AI Data Centers Amplify Grid and Energy Price Risks

On July 23, 2026, the White House expanded a voluntary Ratepayer Protection Pledge to governors, utilities and data center developers. It said the pledge now covers 80% of the electricity delivered to U.S. homes and businesses, with large data center operators expected to fund the generation and grid infrastructure their projects require. Two days earlier, the House Energy and Commerce Committee voted 52-0 to advance a bill requiring state utility commissions to consider ways of making data centers pay their own costs. The measures put one question at the center of the AI boom: who will carry its electricity bill?

A new working paper by Olivier Darmouni, Clemens Lehner and Yuqi Zhang studies this relatively unreported issue and draw a complicated picture. The national spillover to electricity prices and output may remain modest. The costs become much larger in a few exposed regions, while the buildout raises emissions, reduces the reserve capacity that protects grid reliability and increases dependence on volatile natural gas prices.

The paper, titled The Energy Cost of AI and Data Centers, combines plans for 420 U.S. data center developments with regional supply curves built from individual power plants. The model projects how new electricity demand could affect prices, output outside the data center sector, energy profits and emissions through 2035. It isolates the energy consequences of the buildout and leaves any productivity gains from AI outside its scope.

How concentrated is the buildout?

The paper starts from an exceptional increase in demand. The International Energy Agency expects U.S. data center electricity consumption to roughly double to 240 terawatt-hours by 2030. Data centers could then use more electricity than all energy-intensive manufacturing sectors combined and account for up to half of the growth in U.S. electricity demand through the end of the decade.

The researchers identify about 113 gigawatts of planned data center capacity by 2035. The projects are highly concentrated: only 136 U.S. counties, about 4% of the total, host an announced development, while the top 20 counties account for 60% of planned capacity. Just 42 counties account for 80%.

Those counties are populous and already close to fossil-fuel infrastructure. They have more than five times as much natural gas generation capacity as other counties and sit closer to major electricity loads. Their wind and solar resources are not significantly better. The planned buildout is therefore clustering where firm power and transmission already exist, rather than where the local renewable resource is strongest.

Why is the national effect modest?

Under the paper’s 2035 scenario, data centers pay about $85 billion a year for electricity. If no new generation is built, higher electricity prices reduce output in other sectors by about $20 billion a year, or around 0.07% of 2023 U.S. GDP.

Two features of the present power system limit that national effect. Existing fossil-fuel plants operate below capacity in many regions, leaving relatively flat supply curves around current demand. Electricity also represents only about 2% of aggregate production costs, so a moderate price increase produces a limited decline in output across the wider economy.

The distribution of gains is very different. Electricity generator profits rise by about $60 billion a year and upstream fuel providers gain roughly $20 billion. The additional emissions carry an estimated climate and social cost of about $80 billion a year when carbon is valued at $200 per metric ton.

These figures are not a complete cost-benefit account of artificial intelligence. Data center demand is fixed in the model, and the researchers do not estimate the value of the services or productivity gains the facilities might support. Their question is narrower: how does a large new electricity load travel through the existing grid and the rest of the economy?

Why do some regions pay far more?

The national average conceals a sharply uneven map. Texas, Virginia and the Carolinas account for two-thirds of the projected loss in non-data-center output. In these areas, the loss reaches about 0.3% of state GDP and wholesale electricity prices rise by 20% to 40% when no new power is built.

The amount of new demand alone does not determine the outcome. Texas receives about 40% less data center load than Virginia and the Carolinas in the model, yet its output loss is twice as large. The Pacific Northwest also experiences a comparatively strong price effect despite a smaller buildout. Local spare capacity and the shape of the regional supply curve decide how expensive the next unit of electricity becomes.

The researchers test the importance of location by shifting 20% of planned demand from other regions to New York and California. Data center electricity bills rise by about 40% in that scenario. The result helps explain why developers favor areas with available generation and transmission, even when those areas are already heavily exposed to the buildout.

Can new generations contain the spillovers?

