David Carter, industrials senior analyst at RSM US, describes the major supply chain constraints that are bedeviling major U.S. data centers in their efforts to move ahead with innovations in artificial intelligence.

In the last couple of quarters, U.S. data center developers known as “hyperscalers” have begun mentioning in their earning calls problems with acquiring enough microchips and land to drive their continuing efforts at AI innovation, Carter says.

At the heart of the issue is a need for heavy capital investment for the construction of chip manufacturing plants in the U.S. The CHIPS and Science Act was passed in 2022 for that purpose, but it’s a long-term process that won’t see meaningful change for years. “There’s a long, long runway to actually building that out,” Carter says. In the meantime, data centers are finding their supply chains and supporting infrastructure to be increasingly inadequate for the task.

The most complex form of semiconductor manufacturing requires three distinct stages: design in the U.S., production in Taiwan, and access to highly specialized equipment made in the Netherlands. Carter says the U.S. currently has around 10% of the world’s chipmaking capacity, with the goal of growing that to 30% by the end of the decade. But given the constraints that domestic industry faces today, “that’s a tall order.”

Land is also a major issue. Chipmaking plants need to located in areas with access to adequate power, water and labor. Electricity is becoming an especially serious problem. Between 1950 and 2007, Carter says, the U.S. grew its need for power by around 4.7% per year. Growth over the following 15 years was roughly flat, but today it stands at between 2% and 3% annually — a massive increase, given the requirements for power by chip manufacturers, AI, cryptocurrency miners and data centers. “We’ve lost that muscle,” Carter says.