AI bubble, is 2026 the year it bursts? The issue on energy, data centers and “sovereignty of minds”

* President of the Copernican Association

Will 2026 be remembered as the year the AI ​​bubble burst? 2000 is remembered as the year of the “dotcom” explosion, the internet companies whose prices had reached stellar levels that did not take into account their economic fundamentals: Ratings were based on “eyeballs,” a slang way of defining users. We will then see how to monetize them. But in 2000 there was not yet a widespread understanding that technology would be monetized with advertising and that to win the loyalty of the public to be sold to advertisers, it was necessary to exploit the network effect (users are where everyone else is – who would go to a social desert?) and lockin (like the lobster pot: easy to enter but impossible to exit – who can abandon Gmail with the perceived risk of losing all the accumulated mail?).

The push of the survivors

After the bubble bursts, the internet companies that have best interpreted and understood their role as a new technological medium have become prosperous and their prices have pushed the NASDAQ to much higher valuations than those of 2000. Today the most popular browsers are Chrome and Safari, thanks also to the push given by the integration with Android and the entire Apple world, but in the bubble years, between 1995 and 1999, the dominant browser was Netscapea company founded by twenty-three-year-old Marc Andreessen, today a highly successful venture capitalist and powerful influencer of the White House’s technology policy. Netscape allowed free browser downloads and its strategy was to charge for the web server, the software used to serve web pages. When we started distributing Netscape with I.NET, the first Italian business ISP, their web server cost a million dollars. Needless to say, we didn’t even sell one because the Apache web server was distributed free of charge in Open Source mode. Both the browser and the web server, both Netscape and Apache were based on developments that originated at CERN in Geneva, where Tim Berners Lee invented the web for a research project financed by Paolo Palazzi from Trieste.

The bubble was the TLCs

In reality, the dotcom bubble was not so much caused by internet companies but by telecommunications companies which, dragged by the explosion of internet use, invested a quantity of resources never seen before in optical fiber and related devices, in the hope that one day this investment would pay off. It was companies such as Corning (fiber), Lucent and Cisco (network devices) that dominated stock market growth and suffered an almost vertical collapse starting in April 2000. The NASDAQ didn’t actually collapse but suffered a decline that lasted about 2 and a half years before starting to grow again. It then took around 14 years to return to 2000 levels. It is also said that the bursting of the bubble bequeathed us a significant infrastructure of fiber networks, just as the bursting of the railway bubble in the UK in 1847 (after two years of frenetic investment in the face of unachievable expectations) left us a track infrastructure which was subsequently fully utilised. In reality this is not the case: the performance of the fibers depends to a large extent on the devices at the ends. The evolution of electronics has made it possible to squeeze more and more bandwidth from existing fibers, so the growth in demand has not led to the use of all the fiber laid in the 2000s.

Between 1999 and 2001, the growth and collapse of the market were marked by accounting irregularities of some of the giants of the time such as Enron, Qwest and Worldcom, the discovery of which contributed to the collapse of the stock markets. These companies purchased portions of fiber from each other: Enron purchased from Qwest in some areas and Qwest similarly purchased from Enron in others: each recorded an increase in revenue for what it sold and put the costs of what it acquired into twenty-year amortizations, thus generating huge apparent capital gains. A congressional commission wanted to understand whether these circular investments were determined by actual market demand or were simply hidden exchanges aimed at inflating balance sheets. The rest is history: subsequent investigations by the competent US agencies led to numerous arrests and none of the companies involved no longer exists, nor does Arthur Andersen LLP, the global giant of auditing firms which certified the financial statements without reporting these frauds.

Circular operations and the differences with Dotcoms

An internal memo written by Nvidia, the AI ​​star, entitled “We are not Enron” recently reached the press in response to accusations made by Michael Burry, the manager who bet against the market during the subprime mortgage bubble (made popular by the film “The big short”) and which today bets against the AI ​​bubble. Although many analysts have criticized it the circular affairs of AI starsNvidia does not use corporate vehicles to inflate revenues and hide debts; although it invests in companies that then buy its hardware, it always does so together with institutional financial operators. Financial reports show that Nvidia’s top six customers account for more than 80% of its revenue; the fact that almost all of them have extended the amortization of the hardware purchased is a problem that could affect Nvidia only as a source of a prospective decline in demand, not as an accounting trick as Worldcom did. Nvidia may have declining sales, but there’s no sign of 2000s-style wrongdoing.

