# AI Power Demand and Energy Infrastructure: The Numbers, the Bottlenecks, and the Investment Thesis
Data as of August 2026. Sources: EIA, IEA, Lawrence Berkeley National Laboratory, NERC, company filings. Educational purposes only.---
The Scale of the Problem: How Much Power Does AI Actually Use?
U.S. Data Center Electricity Demand Forecast Comparison
| Source | Publication | Baseline | Forecasted Horizon | Projected TWh |
|---|---|---|---|---|
| LBNL | 2024 Report | 176 TWh (2023 US) | 2028 | 325 – 580 TWh (US) |
| [IEA]([energy](/guides/ai-power-demand-energy-infrastructure-stocks)) | 2025 Energy & AI | 415 TWh (2024 Global) | 2030 | 945 TWh (Global) |
| [EIA](/guides/ai-power-demand-energy-infrastructure-stocks) | Annual Energy Outlook | ~7% of commercial | 2050 | 22% - 33% of commercial |
- U.S. data centers consumed approximately 176 TWh (4.4% of total U.S. electricity) in 2023
- Projected to reach 325–580 TWh (6.7%–12.0% of total U.S. electricity) by 2028
- The wide range reflects uncertainty in AI deployment pace and efficiency improvements
- Source: [LBNL 2024 Data Center Report](https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report)
- Estimated 2022 global data center electricity consumption: ≈ 460 TWh
- Forecasted to exceed 1,000 TWh by 2026
- Source: [IEA Electricity 2024](https://www.iea.org/reports/electricity-2024)
- Revised 2024 global data center electricity consumption: ≈ 415 TWh
- Base Case Forecast: Projected to reach ≈ 945 TWh by 2030
- Methodological Evolution: The 2025 modelling incorporates more precise AI accelerator utilization rates and efficiency offsets, adjusting the baseline and extending the forecast horizon to 2030.
- Source: [IEA Energy and AI Report](https://www.iea.org/reports/energy-and-ai)
- Data center servers accounted for ~7% of commercial sector electricity in 2025
- Projected to grow to 22%–33% of commercial building electricity use by 2050
- Source: [EIA Annual Energy Outlook](https://www.eia.gov/outlooks/aeo/)
---
The Power Density Revolution: Why AI Data Centers Are Different
Traditional [cloud](/guides/ai-hyperscaler-capex-infrastructure-winners) data centers operate at 5–20 kilowatts per server rack. A standard 40,000 sq ft data center might draw 5–20 megawatts total.
An AI training data center deploying [NVIDIA](/stocks/NVDA) H100/H200 or Blackwell GPU clusters operates at 400–500+ kW per rack — roughly 25–50x the density. A single AI training cluster of 100,000 GPUs (the scale that [Meta](/stocks/META), [Microsoft](/stocks/MSFT), and Google are building) can require ~300–500 megawatts of dedicated power capacity.
For context: 500 MW is roughly the output of a large natural gas peaker plant, or half the output of a nuclear reactor. Dedicating this capacity to a single AI cluster means either:- Building new generation capacity specifically for that cluster
- Securing dedicated long-term offtake from existing generators
This is why the AI buildout has transformed from an IT capital expenditure story into an energy and real estate story.
---
Grid Infrastructure: The Interconnection Bottleneck
Connecting new large loads to the electrical grid is not as simple as signing a power contract. New data centers require interconnection studies and grid upgrades that take 3–7 years under current processes.
North American Electric Reliability Corporation (NERC) — the government-sanctioned grid reliability organization — identified rapid data center expansion as a primary risk in its 2025 Long-Term Reliability Assessment (LTRA). NERC website: [https://www.nerc.com](https://www.nerc.com) PJM Interconnection (serving 13 states from Illinois to New Jersey, the largest US grid operator) published its Long-Term Load Forecast documenting multi-gigawatt load additions from data center clusters, particularly in Northern Virginia (the world's largest data center market). PJM website: [https://www.pjm.com](https://www.pjm.com) MISO (Midcontinent Independent System Operator) serves the Midwest and has flagged data center growth in Texas, Indiana, and Illinois corridors as a resource adequacy challenge. MISO website: [https://www.misoenergy.org](https://www.misoenergy.org) FERC (Federal Energy Regulatory Commission) has issued new rules on large load interconnection to address the queue backlog — thousands of projects awaiting interconnection studies. FERC Portal: [https://www.ferc.gov](https://www.ferc.gov)The bottleneck is not generation capacity per se — it is the grid interconnection queue and transmission upgrade timeline. This is why hyperscalers are increasingly pursuing nuclear power agreements: nuclear plants are already connected to the grid and can supply dedicated output under a new power purchase agreement without waiting for new interconnection studies.
