AI data centers are projected to consume nearly 1,000 terawatt-hours (TWh) of electricity in 2026 — roughly double 2022 levels — yet utility interconnection timelines of three to five years cannot keep pace. The mismatch is forcing behind-the-meter natural gas plants, delaying coal plant retirements, and raising consumer electricity bills by 1.5% to 5% in affected regions. This is AI’s power paradox.
Why Data Center Demand Is Outpacing the Grid
In 2024, global data centers consumed 415 TWh — about 1.5% of world electricity — and the International Energy Agency projects that figure will double to 945 TWh by 2030. Accelerated AI servers are growing roughly 30% annually. AI data center energy consumption now drives nearly half of the net increase in global data center electricity use.
The constraint is not chips but power. Goldman Sachs Research sees U.S. data center demand jumping from 31 GW in 2025 to 66 GW by 2027, while new projects face grid interconnection queues of four to ten years and transformer lead times of three to five years.
Behind-the-Meter Gas and Delayed Coal Retirements
To bypass the queue, hyperscalers are building “behind-the-meter” generation at data center sites. Behind-the-meter natural gas plants have become the fastest fix: proposals for on-site gas capacity tripled to more than 250 GW in 2025. In December 2025, FERC ordered PJM to allow data centers to co-locate directly at power plants, eliminating the old “first use” rule.
The fossil-fuel ripple effect is measurable. At least 15 U.S. coal plants have postponed or canceled retirements since January 2025, driven largely by AI data center demand, according to a DeSmog analysis. Those plants emitted over 68 million tons of CO2 in 2024 — more than Delaware, Maryland, and Washington, D.C. combined. Coal plant retirement delays now stretch from Georgia to Wyoming, with Southern Company extending two Georgia coal plants to 2039. Energy Secretary Chris Wright calls it winning the AI race; the Sierra Club calls the delays unacceptable.
Grid Prices and Consumer Costs
PJM’s 2026/2027 capacity auction cleared at a record $329.17 per megawatt-day — an eleven-fold jump from $28.92 a year earlier. The next auction produced a 6,623 MW reliability shortfall, the first system-wide gap in PJM history. PJM capacity auction prices are translating into residential bill increases of 1.5% to 5%, on top of earlier hikes that reached 30% in some territories.
Goldman Sachs estimates utilities globally will need $720 billion in grid upgrades by 2030, while EPRI projects data centers could consume 9% to 17% of U.S. electricity by 2030. In Virginia’s “Data Center Alley,” demand hit 12.1 GW in 2025 and connection timelines extend to seven years.
What It Means for Climate Commitments and Policy
Net-zero pledges are colliding with operational reality. Microsoft, Amazon, Google, and Meta have signed nuclear and geothermal deals, including small modular reactors that will not deliver firm power before 2030. In the meantime, natural gas and delayed coal retirements fill the gap. Analysts warn that reliable power access has replaced GPUs as the defining bottleneck for AI expansion.
Frequently Asked Questions
How much electricity will AI data centers use in 2026?
Projections put annual AI data center consumption near 1,000 TWh in 2026, roughly double 2022 levels. The IEA’s base case sees about 945 TWh by 2030.
Why are grid interconnection timelines so slow?
Permitting, siting, and equipment shortages have stretched connection timelines to three to five years, and up to seven to ten years in congested regions like Northern Virginia.
Why are coal plants delaying retirement?
At least 15 U.S. coal plants have postponed retirement because surging data center demand makes their baseload power hard to replace in the short term.
How much are consumer electricity bills rising?
In PJM and other affected markets, residential bills are increasing by 1.5% to 5%, with some utility territories seeing cumulative increases of 30% or more.
What are behind-the-meter power deals?
These agreements let data centers connect directly to adjacent generation, bypassing transmission charges and interconnection queues, though critics warn they can shift costs to other ratepayers.
Future Outlook
Gartner predicts 40% of AI data centers will face power-driven operational constraints by 2027. The companies that secure firm, scalable electricity will set the pace of AI deployment. The grid, not the GPU, is now the critical path.
Follow Discussion