Executive Summary
AI-driven data center power demand is on track to exceed 50 GW in the United States alone by 2028, a figure equivalent to the total electricity generation capacity of countries like Poland or Australia. This surge has triggered a wholesale restructuring of American energy infrastructure: nuclear plants scheduled for retirement are being restarted, natural gas capacity is expanding despite climate commitments, and utility interconnection queues have ballooned to multi-year backlogs. The hyperscalers — Microsoft, Google, Amazon, and Meta — are collectively committing over $200 billion in capital expenditure on data center infrastructure through 2027, creating the largest private infrastructure buildout since the transcontinental railroad.
Current State
US data center electricity consumption reached an estimated 28–32 GW of peak demand in early 2026, up from approximately 20 GW at the start of 2024. The Department of Energy’s most recent assessment projects this figure will reach 50+ GW by 2028 under moderate growth scenarios, and could exceed 80 GW by 2030 if AI training and inference workloads scale at current rates.
To contextualize these numbers: the entire US electrical grid has a total generating capacity of approximately 1,300 GW. AI data centers currently consume roughly 2.5% of total US electricity generation. By 2030, that share could rise to 6–8% — a figure that has prompted FERC commissioners, state utility regulators, and grid operators to issue formal warnings about reliability risks.
Data center power demand by major market (estimated peak MW, early 2026):
| Market | Current Demand | Under Construction | Planned |
|---|---|---|---|
| Northern Virginia (NOVA) | 5,200 MW | 3,800 MW | 6,500 MW |
| Dallas-Fort Worth | 2,800 MW | 2,100 MW | 4,200 MW |
| Phoenix / Mesa | 2,100 MW | 1,800 MW | 3,500 MW |
| Columbus / Central Ohio | 1,400 MW | 1,200 MW | 2,800 MW |
| Chicago | 1,200 MW | 900 MW | 2,100 MW |
| Atlanta | 1,000 MW | 800 MW | 1,600 MW |
| Portland / Hillsboro | 900 MW | 700 MW | 1,400 MW |
Northern Virginia remains the single largest data center market globally, with Dominion Energy’s service territory hosting more than 5 GW of connected data center load. Dominion’s interconnection queue now contains over 30 GW of requested capacity, representing a pipeline that would take a decade or more to fulfill under current grid expansion rates.
Key power procurement deals completed since 2024:
| Deal | Parties | Capacity | Structure |
|---|---|---|---|
| Three Mile Island restart | Microsoft / Constellation Energy | 835 MW | 20-year PPA |
| Susquehanna nuclear | Amazon / Talen Energy | 960 MW | Direct procurement |
| Vogtle expansion | Google / Southern Company | 600 MW | Clean energy PPA |
| Palo Verde allocation | Meta / Arizona Public Service | 500 MW | Behind-the-meter |
| SMR development | Oracle / NuScale | 300 MW (planned) | Joint development |
| Natural gas campus | CoreWeave / Calpine | 1,200 MW | Co-located generation |
| Wind + battery | Microsoft / AES | 900 MW | Hybrid PPA |
| Geothermal pilot | Google / Fervo Energy | 150 MW | EGS demonstration |
Key Dynamics
The Interconnection Queue Crisis. The single largest bottleneck for AI data center expansion is not chip supply, construction labor, or capital — it is grid interconnection. According to Lawrence Berkeley National Laboratory, approximately 2,600 GW of generation and storage capacity sits in US interconnection queues as of early 2026, with average wait times exceeding 5 years. Data center developers now routinely purchase land with existing grid connections (including decommissioned industrial sites and retired power plants) at premium valuations simply to bypass the queue.
Nuclear Renaissance. AI’s need for baseload, carbon-free power has revived the American nuclear industry in ways that seemed inconceivable five years ago. The Constellation Energy deal to restart Three Mile Island Unit 1 (the undamaged reactor, not the one involved in the 1979 accident) was the first US nuclear restart of a fully decommissioned plant. At least 12 similar deals are in various stages of negotiation, including potential restarts at Palisades in Michigan and Duane Arnold in Iowa. Small modular reactor (SMR) development has accelerated, with NuScale, Oklo, and Kairos Power all targeting initial deployments by 2028–2030.
Natural Gas Expansion vs. Climate Commitments. Every major hyperscaler has committed to net-zero carbon emissions by 2030, yet the fastest path to new power capacity remains natural gas. Microsoft, Google, Amazon, and Meta have all signed natural gas procurement agreements or invested in gas-fired generation since 2024, creating tension with their stated climate goals. The International Energy Agency estimates that AI-related natural gas demand could add 50–80 million tons of CO2 annually to US emissions by 2030 if renewables and nuclear cannot scale fast enough.
