Only 35 Percent of the US Data Center Pipeline Is Credible, Bernstein Says, as 47 Projects Stall in 2026

The American AI buildout is no longer a story about land, chips or capital alone. It is increasingly a story about whether a project can be built at all. The Financial Times reported on October 18 that at least 47 large US data center projects have been delayed, withdrawn or blocked in 2026, according to regulatory intelligence firm RegLayer, a figure the firm describes as a conservative floor because it covers only large developments.

The same reporting carried a second, sharper number. Bernstein estimated this month that only about 35 percent of the planned US data center pipeline is credible once grid-connection timelines, transformer shortages and liquid-cooling bottlenecks are priced in. In other words, the gigawatts announced on earnings calls and in economic development press releases are roughly three times the capacity that investment analysts expect to actually arrive.

The friction is local as much as financial. FT's investigation found more than 1,000 US communities have attempted moratoriums or other restrictions on AI infrastructure during 2026. Industry executives told the paper that some utility regulators are refusing to approve interconnection for projects that cannot supply their own power, which pushes developers toward on-site generation and longer schedules.

The money at stake is large enough to matter at the scale of the national economy. A Brookings paper by Columbia Business School economist Stijn van Nieuwerburgh estimates cumulative US AI infrastructure investment of about 10.3 trillion dollars between 2025 and 2032, an average of 3.63 percent of US GDP each year. That share is higher than the railway, highway and electrification waves that previous booms were measured against. Private construction spending on data centers alone has climbed past 5 billion dollars a month in 2026, a figure that excludes GPUs and servers and is adjusted for inflation.

Credit markets have started to price the gap. FT reported that bonds issued by AI infrastructure companies were sold off and that CoreWeave shares fell nearly 25 percent in a week, while the New York Times reported yields on CoreWeave debt above 15 percent and some Oracle-related paper trading at 80 to 90 cents on the dollar. Insurance against SoftBank exposure has also become more expensive. For operators that underwrite new campuses with long-dated leases, a higher cost of capital compounds every schedule slip.

The distributional effect is uneven. Companies that already own energized shell capacity, or that can self-generate, gain pricing power as the credible pipeline shrinks. Developers that depend on a single anchor tenant to make a project bankable carry the most schedule risk, and their lenders are the first to reprice. Chip suppliers face a timing question rather than a demand question: the orders may exist, but the buildings that would host them are slipping into later quarters.

What to watch next is less about announcements and more about interconnection queues and equipment lead times. If the 35 percent figure holds, the capacity that actually energizes by 2028 will undershoot the public pipeline by a wide margin, and the price of delivered AI compute will reflect scarcity rather than the cost of the silicon inside it.