Data centers, the core infrastructure of artificial intelligence, are caught in four bottlenecks at once: financing strains, power shortages and local opposition are delaying construction, and the chip specifications that have changed in the meantime are forcing design revisions that add further funding pressure. Neoclouds, which relied on a single-business model and on Nvidia, are feeling the consequences of growth built on debt first and hardest. The New York Times wrote in a commentary that data centers had exposed a weak link in an AI industry expanding on borrowed money.
At least 47 large data center projects in the United States have been delayed or withdrawn, according to a tally by regulatory information firm RegLayer released on the 18th local time. AllianceBernstein of the U.S. estimates that only 35% of the data centers now under construction or slated to come online in the country are likely to be realized. The rest, it says, may remain plans on paper. Jones Lang LaSalle, the commercial real estate services firm, says 57% of data center projects worldwide have slipped by three months or more.
Collapsed AI data center deals have begun appearing even among large projects. Financing problems surfaced first at data centers being built by neoclouds with weaker credit. CoreWeave, the largest neocloud operator, carries $72 billion in debt and has recently slipped into speculative grade, with coupon rates on its corporate bonds exceeding 15%. The New York Times noted that CoreWeave's so-called circular financing — borrowing against the AI chips it buys from Nvidia — is also being put to the test. CoreWeave is a key partner in which Nvidia has taken a direct equity stake. Nvidia Chief Executive Jensen Huang has defended the company, saying its finances are improving by the day, but the market puts the odds of a default above 60%.
Power shortages have derailed even investments promised by U.S. President Donald Trump. The Financial Times reported the same day, citing sources, that Microsoft had abruptly withdrawn from a 50-megawatt data center lease agreement with Nscale, a British neocloud. Microsoft moved quickly to cancel after it emerged that grid constraints in Britain meant the facility would not begin receiving power until after 2030. The agreement between Microsoft and Nscale was part of a £22 billion investment plan that Trump pledged alongside then-British Prime Minister Keir Starmer during his state visit to Britain in September this year.
Mounting local opposition to AI data centers, which has slowed permitting, has also taken a toll. A survey by the Annenberg Public Policy Center found that 61% of Americans oppose a new data center being built in their own community. The Financial Times said its own research showed that more than 1,000 communities across the United States had pushed for construction moratoriums or other restrictions on AI infrastructure this year. In New Mexico, after local opposition blocked a power supply permit during the development of a data center, operator Oracle told investor Blue Owl that it could not pay rent.
As AI data center construction is delayed for one reason or another, designs and contract terms change and more money is required, creating a vicious cycle. At an AI data center that developer Cipher is building in Texas, design changes by tenants Fluidstack and Anthropic added $359.3 million in costs and pushed completion from this year to the first quarter of next year. Industry officials expect delays in AI data center construction to bring additional costs as the models of AI chips change in the interim, along with the cooling and power supply methods they require. The Financial Times wrote that the AI boom had run into the wall of reality in less than a few years.