The OpenAI logo is being displayed on a smartphone with an AI brain visible. (Photo illustration by Jonathan Raa/NurPhoto via Getty Images)
NurPhoto via Getty Images
America is racing to deploy $1 trillion in AI-driven capital over the next decade, but that investment has some painful stress points: transformers, substations, and transmission lines can take years longer to expand than the data centers they power. This mismatch is not just a logistical headache; it is a ceiling on American productivity and a direct threat to AI sovereignty.
Infrastructure is now the ultimate arbiter of global tech dominance. The United States can design the smartest computer models and servers, but it still cannot run them without the massive amounts of electricity, water, and interconnections required for high-density computing. If the grid cannot keep pace, the AI race becomes less about intellectual innovation and more about a raw "land grab" for existing power, land, and grid capacity.
“The boom in energy consumption comes after years of insufficient investments in electric grids that have left data-center developers concerned they could face power shortages, particularly in 2027 and 2028,” analysts at Morgan Stanley wrote in a 2026 energy-market outlook.
The timeline disparity is the "Economic Chokepoint" of our era: A hyperscale data center campus can be built in roughly 18 to 24 months, yet the grid infrastructure needed to serve it—new transmission lines, upgraded substations, and reinforced distribution circuits—can take 5 to 10 years.
Interconnection queues at grid operators are now bloated with a mix of new power sources, batteries, and large loads. Some of the queue is new generation, such as wind and solar, or additional gas-fired plants. Some of it is storage, which helps smooth out peaks and integrate renewables. And some of it is new demand, such as AI-driven data centers, industrial facilities, and electric-vehicle charging hubs.
MORE FOR YOU
Tens of gigawatts of potential capacity sit in limbo, awaiting engineering studies, approvals, and financing. Data-center projects are delayed, investors grow impatient, and local communities watch promised jobs and tax revenue languish, even as the grid that enables the AI boom remains in the planning phase.
Who pays for all of this? In practice, utilities typically recover the cost of grid upgrades through tariffs spread across all customers, meaning that households and small businesses can end up subsidizing infrastructure for AI-driven loads. The economic risk is asymmetric: households pay for the build-out through tariffs, but if the AI bubble shifts or a project is abandoned, ratepayers are left holding the tab for “stranded” high-voltage infrastructure that serves no one else.
“If a hyperscaler is going to consume the equivalent of a midsize city's power, they should bear the cost of the infrastructure that serves them—not ratepayers,” David Stout, CEO of webAI in Austin, Texas, told me.
Detroit, Michigan, Residents picket DTE Energy, opposing the electric utility's plan to provide power for a proposed $7 billion data center in rural Michigan. They fear that it could raise residential electricity rates and endanger the water supply. (Photo by: Jim West/UCG/Universal Images Group via Getty Images)
UCG/Universal Images Group via Getty Images
What is the cost of failing to build? In the short term, it is constrained AI growth, stranded capital, and slower productivity gains. In the long term, it is ceding leadership in artificial intelligence to countries whose grids are more central-planned, more flexible, or more willing to prioritize one sector above the rest—notably China, Saudi Arabia, and the United Arab Emirates, where the state can effectively overbuild transmission and generation on behalf of AI-driven loads.
The demand is arriving in three overlapping waves:
The Belfer Center at Harvard’s Kennedy School estimates that data-center demand could grow from about 4.4% of total U.S. electricity in 2023 to somewhere between 6.7% and 12% by 2028, driven heavily by AI-driven workloads. Goldman Sachs analysts add that data centers already consume roughly 6% of total U.S. power and that share could nearly double to 11% by 2030.
The grid is no longer dealing with steady, predictable demand from factories and homes. It must now handle clusters of specialized computers that can ramp up to a gigawatt or more of power in minutes. That turns the grid from a relatively stable, predictable system into a more volatile, reactive one—harder to balance, harder to plan for, and more vulnerable to sudden swings.
Those pressures are already visible in Northern Virginia, where “Data Center Alley” has become the front line of the grid-versus-AI conflict. The region processes roughly 70% of global internet traffic, effectively serving as the digital chokepoint. Now it is also facing a torrent of new AI-driven load, with utilities and grid operators preparing for tens of gigawatts of additional demand from data centers over the next decade.
PRODUCTION - 25 August 2025, Hesse, Friedberg: Surveillance cameras are located at the substation on Ovag's open-air site. (Photo by Andreas Arnold/picture alliance via Getty Images)
dpa/picture alliance via Getty Images
Dominion Energy, which serves that corridor, has warned regulators that roughly 70,000 megawatts of new load could come from data centers in the coming years, far exceeding the transmission capacity built for the region.
“If the grid doesn't keep pace, America loses the AI race—full stop,” Mark Meckler, president of Convention of States Action in Houston, told me. “AI is the most energy-intensive technology revolution in human history, and the country that can deliver reliable, abundant, affordable power to data centers will set the rules for the next century of computing, defense, and finance.”
Those new projects bring tax dollars and jobs, but they also strain local water supplies, congest transmission lines, and push up electric bills for other customers. Some states are experimenting with new rate structures that force AI-driven loads to pay more for grid upgrades, hoping to insulate households and small businesses from the full cost.
The story does not end at the utility fence. If the grid is running near capacity just to keep the lights on and the AI models humming, it can suffer sudden outages or become a target for state-sponsored cyberattacks. I just examined the Russia-led cyber campaigns targeting infrastructure in Poland and Italy and found that a grid pushed to its physical limit can’t withstand coordinated state-sponsored disruptions. That vulnerability links energy security directly to AI supremacy.
There is a growing chorus of experts who argue that the grid is not fundamentally short of capacity, just poorly managed. They point to evidence that most advanced-economy grids operate at a small fraction of their theoretical peak on an average day, and argue that better coordination, demand response, and “software-defined grid” tools can squeeze far more value out of existing assets.
However, even the most flexible grid must have sufficient physical capacity to handle peak demand. If data center developers are waiting 4 or 5 years to get a grid connection, then flexibility alone cannot solve the problem. You can shift the load around, but you cannot change the physics of a transmission line that simply cannot carry more megawatts without an upgrade.
The AI age is forcing a reckoning: We are trying to run a 21st-century information revolution on a grid that remains a 20th-century artifact. Indeed, Goldman Sachs estimates that almost all U.S. power grids will lack the capacity to meet data center demand by 2030.
Energy security and AI supremacy are no longer separate conversations. The country that can build resilient, flexible power systems—capable of integrating renewables, storage, and data centers without breaking—will have an edge that pure software innovation cannot replace. In the race for AI supremacy, the copper in the ground matters as much as the silicon in the chip.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。