Buried inside the modeling behind the Environmental Protection Agency’s September 17 partial repeal of the 2024 Carbon Pollution Standards is a consequential assumption about the next power grid: electricity demand will rise sharply, and data centers supporting artificial intelligence will be a primary driver.
In the final rule, EPA removes three pieces of the 2024 framework. It repeals emission guidelines for existing fossil-fuel steam generating units, carbon-capture standards for large modifications to coal plants, and the phase-two carbon-capture requirements for new base-load combustion turbines. It takes effect November 16. A broader repeal remains a separate proposal, so this is a final but partial change rather than the end of the larger regulatory fight.
Within that regulatory case, AI demand did not single-handedly cause the repeal. EPA’s legal and technical argument rests on a reassessment of whether carbon capture and natural-gas co-firing qualify as adequately demonstrated systems of emission reduction, given their cost and practical limits. But the rule’s Regulatory Impact Analysis makes the relationship explicit: its updated power-sector model assumes “significantly higher electricity demand driven primarily by data center use in applications supporting artificial intelligence,” alongside changes from 2025 tax and energy legislation.
Once that forecast enters the model, it becomes part of the policy machinery. A projection about AI infrastructure is no longer confined to utility planning or an investor deck. It is helping define the operating environment in which a federal agency weighs cost, feasibility, reliability, and emissions.
That role makes the accounting boundaries especially important. EPA estimates that the repeal will reduce compliance costs by $160 billion in present value through 2047 using a 3% discount rate, or $95 billion at 7%. Its broader real-resource calculation produces larger savings of $280 billion and $180 billion. Those figures describe different measures and should not be blended into one headline number.
On the other side of that ledger, the analysis projects higher carbon dioxide emissions than under the 2024 standards: 20 million metric tons more in 2030, 406 million in 2035, 533 million in 2040, and 486 million in 2045. EPA did not assign a monetary value to those additional emissions, citing Executive Order 14154. The agency’s roughly $23 billion estimate for annualized social-cost reductions therefore excludes changes in environmental quality. This is a choice about what enters the ledger, not a complete tally of every consequence.
Meanwhile, other federal and grid assessments support the direction of the demand shift while warning against false precision. The Energy Information Administration’s September outlook expects US electricity sales to rise from 4,135 billion kilowatt-hours in 2026 to 4,211 billion in 2027, driven largely by data centers and manufacturing. Its longer-range scenarios vary according to data-center construction, computing power, and efficiency. The North American Electric Reliability Corporation says AI and digital data centers account for most projected demand growth, while also noting that large-load forecasts remain volatile because announced projects can slow, move, or disappear.
Taken together, these documents show forecasts about AI electricity use beginning to carry regulatory weight before the shape of that demand is settled. Agencies will still have to make decisions under uncertainty. The public-interest test is whether people can see which forecast was used, which assumptions changed, which costs were counted, and what evidence would trigger another review.
Verification bottleneck
Three questions deserve a durable record:
- Which demand scenario and model vintage supported the final rule, and how did the result change across lower and higher load cases?
- How quickly can the forecast be revised when a large data-center project is delayed, relocated, or cancelled?
- Where can a reviewer see compliance savings, unpriced emissions, reliability effects, and uncertainty together rather than in separate ledgers?
These questions follow from the RIA’s unusually direct uncertainty warning. It says future outcomes are sensitive to demand, fuel prices, capacity additions, carbon-capture performance and cost, and the feasibility of changes to the power grid. That warning should travel with the headline savings figures.
Opportunities
One practical response would be a public forecast-to-decision receipt for policies built on fast-moving technology demand. It could record the model version, demand range, estimated AI and data-center share, assumptions that moved since the previous run, the decision those assumptions informed, and a scheduled review condition.
This would not eliminate uncertainty or settle the policy argument. It would make a moving premise easier to inspect before it hardens into infrastructure, rates, or regulation.
Sources
- Federal Register: Partial Repeal of the Carbon Pollution Standards for Fossil Fuel-Fired Electric Generating Units
- EPA: Regulatory Impact Analysis for the Final Partial Repeal
- EPA: Final-rule announcement
- EIA: Short-Term Energy Outlook, electricity data
- EIA: Annual Energy Outlook 2026 narrative
- NERC: 2025 Long-Term Reliability Assessment