Plans to build artificial intelligence (AI) data centers are increasing in the United States, but only some have led to actual groundbreakings. With power and construction materials short, a lack of skilled workers, and even permitting delays, the pace of data center expansion is failing to keep up with initial plans.
On the 7th, CNN, citing data center research firm Aterio, reported that AI corporations have released plans for 3,969 new data centers in the United States, but only 802, or 20%, are currently under construction.
JPMorgan said about 60% of data center capacity slated for completion in 2027 has not yet broken ground, and 7% of sites that have already started construction have also seen schedules delayed. Goldman Sachs projected that only about half of the AI compute capacity scheduled to go online from now through 2028 will meet the target timing.
The first reason for construction delays is a shortage of materials and equipment. As demand for data center construction has surged, securing building materials has become difficult, and there is also a shortage of advanced AI Semiconductor chips to be installed inside. In particular, production of key advanced AI Semiconductor chips such as Nvidia's Blackwell and AMD's MI300X is concentrated at Taiwan's TSMC, increasing supply chain dependence.
Securing power is also a problem. Data centers already account for about 8% of U.S. electricity use, and that share could rise to 12% by 2028. While AI corporations have moved to build their own power plants, related equipment is also in short supply. According to JPMorgan, the delivery lead time for step-up transformers for power plants has recently tripled.
There is also a shortage of skilled workers. American Edge Project estimated that to meet current construction schedules, an additional 500,000 electricians, 300,000 welders, and 550,000 plumbers are needed.
Joe Masek, head of U.S. real estate digital infrastructure at Mace Risk, said, "Some clients are pushing projects around the clock, and contractors are busy all day. But if the required workforce is all tied up in existing projects, there is no way to move construction further along."
On top of that, local opposition is significant. In a recent Gallup poll, 71% of Americans opposed data center construction, and about a dozen states pushed for moratoriums on building data centers. Goldman Sachs analyzed that permitting approvals are a bigger hurdle to actual project execution than outright construction bans.
Some also say many of the announced plans had low feasibility from the start. That is because developers often submit applications in multiple regions simultaneously and then choose only the site with the highest business viability.
Stijn Van Nieuwerburgh, a professor at Columbia Business School, projected that of the 565 gigawatts (GW) of compute power planned by AI corporations, only 180 GW will actually be built over the next 10 years.
Professor Van Nieuwerburgh said, "It is very difficult to time large-scale construction projects precisely," and added, "If market expectations grow excessively, excessive liability can pile up, and in the end, many investments may end in failure."