Behind-the-Meter Generation. A growing trend sees data center operators building dedicated power generation on-site, effectively bypassing the public grid entirely. CoreWeave’s partnership with Calpine to build co-located natural gas generation at its Texas data center campus exemplifies this approach. Behind-the-meter generation avoids interconnection delays and transmission losses but raises concerns among utility regulators about grid cost allocation — remaining grid customers may bear a disproportionate share of transmission infrastructure costs as large loads exit the system.
Water Consumption. AI data centers consume enormous quantities of water for cooling. A single large facility can consume 5–10 million gallons per day, equivalent to the water usage of a small city. This has generated community opposition in water-stressed regions, particularly in the American Southwest. Meta’s data center proposals in Mesa, Arizona, and Google’s expansion in The Dalles, Oregon, have both faced significant local resistance over water use concerns.
Who’s Involved
Microsoft has the largest committed data center capital expenditure program, announcing $80 billion in planned spending through fiscal year 2027. The Constellation Energy deal for Three Mile Island power was negotiated personally by Brad Smith (Vice Chair and President) and represents Microsoft’s largest single energy procurement commitment. Microsoft’s strategy emphasizes nuclear and renewable power to maintain its carbon-negative pledge, though it has also signed natural gas agreements.
Google has committed approximately $50 billion in data center capex through 2027 and has pursued the most diverse energy procurement strategy, including utility-scale solar, wind, battery storage, enhanced geothermal systems (through its investment in Fervo Energy), and nuclear power. Ruth Porat (President and CIO, Alphabet and Google) has been the primary executive driving energy strategy. Google’s carbon footprint increased 48% from 2019 to 2024, largely due to data center expansion, prompting internal debate about the sustainability of its growth trajectory.
Amazon Web Services is both the largest cloud provider and the largest corporate buyer of renewable energy globally, with over 25 GW of clean energy capacity contracted. Amazon’s direct procurement deal with Talen Energy for Susquehanna nuclear power drew criticism from grid operators who argued it could raise electricity costs for other consumers. Matt Garman (CEO, AWS) has indicated that energy procurement is now a “top three” strategic priority.
Meta has committed to roughly $40 billion in data center capex for 2025 alone, the largest single-year infrastructure spend in the company’s history. Meta’s approach emphasizes massive single-campus facilities (its Richland Parish, Louisiana campus is planned at over 2 GW) with a mix of renewable PPAs and grid power.
CoreWeave has emerged as the largest independent AI infrastructure provider, with a data center portfolio exceeding 5 GW of planned capacity. Unlike the hyperscalers, CoreWeave focuses exclusively on GPU compute and has been more aggressive in securing natural gas and behind-the-meter generation.
Constellation Energy became the most prominent beneficiary of AI energy demand when its stock price roughly doubled following the Three Mile Island deal announcement. As the largest US nuclear fleet operator (with approximately 21 GW of nuclear capacity), Constellation is uniquely positioned to supply the baseload power that AI data centers require.
What the Data Shows
Capital expenditure on data center infrastructure by the four largest hyperscalers (Microsoft, Google, Amazon, Meta) totaled approximately $190 billion in 2025, up from $120 billion in 2024. Analyst consensus estimates for 2026 exceed $230 billion. This spending is now the single largest line item in these companies’ capital allocation, exceeding all other infrastructure categories combined.
Electricity price impacts are already visible in major data center markets. Wholesale electricity prices in PJM Interconnection (which covers Northern Virginia and much of the eastern US) rose approximately 25% from 2023 to 2025, driven in part by data center load growth. Dominion Energy has filed for rate increases that would pass through a portion of grid expansion costs to all ratepayers in its Virginia service territory.
Renewable energy procurement by tech companies reached record levels, with the five largest corporate renewable energy buyers (Google, Amazon, Microsoft, Meta, Apple) collectively contracting over 70 GW of clean energy capacity. However, the gap between contracted capacity and actual delivered clean energy remains significant — many renewable PPAs take 3–5 years from signing to commercial operation.
GPU power consumption trends show that while individual chip efficiency (performance per watt) has improved with each generation, the aggregate power draw of AI clusters has increased faster. NVIDIA’s B200 GPU draws approximately 1,000 watts under full load, compared to 700 watts for the H100. A single rack of B200 GPUs can draw 120 kW, meaning a 100,000-GPU training cluster requires over 100 MW of continuous power — before cooling, networking, and storage overhead.
