path to productivity.
Perhaps we are simply using the right tool in the wrong way.
AI is not for everything
AI is not the right answer to every question.
How to make a good carbonara. What the word «stochastic» means. What the capital of Chad is. Billions of queries like these are processed every day by systems that consume as much energy as entire cities; that require infrastructure worth hundreds of billions; that are designed to solve problems of far greater complexity. In short, we are using a trillion-dollar calculator to work out 2+2. For centuries, a book has done the same job better, faster and with zero energy consumption.
For any question whose answer is stable, verified and written down, dictionaries, encyclopaedias and Wikipedia are more accurate, quicker to consult and cheaper in infrastructure terms.
This tendency to replace them with a generative system is a symptom of habit. And an expensive habit at that.
AI is an extraordinary tool for problems that require synthesis, reasoning on complex data, processing of non-linear patterns, and decisions that cut across different domains. For every other problem, the solution already exists. And it is often printed on paper.
The difference between those who generate returns and those who do not
The question a CEO should be asking is not “are we using AI?” but: “for which decisions does AI produce a result that no other tool produces better?”
The difference between the 5% of organisations generating measurable returns from Artificial Intelligence and the 95% that are not comes down to method.
A corporate AI governance system answers three questions before any implementation. Which specific problem does this tool solve better than any existing alternative? How is the result measured in verifiable economic terms? What do we stop doing, or do differently, when the tool works as intended? Without these answers, investment in AI is spending without direction. And spending without direction, multiplied across an entire industry, produces the figures we are now seeing.
AI is not a bubble: it is simply a tool that is still being used in an immature way. Yet the same could be said of the American railways in 1873, when the market corrected by 30% in six months.
So caution is needed. Because the technology survives the correction, but the capital burned does not.
New Connections (FAQ)
Is AI just a bubble?
Every major technology (such as the railways, electricity or the internet) has gone through a phase of overcapitalisation before generating widespread returns. The question is not whether the technology is real, but whether current valuations already price in twenty years of adoption that has not yet happened. The gap between the capex invested and the returns measurable today suggests that at least part of that value is anticipation, not reality. Anticipations get corrected.
Should a mid-sized Italian company slow down its AI investments?
It should stop investing in AI in response to the enthusiasm of the moment and start investing in AI in response to specific problems. The distinction is practical: identify two or three business processes where AI produces output that is verifiably better than any alternative, measure the result in economic terms, and scale only what withstands that measurement. Everything else, including queries that could be answered by a manual, an Excel spreadsheet or a phone call, does not require AI. It requires discipline.
Is the fact that AI costs more than employees a temporary or a structural problem?
Partly temporary: inference costs will fall significantly over time. Partly structural: the most capable agentic models consume exponentially more tokens per task than simple models, which could offset the reduction in unit costs with higher consumption per query. For this reason, Gartner expects enterprise AI costs to keep rising even as infrastructure costs fall. The break-even point between AI cost and human cost will depend on the specific function, not on a general law. Those who govern this distinction function by function, rather than applying it uniformly, will have a structural advantage.