Fazlur Rahman
Last verified: October 5, 2026.
AI used about 4.4% of US electricity in 2024 — one in every twenty kilowatt-hours. By 2028, projections put the share between 6.7% and 12%, depending on how fast training runs scale. The doubling of demand in four years is not unusual for new technology; what is unusual is the grid's response: 92% of new US generation in 2025 was solar plus battery, the Three Mile Island restart is sold out to a hyperscaler, and gas plants are being commissioned at a pace not seen since 2007. The grid is being rebuilt for the AI cycle, and the cost of it shows up on every utility bill.
This piece walks through the demand curve, where the new demand is concentrated, the cost stack that hyperscalers pay, the supply response that is mostly solar plus battery, the regional trade-offs that decide where AI training happens, the efficiency gains that eat the demand growth, and whether the grid breaks before it bends.
The Demand Curve, In Numbers
AI data center electricity consumption has grown faster than any single segment of US power demand in the post-2010 record:
| Year | Data center electricity (TWh) | Share of US electricity | Annual growth |
|---|---|---|---|
| 2018 | ~200 TWh | ~4.0% | Baseline |
| 2020 | ~210 TWh | ~4.1% | +2-3% |
| 2022 | ~230 TWh | ~4.5% | +5-7% |
| 2024 | ~176 TWh (AI only) | ~4.4% | +12-15% |
| 2026 (projected) | ~280 TWh (AI + traditional) | ~6.7% | +15-20% |
| 2028 (projected) | ~400-500 TWh | ~9-12% | +20-25% |
| 2030 (projected) | ~600-800 TWh | ~14-18% | +20%+ |
The doubling window is 2024 to 2026 (176 → 280 TWh) and again 2026 to 2028 (280 → 450-500 TWh). That two-year doubling is the rate at which AI training runs scale — a frontier training run in 2024 consumed roughly 5 GW of peak power for two months; a 2026 frontier run consumes 10-15 GW for longer. The implication for the grid: doubling every two years is a pace that the US has never seen outside wartime industrial mobilisation.
Where the Demand Is Building
Eight US data center markets carry most of the operational and pipeline capacity:
| Market | Operational MW (2026) | Pipeline MW | Primary power source | Constraints |
|---|---|---|---|---|
| Northern Virginia (Loudoun, Prince William, Fairfax counties) | 2,600+ MW | 5,000+ MW | PJM grid (mixed) | Severe capacity shortfalls; transmission |
| Texas (Dallas, Austin) | 1,500+ MW | 4,000+ MW | ERCOT (gas, wind, solar) | Summer/winter reliability; interconnection wait |
| Phoenix, Arizona | 1,200+ MW | 3,500+ MW | SRP + APS (mixed) | Water stress; capacity constraints |
| Columbus, Ohio | 1,000+ MW | 2,500+ MW | AEP grid (mixed) | Transmission; coal plant retirements |
| Atlanta, Georgia | 800+ MW | 2,000+ MW | Georgia Power (nuclear, gas, solar) | Capacity tight; nuclear expansions |
| Dallas-Fort Worth | 700+ MW | 1,800+ MW | ERCOT | Reliability; cost allocation |
| Quincy, Washington | 600+ MW | 1,200+ MW | Hydro + nuclear | Limited; environmentally constrained |
| Hillsboro, Oregon | 500+ MW | 1,000+ MW | BPA (hydro, wind) | Power shortages; transmission |
Northern Virginia is the story. 2,600 MW operational plus 5,000 MW pipeline equals 7,600 MW total — roughly 7.6 GW of planned demand concentrated in a region whose grid was sized for retail and commercial customers, not hyperscaler loads. Loudoun and Prince William counties have data center load equal to one or two large nuclear reactors' worth of generation, with most of it landing on the PJM grid's existing capacity. The pipeline is the problem: at 5,000 MW, it would require 4-5 years of build-out at current construction rates.
The Cost Stack: $/MWh Math
The cost-per-kilowatt-hour math of AI data centers explains the supply response:
| Data center size | Capacity (MW) | Annual electricity cost @ $0.10/kWh | Annual electricity cost @ $0.06/kWh (PPA) |
|---|---|---|---|
| Small edge facility | 10 MW | $8.76M | $5.26M |
| Mid-sized enterprise | 100 MW | $87.6M | $52.6M |
| Large hyperscaler campus | 1 GW (1,000 MW) | $876M | $526M |
| Massive AI training campus | 5 GW (5,000 MW) | $4.38B | $2.63B |
A 1 GW hyperscaler campus running at $0.10/kWh pays $876M; at the $0.06/kWh PPA price, $526M. The 350M annual difference is what makes hyperscalers sign 10-15 year power purchase agreements (PPAs) with wind and solar developers — the discount is real enough to cover the off-take risk. A 5 GW AI training campus, the largest scale now being commissioned, runs $2.6-4.4 billion a year just in electricity.
