Analysis: AI data center power demand is approaching grid capacity limits, with electricity supply emerging as the primary bottleneck for further buildout scale.
American data centers driving AI are projected to grow in power consumption from 5GW in 2025—roughly equivalent to New York City's annual electricity usage—to 50GW by 2030. If AI infrastructure continues growing at the same pace, America would need 500GW of AI data centers by 2035, requiring more electricity than all American households, businesses, and other users currently consume in a year.
Having worked in the UK Government's energy department for eight years, including on data center power policy, I know firsthand how difficult it is to build infrastructure quickly. If AI companies want AI to become increasingly capable throughout the 2030s—as they claim is necessary to invent new technologies and cure diseases—they must start investing far more in energy infrastructure now.
AI companies currently invest an estimated $800 billion annually in data center construction, yet face significant logistical barriers. To accelerate deployment, many are sidestepping traditional power grid connections in favor of on-site gas generators. SpaceX's Colossus data center cluster in Memphis, Tennessee, touted as completed in just 122 days, relies on dozens of portable gas-fired turbines.
Growing numbers of Americans oppose these facilities, citing environmental and community concerns. The portable turbines are loud, degrade local air quality, and contribute to carbon emissions. In recent months, New York State imposed a moratorium on new data centers, while Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez proposed a federal moratorium on facilities exceeding 20MW, citing environmental, labor, and AI safety concerns.
Yet even if America embraces AI data centers and can manufacture chips quickly enough—two enormous uncertainties—the electrical grid cannot keep pace. Lead times for large gas turbines now exceed five years, and workarounds using portable turbines are themselves becoming stretched as more data centers pursue the same strategy. Many multi-gigawatt campuses planned for 2026 are already delayed; Bloomberg reports that of 12GW of data center campuses scheduled for this year, only one-third are even under construction.
If AI scaling trends persist, demanding more than 50GW of power by 2030 and hundreds of gigawatts by the mid-2030s, three potential futures emerge:
AI companies invest heavily in compute and energy infrastructure now, enabling uninterrupted scaling through the 2030s; America decides training more capable models is not a priority—perhaps through a global agreement to slow AI development; or American AI scaling slows without deliberate policy action.
Some argue that with appropriate workarounds or modest price increases, data centers can be built in two years or less. Proponents of approaches like the Offgrid AI white paper contend that solar panels and batteries enable rapid scaling. Solar and batteries can indeed be manufactured quickly and could support 100MW facilities. However, this approach fails for gigawatt-scale data centers. All large-scale power networks—grid-connected or behind-the-meter—require custom high-voltage electrical transformers and other made-to-order components. Even with everything else available within months, high-voltage transformers currently take five or more years to deliver.
New manufacturing investment is emerging, but the energy industry is wary of aggressive scaling. In the late 1990s and early 2000s, over 200GW of power plants were added to US grids between 2000 and 2005. When demand suddenly collapsed in the early 2000s, deregulation made it harder to forecast transformer requirements. By 2005, ABB—one of the world's largest transformer manufacturers—was forced to close facilities and eliminate 10 percent of its transformer workforce. "Overcapacity has been the biggest problem in the transformer industry in recent years," ABB's president and CEO Fred Kindle stated at the time. ABB shuttered another manufacturing facility in 2017, which remains closed despite the recent data center boom.
The industry abounds with similar stories: manufacturers of transformers and specialized electrical components ramped production to meet forecasted demand, only to watch that demand flatline for 15 years. AI data centers require these transformers, electrical switchgear, and components to obtain sufficient power, yet manufacturers are not scaling production fast enough to meet projected needs.
Alternatives have been proposed. Elon Musk claimed at Davos in January that "the lowest-cost place to put AI will be in space, and that will be true within two years, maybe three at the latest." Others propose ocean-floating solar data centers. Setting aside their extraordinary costs, current versions are effectively tiny facilities of 1MW or less—nowhere near the multi-gigawatt training runs required for frontier AI models. Researchers at Epoch AI have suggested that training could span multiple data center campuses rather than one, though this still assumes high-bandwidth networking between them. While technically feasible, distributed training across numerous 1MW facilities connected via the internet would substantially degrade performance.
The immediate opportunity lies in optimizing existing infrastructure. Power grids worldwide have line ratings based on worst-case weather conditions. Dynamic line rating—allowing power lines to carry extra load on safe-weather days—combined with data center flexibility to reduce or halt power draws for brief periods annually could significantly increase available capacity. Google has long championed flexible data centers, and the UK electricity regulator recently consulted on establishing new flexibility frameworks for large facilities. Even after such optimization, AI training will eventually be constrained by available physical infrastructure.
To guarantee uninterrupted model-training scaling, AI companies would need to de-risk the possibility of an AI bubble for transformer and electrical component manufacturers considering new factories. They could commit to purchasing large quantities of interchangeable transformers, which, though less efficient, can be manufactured faster and at scale—an advance market commitment analogous to government support for vaccine manufacturers.
Theoretically, AI companies could also mitigate manufacturing risk through milestone-based payments. However, economics make this virtually impossible. Despite projected revenues in the tens or hundreds of billions next year, current capital costs for ongoing compute provision mean companies cannot today afford orders for even a fraction of the power they will need in five years. SemiAnalysis founder Dylan Patel argues this dynamic will slow semiconductor production; the same constraint will throttle power infrastructure. AI companies may require government backing to convince manufacturers to scale up—for instance, the government could guarantee to cover portions of transformer orders if AI companies cannot.
Without immediate action, energy infrastructure will constrain the pace of technological development and economic growth. Conversely, it may provide humanity a crucial moment to understand new AI capabilities. An intelligence explosion may yet arrive—but it will arrive on the grid's timeline, not the industry's.