Every Watt · Tracked
The definitive tracker for AI energy infrastructure. Nuclear deals, grid strain by region, watts per query, water footprint, carbon gap analysis, datacenter capacity growth, and hyperscaler capex · sourced from utility filings, PPA disclosures, sustainability reports, and DOE data.
- Power demand growth90/100
- Grid strain39/100
- PPA price escalation54/100
- Nuclear premium53/100
- Water stress45/100
- Permit delays58/100
- AI power demand surging45% year-over-year growth in AI datacenter power consumption
- Permitting bottleneckAverage 28 month wait for new datacenter grid connections
- PPA prices escalatingAverage PPA prices up 18% year-over-year
- Nuclear capacity race17,770 MW of nuclear deals announced for AI
The Nuclear Ledger
8 energy deals · 18,230 MW total capacity · the race to power AI
| Deal | Type | Capacity |
|---|---|---|
 Microsoft Global renewable portfolio (solar + wind) | Renewable PPA | 10,500 MW |
 Meta Nuclear generation (new build or existing) | Nuclear New Build | 4,000 MW |
O Oracle SMR-powered data center campus | Small Modular Reactor | 1,000 MW |
 Amazon Susquehanna Nuclear Station | Nuclear Direct | 960 MW |
Microsoft Three Mile Island Unit 1 | Nuclear Restart | 835 MW |
 Google Multiple Kairos fluoride salt-cooled SMRs | Small Modular Reactor | 500 MW |
Amazon Xe-100 high-temperature gas-cooled SMRs | Small Modular Reactor | 320 MW |
Google Enhanced Geothermal Systems (EGS) | Geothermal | 115 MW |
Grid strain by region
10 regions · sorted by datacenter share of grid · most strained first
The largest concentration of data centers on Earth. Loudoun County alone hosts over 300 data centers, processing an estimated 70% of global internet traffic. Dominion Energy is struggling to keep up with power demand, requesting emergency natural gas peaker plants. Queue for new grid connections now extends 4+ years.
Amazon · 1400 MWMicrosoft · 1100 MWGoogle · 650 MWEurope's primary data center hub thanks to low corporate tax rates, English-speaking workforce, direct subsea cable connections to the US, and mild climate. But Ireland now faces a reckoning: data centers consume 21% of national electricity, and EirGrid has effectively paused new grid connections for large loads in the Dublin area. The country is racing to build offshore wind capacity to meet both DC and residential demand.
Microsoft · 350 MWAmazon · 300 MWGoogle · 250 MWAmong the cheapest power in the US thanks to Columbia River hydroelectric dams managed by the Bonneville Power Administration. Google built its first massive data center here in 2006. The region offers near-zero-carbon baseload power, but drought years threaten hydro output and cooling water availability.
Google · 500 MWMeta · 350 MWAmazon · 300 MWThe most important data center hub in Asia-Pacific and a critical node for undersea cable connectivity. Singapore imposed a moratorium on new data centers from 2019-2022 due to sustainability concerns, then selectively reopened with strict energy efficiency and renewable requirements. In a tropical climate with zero domestic fossil fuel or renewable resources, every watt is imported. The government now requires new DCs to achieve PUE below 1.3 and source green energy.
Google · 80 MWAmazon · 80 MWNear-perfect conditions for data centers: almost 100% carbon-free power (hydro + nuclear + wind), cold climate enabling free air cooling 8+ months per year, political stability, strong fiber connectivity, and EU data sovereignty. Meta's Lulea campus and Google's Hamina facility are among the most efficient in the world. The challenge is limited total grid capacity and long distances from major population centers.
Meta · 400 MWGoogle · 350 MWMicrosoft · 300 MWHome to the two largest internet exchanges in the world (AMS-IX and DE-CIX). The Netherlands and Frankfurt together form Europe's primary data center corridor. Amsterdam imposed a moratorium on new DCs in 2022 due to grid and land constraints. Frankfurt faces Germany's nuclear phaseout and high energy costs from the Energiewende transition. Both are pivoting to require proof of renewable energy sourcing for new permits.
Microsoft · 450 MWAmazon · 400 MWThe only major US grid operating independently from the national interconnect. ERCOT's deregulated market offers some of the cheapest power PPAs in the country, especially for solar and wind, but Winter Storm Uri (2021) exposed catastrophic fragility. Now the hottest market for new AI datacenter construction, with renewable PPAs below $30/MWh drawing hyperscalers.
