The US is leading the world of artificial intelligence, but it is not guaranteed to stay there. The bottleneck is not talent, ideas, or capital, but electricity.
Electricity is a binding constraint on the construction and use of hyperscale data centers, essential for training today’s advanced AI models. The Ministry of Energy will project that data center electricity demand will almost triple by 2028. That increased demand requires generation and transmission capabilities at unprecedented speeds. However, some AI developers have a hard time getting the power they need. Recently, the Federal Energy Regulation Commission (FERC) blocked Amazon from connecting data centers directly to the Talen Energy nuclear power plant.
One of the major obstacles to increasing power availability is today’s authorization process. A recent survey of the Brookings facility allowed reforms. Focusing on wind and solar power, the author concluded:
(Energy) projects can be delayed at any stage of the process. Local communities may oppose projects that take over property or interfere with opinions. The state could place a high environmental review burden or block unpopular projects. The interstate grid operator is backlogged in connection requests. It may take years for federal authorities to grant any necessary permits for the project.
Even building traditional natural gas plants requires navigating federal, state, and local approval mazes, including the Environmental Protection Agency, FERC, Local Zoning Commission, State Environmental Agency, Utilities and sometimes the US Army Engineer engineer. The assignment does not stop with generation. Major AI providers find electric transmission grids the most challenging challenge. Construction and interconnection requires several federal and state approvals and negotiations between utilities and electrical grid operators.
The timeline is incredible. Utility-scale solar or wind projects can spend up to three years just by negotiating permits and grid connections. The American Public Power Association (APPA), representing city-owned utilities, says that despite environmental regulators often have decision-making deadlines, opponents have learned to avoid deadlines by triggering new procedures, resulting in a “permission not approved” and a “constantly changing process.”
Meanwhile, the global AI race is getting hotter and demanding more from energy resources. Last year, US companies released 40 notable AI models compared to China’s 15. China has its own data center challenges, but countries are keeping up with patents and research production volumes, and powerful AI models are emerging from countries such as Canada, France and South Korea. The advantages of American innovation lie in staying ahead in both computing power and the electricity that drives it.
There have been efforts to streamline permissions by reducing management bloat and simplifying specific processes. Recent FERC decisions aim to streamline transmission plans. The APPA proposes implementing decision deadlines and eliminating overlapping environmental studies and redundancy when multiple agencies are reviewing projects. The Institute for Progress and the R Street Institute have additional suggestions. However, policy changes need to be stuck. Environmental requirements have been changed from Obama to Trump I, Biden and Trump II. Such flip-floping increases business costs.
Recent improvements in government efficiency in Iowa could serve as a template to allow reforms. Iowa Governor Kim Reynolds requested that agencies require that their central mission, the effectiveness of their programs, and identify where replicating and misorganization occurred. As a result, 21 agencies were removed from her cabinet, reducing the time it took to integrate licensing functions and to make regulatory decisions by 90%. Lawmakers can follow this model to test the usefulness and effectiveness of permitting and licensing procedures.
Texas offers a different model. The Texas Sunset Act of 1977 requires state agencies to be reviewed every 12 years. This process completely eliminated 42 agents and a further 53 by integrating or forwarding features. Legislatures and state legislatures may sunset the procedures for permitting problematic issues that exist in related agencies demonstrating that their requirements and procedures will provide economic benefits. This should be a truly valuable regulation problem.
If the US wants to lead the future of AI, it will need to clear the brushes to build electrical infrastructure.