The Supply Response: 92% Solar + Battery
Capacity additions in 2024-2026 track the AI demand curve. Solar capacity added 32 GW in 2024, projected 50+ GW for 2025-2026. Battery storage added 10 GW in 2024, projected 18+ GW in 2025-2026. EIA's 2025 data shows 92% of new US electricity generation was solar plus battery storage, the highest share for any single source category in modern US power history.
Gas peakers fill the flex role for when solar is offline. New gas plants are being commissioned at a pace not seen since 2007, mostly as peaking capacity that runs a few hours a day. The Three Mile Island restart — Unit 1 returning to service for Microsoft-Constellation — is the highest-profile nuclear revival, though small modular reactor (SMR) announcements have not converted to commercial orders at the pace the industry projected.
The Regional Trade-Offs
Each region trades different costs. Phoenix trades water: data center cooling consumes 50-200 million gallons a day for a 100 MW data center, and the Southwest's water stress makes that a hard limit on Phoenix's growth. Atlanta trades nuclear: Plant Vogtle Unit 3 and 4 expansions are powering the Georgia Power capacity additions that the Atlanta pipeline depends on. Quincy and Hillsboro trade environmental constraints — Pacific Northwest hydro is plentiful, but new transmission across the Cascades is contested.
Texas trades intermittency: ERCOT's wind and solar mix is volatile, and the data center load is concentrated enough that reliability standards are getting revisited. Northern Virginia trades transmission: PJM has a 4-5 year interconnection queue, and the capacity shortfall is real enough that Dominion Energy is commissioning gas peakers specifically to backstop hyperscaler loads.
The Efficiency Curve
The demand growth gets eaten by efficiency:
| Efficiency metric | 2022 | 2026 | 2030 (projected) |
|---|---|---|---|
| Training compute efficiency (FLOPs/W) | Baseline | 4-5x better | 10-20x better |
| Inference efficiency (FLOPs/W) | Baseline | 3-4x better | 8-15x better |
| Chip (NVIDIA Blackwell vs Hopper) | Hopper | Blackwell (2-3x perf/W) | Next-gen (further 2-3x) |
| Cooling PUE | 1.4-1.6 | 1.1-1.2 (liquid cooling) | 1.05-1.10 (immersion) |
Training compute efficiency at 4-5x better than 2022 means each frontier training run consumes the same electricity for 4-5x more useful compute. Inference efficiency at 3-4x better means production AI services are paying 30-40% less than they would have for the same quality. Hopper to Blackwell alone moved 17 percentage points of perf/W; the next generation adds another 2-3x.
PUE (power usage effectiveness, the ratio of total facility power to IT power) tells the cooling story: from air-cooled 1.4-1.6 to liquid-cooled 1.1-1.2 to immersion-cooled 1.05-1.10. Each step is a 20-30% reduction in cooling overhead. Together the efficiency curve eats about half of the demand growth before it lands on the grid.
Will It Break the Grid
Three concerns decide whether the grid breaks before it bends. Interconnection queues: average wait for grid connection is 4-5 years in major data center markets, per FERC. Capacity shortfalls: Northern Virginia, Phoenix, Atlanta, and Dallas face near-term capacity shortfalls that may force temporary moratoria on new data center construction. Transmission constraints: new high-voltage transmission build-out costs $30-50 billion 2025-2030 and runs into multi-year permitting cycles.
Locational marginal pricing (LMP) spikes during peak hours are becoming the visible cost: hyperscalers and utilities are signing demand-response contracts where data centers throttle or shift load during peak hours, which spreads the grid cost across more end users.
What to Watch Next
Five signals settle the rest of 2026 and the first half of 2027. The EIA monthly electricity data series is the cleanest reading on whether the doubling is on track. FERC Order 2023 interconnection reform is the regulatory lever that decides whether the 4-5 year queue compresses. AI hyperscaler power-purchase agreements disclose capacity decisions that move the demand forecast. Reliability standards for data center load participation in demand response reshape the cost stack. And small modular reactor announcements track whether the SMR industry converts to construction or stays at the announcement stage.
The two lines that decide whether the grid bends or breaks: transmission build-out completion and hyperscaler demand-response participation. Both are running late.
Read next
NVIDIA Physical AI: Robotaxi + Skild AI Demo 2026 tracks the GPU-and-robotics pipeline that drives the demand curve above, and OpenAI Expands Government AI Access + Cyber Defense tracks the federal procurement side of the same cycle.