---
The Nuclear Renaissance: Why Tech Companies Are Buying Nuclear Power
Three major hyperscaler nuclear agreements underscore the structural shift:
Constellation Energy + Microsoft (Three Mile Island, Crane Clean Energy Center)
- Deal: 20-year Power Purchase Agreement for 835 MW of nuclear power
- Source: Restart of Three Mile Island Unit 1 (closed in 2019), renamed the Crane Clean Energy Center
- Official announcement: [Constellation Energy press release](https://www.constellationenergy.com/news/news-releases/2024/constellation-to-launch-crane-clean-energy-center--restoring-jobs.html)
- Significance: First-ever commercial nuclear plant restart in the US, driven by AI data center demand. The 835 MW goes entirely to Microsoft under a long-term contract.
[Amazon](/stocks/AMZN) + X-energy / Talen Energy
- Deal: Agreement with X-energy to bring >5 GW of Small Modular Reactor (SMR) nuclear capacity online by 2039; PPA with Talen Energy for up to 1,920 MW from the Susquehanna nuclear plant in Pennsylvania
- Official source: [Amazon corporate announcement](https://about.amazon.com/news/about-amazon/news/company-news/amazon-signs-agreements-for-innovative-nuclear-energy-projects-to-address-growing-energy-demands)
- Significance: Amazon is hedging between existing nuclear (near-term capacity) and SMR development (long-term capacity), mirroring its diversified cloud infrastructure strategy
Google + Kairos Power
- Deal: Master Plant Development Agreement for 500 MW of advanced nuclear reactor capacity by 2035, starting with a 50 MW Hermes 2 reactor pilot in Oak Ridge, Tennessee
- Official source: [Google blog announcement](https://blog.google/outreach-initiatives/sustainability/google-kairos-power-nuclear-energy/)
- Significance: Google is betting specifically on advanced fission reactors (fluoride salt-cooled) — a different technology path from conventional light water reactors or SMRs being pursued by other vendors
---
The Energy Infrastructure Investment Thesis
The AI power demand story creates several distinct investment angles:
Utilities and Independent Power Producers
Constellation Energy (CEG):- The nuclear PPA with Microsoft dramatically improved the value of Constellation's existing nuclear fleet
- Nuclear plants that were economically marginal now have secured long-term counterparties
- [Constellation IR](https://www.constellationenergy.com/investors)
- FPL (Florida Power & Light) utility CapEx plan of $90–100 billion through 2032
- NextEra Energy Resources targeting 15–30 GW of capacity development by 2035 for large data center loads
- 36.5–46.5 GW of renewables and storage planned through 2027
- [NextEra IR](https://investor.nexteraenergy.com)
- Operates Comanche Peak Nuclear Power Plant (2,400 MW, license extended through 2053 by NRC)
- Acquired Energy Harbor Corp in March 2024, adding ~4,000 MW of nuclear capacity (Perry, Davis-Besse, Beaver Valley plants)
- Total competitive nuclear fleet: >6,400 MW — the largest competitive nuclear portfolio in the US
- [Vistra IR](https://investor.vistracorp.com)
The Grid Upgrade Chain
The power delivered to data centers must traverse the transmission and distribution grid. As grid operators upgrade substations, add transmission lines, and deploy grid-scale battery storage to manage AI load variability, the following categories benefit:
- Transmission line manufacturers: Companies making high-voltage conductors and cables
- Transformer manufacturers: Large power transformers face 2–4 year lead times (supply constraint)
- Grid-scale battery storage: For managing AI load intermittency (AI training runs are not continuous)
- Engineering, Procurement & Construction (EPC) firms: Building the data center physical plants
---
The Sustainability Tension
Every major hyperscaler has made public commitments to carbon-neutral or net-zero operations. Google, Microsoft, and Amazon all have 100% renewable energy commitments for their data centers. However, the AI buildout creates a fundamental tension:
- The timing mismatch: New solar and wind capacity takes time to permit and build. The AI compute buildout is happening now and drawing from the existing fossil-fuel-heavy grid in the interim.
