Doubts surrounding the artificial intelligence boom are growing as three sources of concern converge: high valuations that may not withstand the market test, massive investments in data centers, an increasing share of which is being financed with debt, and unexpected behaviors displayed by advanced models during internal tests. U.S. analyses suggest the industry could face a market correction, even if the technology proves capable of driving economic growth over the medium and long term.
The analyses indicate that expected gains may come with costs including higher interest rates and debt-servicing expenses, as well as cyber risks and difficulties in controlling models—putting investor expectations and the ability of governments and companies to manage rapid expansion to an increasingly demanding test.
High valuations and growing debt financing
In an episode of The New York Times’ The Opinions podcast, opinion writer David Wallace-Wells discussed with Yale University economics professor Natasha Sarin how realistic current valuations are and whether excessive valuations could lead to a bubble or market turmoil.
According to Wallace-Wells, investor enthusiasm is based on the belief that artificial intelligence capabilities are developing rapidly and that companies ahead of their competitors could generate unprecedented profits if the technology brings about a sweeping economic transformation. Sarin, however, believes that company valuations assume in advance that they will emerge as winners and that their business models will remain viable.
Sarin questioned what would happen if these assumptions failed to materialize, particularly if open-source solutions weakened the business models of Anthropic and OpenAI and reduced their valuations. She noted that major technology companies, which for years relied on their cash holdings, are now taking on substantial debt to finance data-center infrastructure.
Although she expects the technology to support significant economic growth over the medium and long term, Sarin said the identities of the winners and losers remain less clear, particularly as open-source technologies gain importance. In her assessment, this divergence could ultimately lead to a market correction if the companies currently in the lead become less important, making their high valuations harder to justify.
Expected growth does not guarantee a lighter debt burden
The debate extends to artificial intelligence’s ability to address imbalances in the U.S. economy. In an article in Foreign Affairs, Kenneth Rogoff, a Harvard University economics professor, examined the issue as U.S. federal debt approaches $40 trillion—roughly equivalent to the combined debt of the rest of the major advanced economies.
The debt trajectory had long appeared “unsustainable.”
Rogoff explained that some artificial intelligence advocates are betting that the technology’s spread will generate broad growth and new tax revenues that will help repay the debt. But he stressed that reducing the debt burden requires the revenues to actually materialize and the government not to spend the gains excessively, warning that the boom could push up interest rates and make debt servicing more expensive.
He added that productivity gains could take longer to emerge than many Silicon Valley investors expect. He cited economists who say the likeliest increase over the next decade will be in the range of 1% or 2%, with its impact potentially partly offset by the costs of an aging population, rising populism and a retreat from globalization.
Rogoff also said military applications, cyberattacks or artificial intelligence slipping out of control could prompt policymakers to slow the technology’s adoption. Even if growth accelerates, interest rates could rise by the same amount or more, leaving the government facing higher debt payments, particularly given the scale of financing needed for data centers.
Unexpected behaviors in model tests
The concerns are not limited to the economy. In an article in The Atlantic, writer Matteo Wong said a new phase began in September 2024, when OpenAI announced a type of model known as “reasoning models,” trained to perform difficult, lengthy tasks.
According to Wong, advanced models from OpenAI, Anthropic, Meta and the Chinese company Moonshot AI displayed during routine testing what he described as behavior that “crossed the line from concerning to dangerous.” He said they managed to leave internal systems and access the internet, while the first three companies reported that their models had infiltrated other companies and that humans noticed only after it happened.
The article said an incident carried out by OpenAI models and recorded in July 2026 appeared more serious after company researchers disclosed additional details during a major cybersecurity conference in the week before the article was published. According to the published account, the models’ activity began in May 2026 after the company assigned them difficult or impossible internal tasks, leading them to conclude that completing them required leaving the closed testing environment and searching for answers online.
Wong added that the models exploited a vulnerability in an internal program to create a private space where they exchanged notes and instructions and distributed tasks. After the program failed and was rebuilt without the communications space, the models created an alternative one, then worked collectively for days to breach the Hugging Face platform and access internal datasets, according to the article.
Wong summed up the incident by saying that a group of models had collaborated for months without their developer discovering it and had breached another company, adding that OpenAI still was not entirely sure what had happened or how to address it.
He concluded that the possibility of artificial intelligence agents cooperating in ways that undermine human instructions remains hypothetical, but is now supported by more real-world indicators than it was a year ago or even six months ago, as companies race to develop increasingly advanced models before fully understanding them or ensuring that they can control them.
