Are we really witnessing the onset of an AI crash?

French debt is not the only one to be attacked on the markets: that contracted by the AI ​​giants also arouses mistrust among investors. The value of bonds issued by large American tech companies to finance their huge spending on the construction of data centers has recently declined.

Neoclouds, this new generation of providers tailored for the AI ​​era, are also in difficulty. Bonds issued by CoreWeave, one of the most famous players in this market, have fallen and its stock market value has fallen by a third since the beginning of October, as has that of its rival, Nebius.

Another sign of turbulence in the AI ​​market: OpenAI pushed back its IPO to 2027 in September, with Sam Altman citing concerns about the risks posed by the technology. As for Firmus, an Australian data center operator financed by Nvidia, it simply canceled an IPO which was expected to be the biggest of a generation in Australia, in the face of market reluctance.

At the origin of the turmoil currently shaking the AI ​​market are the unprecedented investments that the American giants of this technology have deployed in the construction of data centers to remedy the shortage of computing power, as evidenced in particular by the price that must be paid to rent a graphics processor. A B200 chip from Nvidia currently rents for more than six dollars an hour, double that of six months ago, according to SemiAnalysis, a research firm specializing in semiconductors. Alphabet, Amazon, Meta, Microsoft and Oracle plan to spend $800 billion on infrastructure this year, double last year.

To finance these expenses, these companies can no longer rely on their cash reserves, although considerable, and must resort to debt, which has led them to issue an unprecedented quantity of bonds. However, the markets have doubts about the profitability of these investments: they fear in particular that the increase in turnover enabled by AI will not be rapid enough to justify them. On October 22, Meta, whose investments exceeded the money generated by its activity in the third quarter, fell by 7% on the stock market, the day when the Nasdaq experienced its worst session since June.

According to a paper presented at the end of September by economist Stijn Van Nieuwerburgh, who is currently a reference, for all data center owners to reach the break-even point for their investment in AI, it would need to generate $3,725 billion in annual revenue by 2032, or approximately 9.2% of the US GDP projected for 2032. This corresponds to the overall turnover from the sale of power of calculation by operators, whoever they may be, of the 182.7 GW of American capacity completed by 2032.

"Today, U.S. computing power revenues are on the order of $300 billion. To reach $4 trillion by 2032, total data center revenue must grow about 45 percent annually for seven years," Stanford economist Hanno Lustig writes on his blog.

Machine-translated from French