Building local power supply reduces the increase in prices and brings the projected output loss close to zero. Meeting all of the new demand with additional capacity, however, requires between $300 billion and $800 billion in capital expenditure. The lower estimate corresponds to gas generation; the upper estimate uses renewables with storage.

That investment has become more expensive. Compared with projections made before 2022, current capital-cost estimates are 42% higher for wind, 53% higher for solar and storage, and 65% higher for combined-cycle gas generation. Turbine shortages and wider supply-chain pressures also lengthen the time needed to bring new plants online.

In the model, moving generation behind the meter takes both the data center load and its new power supply off the public grid, leaving output losses in other sectors virtually unchanged. If that private supply relies heavily on gas, emissions and fuel-provider profits increase. A high renewable share is needed to limit the additional fossil-fuel use.

Do renewables lower electricity prices?

Replacing new gas plants with renewable generation has little effect on average electricity prices or output in the model. Both technologies add supply, while an existing fossil-fuel plant often remains the marginal generator setting the market price. The choice of technology matters far more for emissions and for who receives the energy-sector profits.

Renewables are dispatched first because their marginal costs are low. They reduce purchases from fuel providers, shift profits toward electricity generators and sharply lower the climate cost of the buildout. They also provide protection against a rise in natural gas prices.

When the researchers increase gas costs by 20%, non-data-center output falls by roughly 0.15% of GDP, an effect comparable to the entire projected data center demand shock in the scenario without new generation. Building renewable power reduces that amplification, while new gas capacity reinforces the economy’s exposure to the same fuel-price risk.

What does the average price miss?

The spare generation capacity that keeps the average price effect modest also serves as a buffer during extreme weather and other exceptional demand peaks. Large, constant loads from data centers use part of that margin. The grid may therefore appear capable of absorbing the new annual demand while becoming less able to manage the hours when the system is under greatest strain.

The model works with annual electricity quantities and does not simulate hourly dispatch, storage, ramping constraints or blackouts. It cannot place a number on the reliability risk. The paper nevertheless identifies reduced grid reliability as one of the two main structural risks amplified by AI demand, alongside dependence on volatile fossil-fuel prices.

The exposure also grows as the rest of the economy electrifies. In a scenario where electricity’s share of production costs doubles from 2% to 4%, the same data center demand produces substantially larger output losses. Electric vehicles, heat pumps and cooling systems would then compete more directly with AI infrastructure for power.

In a recent HEC Paris Breakthroughs podcast, Darmouni discussed how the same buildout could reshape investment in power infrastructure and force Europe to think harder about energy security. The paper itself focuses on the United States, but its regional method shows why a national average can conceal the locations where the pressure is greatest.

Who should pay for AI’s power?

The study does not model the retail rules that determine how utilities divide infrastructure costs among data centers, households and other businesses. Those rules are already changing. Virginia and Ohio have introduced special minimum-payment requirements for very large electricity users, while the federal pledge and the bill advanced in July 2026 seek to keep new data center costs away from other ratepayers.

The model identifies what those rules must allocate: hundreds of billions of dollars in new generation, localized price and output effects, higher emissions, reduced reserve capacity and greater exposure to natural gas prices. Each burden changes with the location of the project and the type of power built to serve it.

The study’s central conclusion is that the largest economic cost lies in the way AI demand amplifies two structural risks: less room on the grid when demand surges, and greater exposure to volatile fossil-fuel prices. That vulnerability is already becoming visible. The 2026 Strait of Hormuz crisis has disrupted global oil and liquefied natural gas flows and produced sharp swings in energy prices, while electricity demand is reaching new peaks in the regions most exposed to data-center growth. ERCOT set an all-time record of 91,089 megawatts on July 22, and PJM used emergency demand response as extreme heat pushed its system close to record demand earlier that month. Strong U.S. gas production has so far provided some protection, but the overlap between geopolitical fuel risk and rapidly rising electricity use closely resembles the pressure identified in the research.