Amazon · 600 MWMicrosoft · 500 MWGoogle · 400 MWThe American wind belt. Iowa generates over 60% of its electricity from wind, the highest share of any US state. Meta operates its largest global data center campus in Altoona. Microsoft, Google, and Apple all have major facilities. Wind PPAs here are among the cheapest in the world, and the flat terrain means excellent wind capacity factors.
Meta · 350 MWMicrosoft · 200 MWGoogle · 180 MWThe Gulf states are building sovereign AI infrastructure at extraordinary speed. The UAE's G42 and MBZUAI partnerships, Saudi Arabia's NEOM project, and Qatar's National AI Strategy represent a geopolitical bet that AI infrastructure is national security infrastructure. Abundant solar provides cheap daytime power, but the extreme heat (40+ C summers) means enormous cooling costs. Desalination provides cooling water but adds energy overhead.
Microsoft · 150 MWPost-Fukushima Japan is restarting nuclear reactors partly to power AI data centers. The government's GX (Green Transformation) strategy explicitly links nuclear energy to AI competitiveness. SoftBank, NTT, and Sakura Internet are building sovereign AI infrastructure, while Amazon and Google are expanding. Japan's challenge is the highest industrial electricity prices among developed nations.
Amazon · 400 MWGoogle · 300 MWWatts per query
The normalization that makes AI energy tangible
| Year | Projected GW |
|---|---|
| 2025 | 4.5 GW |
| 2026 | 8 GW |
| 2027 | 14 GW |
| 2028 | 22 GW |
| 2030 | 45 GW |
Water footprint
The hidden cost · 660B gallons per year globally
| Company | 2024 Water (M gal) | YoY |
|---|---|---|
| Microsoft | 7,800 | +22% |
| Google | 6,350 | +21% |
| Amazon | 5,200 | +18% |
| Meta | 4,200 | +15% |
A Apple | 1,800 | +8% |
Carbon gap
Claimed renewable % vs actual 24/7 carbon-free energy · the greenwashing dashboard
Datacenter growth
Global capacity trajectory · 2020 to 2030
| Year | Global GW | AI share % |
|---|---|---|
| 2020 | 20 GW | 8% |
| 2021 | 24 GW | 0% |
| 2022 | 28 GW | 15% |
| 2023 | 35 GW | 0% |
| 2024 | 42 GW | 30% |
| 2025 | 52 GW | 35% |
| 2026 | 68 GW | 40% |
| 2027 | 90 GW | 0% |
| 2028 | 120 GW | 50% |
| 2030 | 200 GW | 55% |
Capex race
The $283B hyperscaler arms race · 2025 capex commitments
MicrosoftAmazonGoogleMetaAll regions
10 datacenter power regions · each with a detail page
Northern VirginiaSaturatingThe largest concentration of data centers on Earth. Loudoun County alone hosts over 300 data centers, processing an estimated 70% of global internet traffic. Dominion Energy is struggling to keep up with power demand, requesting emergency natural gas peaker plants. Queue for new grid connections now extends 4+ years.
AmazonMicrosoftGoogle+5Oregon · Pacific NorthwestGrowingAmong the cheapest power in the US thanks to Columbia River hydroelectric dams managed by the Bonneville Power Administration. Google built its first massive data center here in 2006. The region offers near-zero-carbon baseload power, but drought years threaten hydro output and cooling water availability.
GoogleMetaAmazon+1Texas · ERCOTGrowingThe only major US grid operating independently from the national interconnect. ERCOT's deregulated market offers some of the cheapest power PPAs in the country, especially for solar and wind, but Winter Storm Uri (2021) exposed catastrophic fragility. Now the hottest market for new AI datacenter construction, with renewable PPAs below $30/MWh drawing hyperscalers.
AmazonMicrosoftGoogle+6Iowa · Midwest Wind CorridorGrowingThe American wind belt. Iowa generates over 60% of its electricity from wind, the highest share of any US state. Meta operates its largest global data center campus in Altoona. Microsoft, Google, and Apple all have major facilities. Wind PPAs here are among the cheapest in the world, and the flat terrain means excellent wind capacity factors.
MetaMicrosoftGoogle+1Near-perfect conditions for data centers: almost 100% carbon-free power (hydro + nuclear + wind), cold climate enabling free air cooling 8+ months per year, political stability, strong fiber connectivity, and EU data sovereignty. Meta's Lulea campus and Google's Hamina facility are among the most efficient in the world. The challenge is limited total grid capacity and long distances from major population centers.