Efficiency metrics:
| Metric | Industry Avg | Best-in-Class | Target 2028 |
|---|---|---|---|
| PUE (Power Usage Effectiveness) | 1.30 | 1.06 | 1.10 |
| WUE (Water Usage Effectiveness) | 1.8 L/kWh | 0.2 L/kWh | 0.5 L/kWh |
| Carbon Intensity | 350 gCO2/kWh | 25 gCO2/kWh | 50 gCO2/kWh |
| Server Utilization | 40% | 70% | 60% |
Outlook
The central tension in AI energy demand is timing: AI companies need power now, but new generation and transmission capacity takes 5–10 years to build. This mismatch will persist through at least 2030, creating sustained upward pressure on electricity prices, increasing grid reliability concerns, and generating significant regulatory and political friction.
Several developments could alter the trajectory. Advances in model efficiency (such as mixture-of-experts architectures and quantization) could reduce inference energy costs by 50–80% per query. Successful deployment of small modular reactors by 2029–2030 would add a new source of scalable baseload power. Conversely, a breakthrough in artificial general intelligence could trigger demand growth that dwarfs current projections.
FERC and state regulators are beginning to impose conditions on large data center interconnections, including requirements for on-site generation, energy storage, and demand response participation. These regulatory frameworks will shape the geography and economics of AI infrastructure deployment for the next decade.
The energy constraint is increasingly recognized as the binding limit on AI scaling. Jensen Huang (NVIDIA CEO) has stated publicly that energy availability — not chip manufacturing — will determine the pace of AI advancement. This view is shared by the CEOs of all four major hyperscalers and has shifted AI infrastructure planning from a technology problem to an energy procurement and regulatory challenge.
Frequently Asked Questions
How much electricity does AI actually consume? AI data centers in the United States currently consume an estimated 28–32 GW of peak power, representing approximately 2.5% of total US electricity generation. This figure is projected to exceed 50 GW by 2028 under moderate growth scenarios. To put this in perspective, 50 GW is roughly equivalent to the total electricity generation capacity of a country like Poland. A single large AI training run on a 100,000-GPU cluster can consume 100+ MW of continuous power for months — comparable to a small city’s total electricity demand.
Why are tech companies buying nuclear power plants? Nuclear power provides what AI data centers need most: large-scale, reliable, carbon-free baseload electricity that operates 24/7 regardless of weather conditions. Solar and wind are intermittent and require massive battery storage to provide constant power. Natural gas is reliable but produces carbon emissions. Nuclear uniquely satisfies both reliability and sustainability requirements. The economics have also shifted: building new nuclear is expensive and slow, but restarting or extending existing plants is far cheaper and faster, making nuclear restarts an attractive bridge while new generation capacity is built.
Will AI cause electricity prices to rise? In major data center markets, this is already happening. Wholesale electricity prices in the PJM Interconnection region (covering Northern Virginia and the eastern US) rose approximately 25% from 2023 to 2025, driven partly by data center demand growth. Utility regulators in Virginia and Georgia have approved or are considering rate increases that allocate grid expansion costs across all ratepayers. The net impact on residential electricity prices is projected at 2–5% nationally by 2028, with higher impacts in data center-dense regions. However, some economists argue that the tax revenue and economic development from data centers may offset ratepayer costs in the long run.
Can renewable energy alone power AI’s growth? Not at current deployment rates. While the major tech companies have contracted over 70 GW of renewable energy capacity, much of it will not be operational for 3–5 years. Renewable energy also faces intermittency challenges — AI data centers typically require 99.999% uptime, which solar and wind alone cannot provide without massive battery storage investments. The practical answer is that AI will require a mix of renewables, nuclear, and natural gas through at least 2035, with the proportions shifting toward cleaner sources over time as storage technology matures and new nuclear capacity comes online.
What happens if the grid cannot keep up with AI demand? Grid constraints are already forcing AI companies to delay or relocate projects. Several planned data center campuses in Northern Virginia have been pushed back 2–3 years due to transmission capacity limitations. If grid expansion continues to lag demand, the likely outcomes include geographic redistribution of data centers to less constrained markets (such as the Midwest and Southeast US), increased adoption of behind-the-meter generation (on-site power plants), higher electricity costs for all consumers in data center-dense regions, and potential regulatory limits on new data center interconnections. Some analysts have suggested that energy constraints could effectively cap the rate of AI capability improvement by limiting the scale of training runs.