Some concerns have also been expressed by analysts regarding the data center boom. More than one has pointed out that, even if the demand were such as to justify the installed computing power, there is not enough energy in the US to power them: many data centers cannot be turned on. It’s not obvious that these buildings full of Nvidia chips will remain as a legacy of an eventual burst of the bubble: how many years can they be competitive? If the answer was “for a long time”, questions should be asked about the sales possibilities of the next generations of Nvidia chips. Nvidia’s growth potential contrasts with the useful life of data centers and therefore with their profitability.

In the end, we ask ourselves, all the circular investments that are taking place openly are in response to a market need? That the end-user market does not appear to be brilliant seems to be demonstrated by Deutsche Bank which, by analyzing credit card transactions, found stagnation in the growth of OpenAI in Europe. Microsoft, more recently, has had to lower its sales targets. The financial director of OpenAI said that she believed “some ‘parachutes’ from the US government for infrastructure investments were necessary”, shaking the markets which interpreted these words as a sign that revenue growth does not seem to be aimed at exceeding costs. (He also said he was talking about these things with the White House.) The CEO, Sam Altman, had to intervene to specify that OpenAI was not asking for a guarantee or subsidy for its investments but that the discussion concerned the sector as a whole: “in order to win the AI ​​race against China” it is necessary to lower the cost of capital to encourage investments and in this the government can intervene in order to broaden demand.

The US administration didn’t have to wait and ask has launched three important supportsafter all, as Trump says “50-60% of the US economy is now driven by AI”. The first is the “Genesis Mission,” a public initiative led by the Department of Energy and compared in relevance to the Manhattan Project – the program that created the atomic bombs. The second was a decree – of dubious legality – which prohibits states from regulating AI. The third was the lifting of the ban on chip sales to China, thus significantly expanding Nvidia’s market. Michael Burry may have viewed the federal support and market widening with concern: his prospect of the bubble bursting has become much longer.

Meanwhile, the Chinese Deepseek continues to produce models with minimal investments compared to the US giants and makes them Open Source. Are OpenAI and its partners facing a repeat of what happened to Netscape in 2000? Many startups in Silicon Valley and beyond are using these Chinese models for free, a fact that is reminiscent of what happened with web servers in 2000. Thanks to the fact that there are free open source ones, today web servers are everywhere: from thermostats to light bulbs. It is not certain that whoever leads a market (then Netscape) is the person who will then be able to monetize it. And it is not certain that SLMs (Small Language Models) will not prevail in diffusion compared to LLMs, probably hosted directly on users’ computers or on company servers, as predicted by a recent article by the IEEE (Institute of Electrical and Electronics Engineers – international association of scientists with the aim of promoting technological sciences). Article that should be read with a bit of discomfort by those who have invested in data centers.

If diffusion in the B2B market could be oriented towards custom linguistic models, distilled from Open Source models, thus also mitigating privacy and security problems that plague larger models, we must ask ourselves what will happen in the B2C market.

The cost of querying LLMs (called inference) will reduce, thanks to the evolution of specific chips and algorithms. Could this reduction lead to a cost level compatible with a remuneration from advertising, making the Internet business model persist as it emerged after the 2000s?

The excessive power of the Llm

Will LLMs become a medium for all Internet users? When? Before that happens, we should start worrying seriously about the power and responsibilities of those who manage them. The power of manipulation by chatbots is much greater than that of advertisements. Unfortunately, the news increasingly reports news of cases of serious conditioning of people by chatbots, conditioning which occurs in a much less visible and more surgical way in the secrecy of personal chats.

Upon closer inspection, this is also a question of sovereignty: sovereignty of minds.