Whatever you hear about the United States as a fading power, it is ahead in the race to gain dominance in artificial intelligence over China, its major rival. That’s a product of Silicon Valley’s unique ability to bring together scientists, entrepreneurs and risk capital. Yet the prospect of continuing dominance in this 21st-century technology hinges on harnessing a 19th-century one: electricity.
That’s where America has a long way to go.
The development of advanced A.I. systems requires vast amounts of energy. At the heart of training these systems are large numbers of specialized computer chips. One estimate suggested that training GPT-4, the latest ChatGPT A.I. system, consumed roughly the same amount of electricity as several thousand U.S. households use in a year.
These extraordinary demands are already pushing up against real-world constraints. America’s power grid, hindered by decades of underinvestment and regulatory logjams, isn’t equipped for the rapid growth in A.I.’s electricity needs. Across the country, energy investors are waiting to develop 2.6 terawatts of new electrical capacity, mostly in wind, solar and battery farms. The total generating capacity that is not connected to the grid and is waiting to be connected has grown roughly eightfold since 2014. Adding that would about triple national generating capacity and help address the future needs of A.I.
In Virginia — a hotbed for data centers — the wait time for data centers to connect to the grid could be seven years. Some counties in the state are introducing limits on data centers.
These projects are stalled by a web of fragmented policies and lengthy approvals. This isn’t China holding back U.S. competitiveness; it’s America’s patchwork of conflicting regulations, outdated structures and misaligned investment incentives that discourage coordinated grid expansion.
In stark contrast, China can move infrastructure projects — from power plants and transmission lines to entire data centers — from blueprint to reality in a small fraction of the time. Over time, this sheer capacity might help China close the A.I. gap or even overtake the United States. As of December 2023, China had built 34 ultra-high-voltage transmission lines — totaling tens of thousands of miles — enabling efficient long-distance power delivery, while the United States had none.
Chinese A.I. companies currently struggle to gain access to the latest chips designed in the United States because of export controls. However, a gray market for these chips is forming, and it is only a matter of time before more competitive Chinese-built alternative chips emerge, through innovation out of necessity. If this happens, the companies’ ability to rapidly expand power capacity could allow them to operate huge training clusters.
Since 2000, China has expanded its electrical capacity roughly sevenfold; In 2023, according to the energy think tank Ember, it added 355 gigawatts of new capacity, compared with 29 gigawatts in the United States. In the past decade, China added more than three dozen nuclear reactors, while the United States managed to add only three. China’s nuclear projects often go from approval to operation in seven years or less; a couple of the most recent reactors built in the United States took more than a decade to become operational.
The reasons for this sloth in the United States are varied: a skills gap and labor shortages in construction and engineering, a complex permitting process trapping projects in years of bureaucratic review across multiple agencies and high costs of capital. Local opposition compounds these delays, significantly slowing the rollout of new energy infrastructure.
American tech giants can’t ignore these problems and, in some cases, are taking matters into their own hands. Microsoft recently secured a deal to power future data centers with electricity from one of the previously decommissioned reactors at the Three Mile Island nuclear facility — enough to power hundreds of thousands of homes. Google announced a partnership that would put $20 billion toward building renewable energy projects next to new data centers.
The urgency is also forcing Big Tech to turn to dirtier power sources, including coal, often at a cost to their decarbonization ambitions and potentially to the detriment of the communities near coal plants.
Meta’s plans have included bringing on board a huge new gas power initiative in Louisiana; Microsoft, Google and Elon Musk’s xAI are similarly relying on fossil fuels. These projects may ultimately have a marginal effect on America’s efforts to reduce carbon emissions, but they are not enough to overcome the power constraints for A.I. or to resolve the bottlenecks.
This is where the government can step in, with a clear declaration of intent to meet the demands of the 21st century. America’s electricity system needs to modernize and expand using clean technologies like solar and nuclear. (Disclosure: I am an investor in a company that helps develop renewable power projects in Britain.) Policy and regulations must change to foster faster approvals for power plants, grid connections and transmission. The intersection of federal, state and local rules creates additional complexity, if not outright gridlock.
The United States has already established the importance of semiconductors through the CHIPS and Science Act and export controls aimed at boosting an A.I. economy and safeguarding national security. Now it must apply similar focus and urgency to upgrading its electrical infrastructure. Maybe it is time for the creation of an Energy Acceleration Authority that would have a mandate to streamline approvals for critical clean energy projects.
Electricity is more than just a utility; it’s the bedrock of the digital era. If the United States truly wants to secure its leadership in A.I., it must equally invest in the energy systems that power it.
The post The 19th-Century Technology That Threatens A.I. appeared first on New York Times.
