AI Data Center Energy Consumption and the Everything-to-Grid Shift

AI data center energy consumption is currently a small share of total electricity use and a very large share of its growth. The International Energy Agency put global data centre consumption at about 415 terawatt-hours in 2024, roughly 1.5 percent of world electricity, and projects around 945 TWh by 2030, just under 3 percent. The more consequential number is at the margin: the IEA reported that data centres accounted for roughly half of all US electricity demand growth in 2025 and expects that to hold through 2030. That is why the grid conversation has shifted from generation alone to “everything to grid,” the idea that data centers, EVs, home batteries and thermostats all become dispatchable resources.

The Numbers, and How Much Confidence to Place in Them

Every credible forecast of data center electricity demand is a range rather than a point, and the ranges are wide enough that all of them are estimates.

Why AI Load Is Not Like Any Load Utilities Have Planned For

Utilities spent two decades planning around flat demand. AI data centers break several of those planning assumptions at once.

The Supply Side: What Is Under Contract Versus What Is Built

The near-term answer is gas, and gas is supply-constrained. GE Vernova has described its heavy-duty gas turbine slots as effectively sold out through 2030, and Bloomberg reported in 2025 that Siemens Energy and Mitsubishi Power face similar difficulty keeping up with AI-driven orders. A turbine order placed today generally does not produce electrons before 2030.

Nuclear is the headline commitment, and it is almost entirely a 2030s story. The publicly announced hyperscaler agreements include:

Two caveats. Agreements on existing reactors mostly reallocate carbon-free power already on the grid rather than adding supply, and no US small modular reactor is generating commercial power today, so every SMR gigawatt above is contracted intent rather than capacity. The bridge is gas, existing nuclear, batteries, and on-site generation installed behind the meter to skip the interconnection queue.

Everything to Grid: EVs, Home Batteries and Thermostats as Capacity

“Everything to grid” is shorthand for treating every controllable device as a grid resource: vehicle to grid and vehicle to home for EVs, battery-to-grid for home storage, and thermostat or water-heater control as dispatchable load. The mature and immature parts differ sharply.

Vehicle-to-grid is genuinely early. The concrete 2025 and 2026 milestones are regulatory and pilot-scale rather than gigawatt-scale:

• Maryland adopted what are described as the first comprehensive US V2G interconnection rules in June 2025.
• Sunrun and Baltimore Gas and Electric ran a residential V2G pilot using Ford F-150 Lightnings that linked dozens of households, not thousands.
• Nuvve, Resource Innovations and ComEd began exploring bidirectional electric school bus charging in Illinois in February 2025.
• Roughly eight states, including California, Maryland, Massachusetts and New Jersey, have enacted or proposed V2G and battery aggregation legislation.

The realistic read: virtual power plants can shave peaks for hours, which is valuable and cheap, but they do not supply a gigawatt of continuous training load. They are a reliability tool, not a substitute for generation.

Demand-Side Flexibility Is the Cheapest New Capacity Available

The most useful research here asks a different question: how much new load could the existing grid absorb if that load agreed to back off occasionally?

The Nicholas Institute at Duke University published “Rethinking Load Growth” in February 2025, finding that the 22 largest US balancing authorities could host roughly 76 to 126 GW of new flexible load depending on accepted curtailment:

The authors noted the analysis excluded transmission constraints and worked from peak demand levels, so it is an upper bound. Field results so far are encouraging but small:

• EPRI demonstrations have reported flexibility of up to about 40 percent of a facility’s draw.
• Emerald AI, working with Salt River Project, reported a 25 percent load reduction sustained across a three-hour peak.
• Portland General Electric reported about a 20 percent power reduction in a simulated emergency scenario.
• Boston University modeling suggests AI workloads could offer 18 to 55 percent flexibility relative to average power draw.