- The intermittency problem: AI training workloads run continuously, 24/7. Renewable energy (solar, wind) is intermittent. Nuclear power (24/7, zero-carbon) solves the intermittency problem for around-the-clock carbon-free power — which is why hyperscalers are pursuing nuclear agreements despite traditional commitments to renewable-only procurement.
- The grid attribution complexity: When Microsoft says "100% renewable energy," this is typically achieved through Renewable Energy Certificates (RECs) — purchased credits representing renewable generation that may not be temporally or geographically aligned with their actual consumption. The emerging "24/7 clean energy" standard (which Google pioneered) requires renewable energy to match consumption in each hour of each day, in the same grid region — a much stricter and more expensive standard.
---
Catalayer Analysis: What the Market Is Mispricing
Non-obvious observation #1: The transformer shortage is the most underappreciated bottleneck.The US electricity system requires large power transformers (LPTs) to step up/step down voltage across the transmission grid. LPT lead times have extended to 2–4 years. This supply constraint limits how fast new generation capacity can be connected to the grid — meaning even if a hyperscaler signs a nuclear PPA today, it may take 3–5 years before that power can reliably reach their data center.
Non-obvious observation #2: Northern Virginia dominates — and is reaching limits.Northern Virginia (Loudoun County, the "Data Center Alley") is the world's largest data center market. It hosts a disproportionate share of cloud and internet infrastructure. Multiple [utilities](/guides/ai-power-demand-energy-infrastructure-stocks) have warned that new data center load additions in Loudoun County are approaching physical grid limits, with no clear near-term resolution. This geographic constraint is driving the geographic diversification of data center development to Ohio, Indiana, Texas, and international markets.
Non-obvious observation #3: The energy price exposure is a double-edged sword for hyperscalers.AI data center power consumption is so large that hyperscalers are moving from passive electricity consumers to counterparties in the energy market. Long-term nuclear PPAs at fixed prices are hedges against rising power prices — but they also represent long-term fixed obligations that could become expensive if AI efficiency improves dramatically (requiring less power for the same compute output). This is a real optionality question embedded in 20-year nuclear power contracts.
---
Related Guides and Content
- Guide: [The Economics of AI Hyperscalers: CapEx Trends and Infrastructure Winners](/guides/ai-hyperscaler-capex-infrastructure-winners) — The demand side generating the power requirement
- Guide: [AI Semiconductor Value Chain](/guides/ai-semiconductor-value-chain-analysis) — The hardware requiring the power
- Topic: [Energy](/guides/ai-power-demand-energy-infrastructure-stocks) — Real-time energy sector news
- Topic: [Utilities](/guides/ai-power-demand-energy-infrastructure-stocks) — Utility stock coverage
---
Primary Sources
- [Lawrence Berkeley National Laboratory: 2024 Data Center Energy Usage Report](https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report)
- [IEA: Energy and AI Report (April 2025)](https://www.iea.org/reports/energy-and-ai)
- [EIA Annual Energy Outlook](https://www.eia.gov/outlooks/aeo/)
- [NERC: Long-Term Reliability Assessment](https://www.nerc.com)
- [PJM Interconnection](https://www.pjm.com)
- [MISO](https://www.misoenergy.org)
- [FERC — Federal Energy Regulatory Commission](https://www.ferc.gov)
- [Constellation Energy — Crane Clean Energy Center Announcement](https://www.constellationenergy.com/news/news-releases/2024/constellation-to-launch-crane-clean-energy-center--restoring-jobs.html)
- [Amazon — Nuclear Energy Agreements](https://about.amazon.com/news/about-amazon/news/company-news/amazon-signs-agreements-for-innovative-nuclear-energy-projects-to-address-growing-energy-demands)
- [Google — Kairos Power Nuclear Agreement](https://blog.google/outreach-initiatives/sustainability/google-kairos-power-nuclear-energy/)
- [NextEra Energy Investor Relations](https://investor.nexteraenergy.com)
- [Vistra Corp Investor Relations](https://investor.vistracorp.com)
- [DOE Electricity Demand Resource Hub](https://www.energy.gov/electricitydemand)
- [EPA eGRID Database](https://www.epa.gov/egrid)
---
Disclaimer: This guide is for informational and educational purposes only. Energy infrastructure investing involves regulatory, commodity, and long-duration capital risks. This is not investment advice.