AI’s Impact on the US Electric Grid

This piece was originally written for the Thomson Reuters Institute. You can read the original here.

The challenges and implications of the growing energy demands of AI-driven data centers on the US electric grid highlights the need for significant infrastructure investments and legislative measures to mitigate risks to consumers.

When President Trump declared a national energy emergency in January, it was the first national emergency declared as such in the history of the United States. While the US continues to be a net exporter of fossil fuels, there are growing concerns surrounding the domestic electric grid’s capacity to sustain the rapidly increasing demand for electricity.

Indeed, energy usage from data centers alone is anticipated to double within five years, with forecasts suggesting they will account for about 12% of US electricity usage by 2028. The pace of AI innovation is simply speaking, soaring past the capacity of physical infrastructure as chip supply chains and energy demands have put added strain on the US power grid.

AI-driven data centers — operated by hyperscalers, a type of large-scale data center that offers massive computing resources, and utilizing large language models that are trained on vast quantities of information — require power-intensive processors and higher energy use. These AI workloads require high-power, high-density racks, and liquid cooling that result in significantly higher energy needs than general cloud computing. Data centers also need to be operational 99.999% of the time, requiring constant redundancies.

To meet future anticipated power needs, the US grid would have to quintuple long-distance transmission capacity within the next decade, which would come at an estimated investment of $720 billion.

Phantom load challenges and speculative growth

Already, AI data center growth is being compared to the real estate bubble of the 2000s as Big Tech and private equity rush to invest billions of dollars into data infrastructure. Quality Technology Services, a major data center operator that leases space to Amazon and Meta, was acquired by Blackstone for $10 billion in 2021. Alphabet, which owns Google, and Meta announced plans to invest upwards of a combined $125 billion into AI infrastructure this year alone.

The investment surge poses serious challenges for the nation’s energy utilities. Only a fraction of proposed data centers actually get built, making it difficult for grid operators and power system regulators to correctly estimate future electricity loads. So-called phantom data centers — projects that are proposed but never materialize — can create speculative interconnection requests. Developers often submit load requests in multiple markets, with the intention of only building in the market that offers the best incentives.

Thus, an overestimated load growth through duplicate requests develops and in fact, could trigger a utility to overbuild capacity without subsequent demand all at the expense of ratepayers.

And as infrastructure investment expands into both and red and blue states — such as Arizona, California, Georgia, Iowa, Nevada, North Carolina, Oregon and Texas — legislators are seeking to insulate ratepayers and ensure greater equity and transparency.

State legislative approaches to mitigate ratepayer risk

Now, utilities and lawmakers are stepping in to ensure that the risk of overbuilt infrastructure is not passed onto consumers. Georgia’s Senate Bill 34 would prevent Georgia Power from passing on costs of data center connection to residential and small business customers; and California’s Assembly Bill 222 proposes similar protections, but would additionally require energy consumption reporting annually. Even historically business-friendly Texas has sought to require developers to disclose duplicative requests and fund some interconnection requests through its Senate Bill 6.

Northern Virginia’s Data Center Alley is home to the world’s largest concentration of data centers with more than 250 facilities supplying 13% of all global and more than 25% of all domestic capacity. Yet, in Virginia, efforts to regulate after investment has occurred have largely failed. Bicameral legislation has failed or been abandoned that would require data centers to source from clean energy sources and sunset retail and sales tax incentives. Where legislation has failed, however, utilities have stepped up directly to protect ratepayers: Northern Virginia utilities including Dominion Energy, Appalachian Power, and Rappahannock Electric Cooperative have proposed new large-load rate classes to insulate residential and smaller load commercial ratepayers from higher costs.

DeepSeek, chip production & global trade tensions

Data centers not only rely on large quantities of electricity for their operations, but they also rely heavily on chips as the building blocks of their servers, which means that the data center market and the chip market grow in tandem. The rapid pace of demand for AI will inevitably lead to the Jevons paradox — the idea that as demand for a resource increases, its production will evolve to greater efficiency.

Currently, the chip and semiconductor markets are dominated by US-based Nvidia. This past winter, however, Chinese startup DeepSeek released R1, a free open-source AI assistant that uses less data and is a fraction of the cost of competitor products. This release had a short-lived, but the seismic market impact on Nvidia, as this model demonstrates that AI will in time require less data and become less costly, was significant. Indeed, it is also anticipated that as AI tasks become commoditized, XPU chips (designed for performing tasks) will be in greater demand than GPUs (chips designed for training.)

Chip access in the Trump Administration remains in flux as global tariffs loom over the tech supply chain. Malaysia has invested heavily in infrastructure for both chip and semiconductor production to meet the US desire for non-China sourced tech components. Last year Malaysia exported more than $16 billion in chips and nearly 20% of all semiconductors to the US. The proposed 24% reciprocal tariff on Malaysian imports into the US was postponed through early July, but it remains to be seen if tariffs will actually shift consumer behavior in amid AI’s robust demands — or if the cost will simply be passed onto end customers.

As AI-driven demand continues to reshape energy and infrastructure in the US, the balance between policymakers, utilities, and Big Tech investors will shape the speed of innovation and the ultimate cost to the public.