A 2026 Duke analysis estimated that a 1 to 2 percent reduction in data center peak demand could lower electricity rates by roughly 0.5 to 2.8 percent. The nuance that gets lost: training is far more flexible than inference. A training run can pause and checkpoint. A user waiting on a response cannot. As inference grows as a share of total compute, the flexible fraction of the load shrinks, which is one reason the cost pressures described in AI Inference Cost Economics and the grid pressures described here are the same problem viewed from two ends.

Interconnection Queues, Phantom Demand, and Who Pays

Queue figures are large and partly fictional. Texas has been reported carrying a large-load interconnection queue in the hundreds of gigawatts, far beyond anything that will be built, because a single project often files with several utilities in several states while shopping for the fastest power. Utilities are responding with large-load tariffs requiring deposits, minimum-take commitments and collateral, precisely to filter speculative requests. For grounded numbers, one 2026 analysis put actual US data center demand near 31 GW in 2025 rising to about 66 GW by 2027, with the data center share of summer peak going from roughly 4.1 to 8.5 percent.

On cost allocation, the evidence is more contested than the headlines suggest:

A December 2025 PJM auction reportedly fell short of its reliability requirement by roughly 6,600 MW, described as a first for the market. An E3 whitepaper published May 18, 2026 attributed about half of the PJM capacity price increase to load growth, with the balance coming from market design changes, plant retirements and reduced accreditation of fossil resources. E3 also found no evidence of a historical cost shift from data centers to residential customers, and noted that states with the largest load growth, Texas and Virginia, saw some of the smallest rate increases while California and New York saw the largest with flat or falling load.

A January 2026 PowerLines analysis reported US power bills up about 40 percent from 2021, with utilities requesting more than $30 billion in rate increases during 2025, roughly double the prior year. So data centers are a real driver of capacity prices in specific markets and not the sole driver of retail bills. Pew polling in March 2026 found Americans more negative about data centers’ energy and environmental effects than positive about jobs and tax revenue, which is why ratepayer-protection conditions are now standard in state proceedings.

Water deserves a caveat rather than a number. Published figures vary enormously depending on whether a site uses evaporative cooling, closed-loop liquid cooling or dry coolers, and facility-level disclosure is still rare. It is a genuine siting constraint in arid regions and an area where public data is weaker than the electricity data.

What Plausibly Gets Built by 2030

Stripping out the announcements that will not clear permitting or financing, the shape of the buildout is reasonably predictable:

1. Gas carries the load, constrained by turbine delivery slots that are already committed through the end of the decade.
2. Existing nuclear gets contracted, not expanded, with the Crane restart in 2027 as the notable addition and SMRs arriving in the 2030s at the earliest.
3. Behind-the-meter generation and storage grow fast, because waiting in an interconnection queue is more expensive than buying your own turbines and batteries.
4. Flexibility becomes a tariff category. Curtailable large-load contracts in exchange for faster energization are the single cheapest lever available, and 2026 pilots are the proof-of-concept stage.
5. Virtual power plants become normal procurement, at gigawatt scale for peak shaving, while vehicle-to-grid stays in pilots for most of the decade.

The reason this gets so much attention is that the alternatives are worse. Moving compute off the planet entirely is a real engineering program, covered in Space Data Centers: SpaceX, Nvidia and the Orbital AI Bet, but the published math puts cost parity in the 2030s at best. On the ground, the honest framing is that AI energy consumption is a solvable engineering and regulatory problem that has been made harder by a mismatch in clock speeds between software and infrastructure.

For most businesses the implications are small but not zero. Power constraints show up as regional capacity limits and price gaps between cloud regions, which is a real input into application development decisions about where workloads run. A generated AI answer also costs meaningfully more compute than a classic ranked result, which is worth understanding alongside how search engines work, and even routine systems like marketing automation now sit on infrastructure whose marginal cost is set partly by electricity prices in a handful of states.

Figures in this area move quarterly, so anchor on dated primary sources: the IEA’s Energy and AI and Electricity 2026 reports, Berkeley Lab’s report for the Department of Energy, EPRI’s Powering Intelligence, Duke’s Rethinking Load Growth, and PJM’s auction results.