After having greatly contributed to feeding it, Sam Altman and Jeff Bezos believe that euphoria around artificial intelligence created a bubble. “It is a kind of industrial bubble. Investors are struggling, in this madness, to distinguish between good and bad ideas,” said the chief of Amazon, this Friday, October 3, during a conference in Italy, while insisted that technology is “real and will change all sectors.”
Sam Altman, the creator of Chatgpt, had preceded him, in mid -August, sliding that “when a bubble is formed, intelligent people are carried by a context of truth.” “Someone will lose a colossal sum. We do not know who, and many will get colossal sums,” added the head of Operai, hoping to be in the second category, while his company reached this week an assessment of $ 500 billion.
To listen to these Silicon Valley magnates, the prices of certain assets in AI, therefore, would exceed their intrinsic value. Since each one is careful to specify that they are not worried, it cannot be excluded that these exits are above all a way of discrediting their rivals, in a context of intense competition.
But the fears that a colossal bubble has been formed has been reinforced in recent weeks. It has become obvious that American growth is now driven by the colossal investments of the AI giants, more than $ 250 billion this year, in its infrastructure (“Stargate” for Openai, “Prometheus” of Meta …). Some wise observers are worried. In this case, the big question is whether the IA mastodons will generate income corresponding to their valuations and investments.
The spectrum of the Internet bubble
Torsten Sløk, the chief economist of Apollo Global, a giant from the American risk capital, believes that history is stuttering, 25 years after the breakdown of the Internet bubble. Sløk recently wrote that the ten largest companies in the American Stock Exchange were even more overvalued than in the 1990s, noting that “the Tesla course/benefit ratio is close to 200, and Nvidia’s is around 60”. Other specialists are more optimistic, such as businessman Azheem Azhar, who responds that “the current development of AI seems more solid” than the Internet bubble.
Until now, investors have reassured the increase in “Seven Magnificent” profits, which will soon join other companies such as Oracle or Palantant. In the same way, madness around Openai, which is not yet in the stock market, continues to be fed by the dissemination of its technology. ChatgPT hopes to spend one billion weekly users by the end of the year and has announced that a billing network increased to $ 12 billion this year.
“A deficit of 1,500 billion”
At the moment, technological magnates have reinforced their expenses to build the colossal infrastructure that they consider necessary for the development of their models. At this point, they are not afraid to invest too much, but not do enough. This year more than $ 350 billion must be injected into AI factories, including 70% in the United States.
By 2029, the accumulated expenses in the data centers are expected to increase to $ 2,800 billion, according to the latest analysis of the CITI Investment Bank. In the same orders of size, Morgan Stanley Bank estimates that the expense of the data center will reach 3,000 billion in 2029.
On this horizon, income is probable. At the end of August, Citi analysts predicted that they would reach $ 780 billion in 2030. This would represent a spectacular leap compared to the $ 43 billion generated today, but it would always be less than the first expenses. If the financing needs are growing, some analysts believe that the Hyperscalers – Microsoft, Nvidia, Meta or Alphabet – They probably cannot continue financing most investments.
“Currently, stability is based on solid cash flows Hyperscalers. If they stop covering most of the investment spending, debt and security will accumulate. This is not good news, “says Azheem Azhar.
The unknown productivity profits
Therefore, the decisive factor will be the growth of income. In the medium term, its increase assumes that AI provides significant productivity profits, which would justify that companies pay a lot to access. Problem, some recent works have been largely put into perspective, at this stage, the promises of mixture productivity profits, particularly formulated by Sam Altman. A study, conducted by the METR specialized Institute, even showed that developers who use AI tools were losing productivity.
Nothing more logical to read Erik Brynjolfsson, economist at Stanford University, interviewed by the Washington Post. His research has shown that new technologies are experiencing a “product in productivity. In a first phase, companies are struggling to deploy it, which leads to a temporary fall in productivity, before eventually significantly increases. This flight also implies that AI models continue to progress.
“The true bottleneck”
More obstacles from Earth to Earth could also be on the road to AI giants, which seriously complicates the search for profitability. “The true bottleneck can be electricity consumption, gigantism attracts me,” observes Jean-Edwin Rhea, Fund Manager of Sunny Asset Managment, interviewed by BFM Business.
According to Citi analysts, “the global calculation demand for AI will require an additional energy capacity of more than 55 gigawatts by 2030”, a little less than the capabilities of the French nuclear park (63 GW), which, in the United States, could generate an additional expense of $ 1,400 billion in the calculation of AI.
Unless there are strong tensions on the network and unbridled prices, the supply of AI factories requires the rapid construction of new electrical capacities. Due to the duration of nuclear construction times, the solution is probably found in gas power plants. But there too, a gas turbine takes at least three years, according to Bloomberg. The construction of data centers in the United States could also decrease due to lack of workforce, while the US manufacturing sector is likely to lack 1.9 million workers by 2033.
America runs the risk of losing its engine
Therefore, the risk of losing the shirt in AI’s career is numerous. Basically it is the lot of all technological revolutions, synonymous with overall bubbles and stock market, according to Carlota Pérez’s work, reference in the matter. The latter explains, however, that these bubbles are generally useful. In the mid -nineteenth century, the excessive optimism of certain investors for the railroad certainly caused their loss, but the railroads remained and then served.
But it is not easy for this to be the case with the gigantic data centers, mainly compounds of GPU chips, used to train models. This age “in the canine years, its useful useful life for advanced applications as model training is about three years,” says Azheem Azhar. If the income does not follow and that a bubble explodes, the latter believes that it could only remain “warehouses full of obsolete chips of GPU”.
In addition, the potential cost of implosion of the stock market has “increased alarming”, alerts the economist. Without investments in AI factories, “the United States would be close to the recession, or even in the recession, this year,” according to a recent note from Deutsche Bank. Factories construction to generate AI models constitutes “a mass recovery program in the private sector,” according to economist Paul Kedrosky. If a bubble explodes, the engine of the first world economy could take advantage of.
Source: BFM TV
