AI Infrastructure Cracks Show as 47 Data Center Projects Stall

The Crack Is Real, and It’s Not Where You’d Expect

Forty-seven. That is not a projection. It is the count of large-scale data center projects in the United States that have been delayed, withdrawn, or blocked at the permit stage this year alone, according to regulatory data reported by the Financial Times. Bernstein estimates only about 35 percent of all planned US data center projects will actually materialize. Over a thousand American communities are actively pushing moratoriums or restrictions on AI infrastructure construction.

The headline numbers sound moderate against a $10 trillion projected spend. The pattern underneath them is not. What is breaking is not demand for AI compute. Demand is insatiable. What is breaking is the financial and regulatory architecture built to deliver it.

Four Layers of Fracture

Regulatory friction is the visible layer. Communities are not abstract opponents. They are local governments wielding zoning power, environmental review, and public pressure against facilities that consume enormous amounts of water and electricity. A thousand municipalities moving toward restrictions is not a protest movement. It is a structural headwind hitting the timeline that every AI company claims is critical. In Texas and Virginia, two of the country’s biggest data center corridors, county commissions have begun demanding impact fees and grid-interconnection guarantees that were never part of the original business case. The cost per megawatt is quietly climbing.

Capital is the second layer. CoreWeave shares fell 23.6 percent last week. One of its corporate bonds is trading in distressed territory. The company leased data center space while betting it could fill it quickly enough to service the debt. When occupancy slows, the leverage flips against it. Nebius, Irène, and Enscale face a different problem. They own the physical assets. Their balance sheets carry the construction risk directly, and any stumble there flows straight into Nvidia’s supply chain and revenue forecast. Lenders are already tightening covenants on project finance deals that once closed in weeks. Due diligence periods have doubled.

Operational inexperience is the third layer. Companies like OpenAI, Oracle, and SoftBank assembled the Stargate project with unprecedented speed and capital. Now multiple Stargate sites are delayed. Anthropic is increasingly likely to terminate a 450 billion dollar neocloud contract with Enscale. The rush to secure GPU capacity forced AI firms to partner with inexperienced operators. Speed was the strategy. Execution risk was the tax. Three separate projects in Georgia and Tennessee have already required engineering reassessment after initial build plans proved incompatible with local grid constraints—something no one flagged during the fundraising phase.

Off-balance-sheet debt is the fourth, quietest layer. Sjoerd van Nieuwerburgh at Columbia Business School calculated that US AI infrastructure spending will average 3.6 percent of GDP from 2025 through 2032. That is 1.5 times the peak of the 19th century railroad boom and three times the internet boom. His warning is specific: almost all of that debt sits outside company balance sheets. If cash flows drop even 10 percent, equity values can evaporate. The structure is leveraged enough that a modest revenue miss becomes a solvency event.

Second-Order Effects Are Already Appearing

The ripple effects extend well beyond the stalled projects. Utility companies are recalibrating their grid-connection queues, which pushes back timelines for every applicant in line. A data center that once expected a two-year interconnection wait now faces four. That delay cascades into every downstream contract tied to those facilities.

Insurers are beginning to treat AI data center projects as higher risk, adjusting premiums and reinsuring less of the exposure. That means sponsors who relied on insurance-backed project finance to close deals are finding the terms materially worse. Some lenders are requiring equity partners to post additional collateral before releasing construction draws.

Local labor markets are tightening faster than expected. Data center operators competing for the same electrical engineers and mechanical contractors have driven wage inflation in key markets. The people building these facilities are in shorter supply than anyone anticipated, and the bottleneck is not bricks and mortar—it is skilled labor.

Who Wins and Who Loses

The winners so far are clear: Nvidia continues to sell chips because the orders were placed before the cracks appeared. The backlog sustains revenue even as individual projects stall. But the next wave of orders depends on whether the projects themselves survive long enough to need them.

The losers will emerge in layers. First go the neocloud operators who bet heavily on leasing and rapid occupancy growth. CoreWeave is one example. Then come the asset owners carrying construction debt, because lenders do not forgive missed deadlines. Finally, the AI applications companies that signed contracts assuming capacity would be ready, only to watch those commitments unravel or price themselves out.

Consolidation is coming. The Information reports that the hundreds of neocloud providers active today will likely compress to seven or ten within a few years. David Driggers of Sierra Wireless noted that asset managers, telecom companies, or fiber operators are the most likely acquirers. These buyers have cheap capital and existing infrastructure relationships. They will buy distressed capacity at discounts and resell compute slots at margins that surviving builders cannot match.

The Reckoning Is a Selection Mechanism

This is not the end of the AI infrastructure buildout. The demand signal is real and growing. What is ending is the assumption that capital available today will remain available tomorrow at the same terms, or that projects promised today will receive permits next year.

The market is running a brutal selection process. Companies that financed aggressively through off-balance-sheet structures will face margin calls or dilution. Companies that rushed partnerships with inexperienced builders will face cost overruns and delays. Companies that ignored local opposition will face litigation and legislative blockage. The survivors will be those with secured permits, conservative leverage, and established relationships with regulators and utilities.

For the broader tech sector, the implication is a pause in the most aggressive parts of the infrastructure cycle, followed by a reshaping. Valuations in the compute and hosting space will correct. Some high-profile projects will die entirely. Others will restart at lower capacity and higher cost per unit. The 10 trillion dollar number from 2025 to 2032 still stands, but the path to it is no longer a straight line. It is a series of bottlenecks, bankruptcies, and bailouts that will determine which companies shape the next decade of AI computing.

The boom is not over. The free money era is.