MetaGoogleMicrosoft+1
IrelandConstrainedEurope's primary data center hub thanks to low corporate tax rates, English-speaking workforce, direct subsea cable connections to the US, and mild climate. But Ireland now faces a reckoning: data centers consume 21% of national electricity, and EirGrid has effectively paused new grid connections for large loads in the Dublin area. The country is racing to build offshore wind capacity to meet both DC and residential demand.
MicrosoftAmazonGoogle+2
SingaporeRestrictedThe most important data center hub in Asia-Pacific and a critical node for undersea cable connectivity. Singapore imposed a moratorium on new data centers from 2019-2022 due to sustainability concerns, then selectively reopened with strict energy efficiency and renewable requirements. In a tropical climate with zero domestic fossil fuel or renewable resources, every watt is imported. The government now requires new DCs to achieve PUE below 1.3 and source green energy.
GoogleAmazon+1The Gulf states are building sovereign AI infrastructure at extraordinary speed. The UAE's G42 and MBZUAI partnerships, Saudi Arabia's NEOM project, and Qatar's National AI Strategy represent a geopolitical bet that AI infrastructure is national security infrastructure. Abundant solar provides cheap daytime power, but the extreme heat (40+ C summers) means enormous cooling costs. Desalination provides cooling water but adds energy overhead.
Microsoft
JapanGrowingPost-Fukushima Japan is restarting nuclear reactors partly to power AI data centers. The government's GX (Green Transformation) strategy explicitly links nuclear energy to AI competitiveness. SoftBank, NTT, and Sakura Internet are building sovereign AI infrastructure, while Amazon and Google are expanding. Japan's challenge is the highest industrial electricity prices among developed nations.
AmazonGoogleHome to the two largest internet exchanges in the world (AMS-IX and DE-CIX). The Netherlands and Frankfurt together form Europe's primary data center corridor. Amsterdam imposed a moratorium on new DCs in 2022 due to grid and land constraints. Frankfurt faces Germany's nuclear phaseout and high energy costs from the Energiewende transition. Both are pivoting to require proof of renewable energy sourcing for new permits.
MicrosoftAmazonFrequently asked
Pulled from the live dataset · schema-ready for AEO
How much power does AI consume globally?
AI data centers currently consume approximately 18 GW of power globally, growing at 45% year-over-year. Total global datacenter power capacity is 52 GW, with AI workloads accounting for a growing share. By 2030, AI inference alone is projected to require 45 GW.
How much energy does a single ChatGPT query use?
A single ChatGPT query consumes approximately 2.9 watt-hours, roughly 10x the energy of a Google search (0.3 Wh). AI image generation uses about 12 Wh per image, and AI video generation uses approximately 50 Wh per minute. With 1.2 billion daily AI queries, this adds up to 4.5 GW of continuous inference power.
What is the nuclear renaissance for AI?
8 major nuclear energy deals have been announced to power AI infrastructure, totaling 18,230 MW of capacity. Key deals include Microsoft restarting Three Mile Island (835 MW), Amazon's direct nuclear connection at Susquehanna (960 MW), and Google's pioneering SMR deal with Kairos Power (500 MW). Nuclear provides 24/7 carbon-free baseload power that intermittent renewables cannot match.
How much water do AI data centers consume?
The top 5 hyperscalers consumed approximately 25,350 million gallons of water in 2024 for cooling. Microsoft led at 7,800 million gallons (+22% YoY). Evaporative cooling uses about 350,000 gallons per MW per day. Liquid cooling systems can reduce water consumption by 40%.
What is the Energy Price Index (EPI)?
The EPI measures overall energy pressure on AI infrastructure on a 0-100 scale. Currently at 59 (elevated). It combines six factors: power demand growth (25%), grid strain (20%), PPA price escalation (20%), nuclear premium (15%), water stress (10%), and permitting delays (10%). Higher scores indicate greater infrastructure strain.
Are hyperscalers really 100% renewable?
Most hyperscalers claim 100% renewable energy matching, but this uses annual Renewable Energy Certificates (RECs) — buying enough credits to offset annual consumption. Actual 24/7 carbon-free energy (CFE) is significantly lower. Google leads at 64% actual CFE, while Microsoft is at 52%. The gap between claimed and actual ranges from 28 to 48 percentage points.
Which regions have the cheapest datacenter power?
Iowa/Midwest offers the cheapest PPAs at ~$22/MWh thanks to abundant wind. The Nordics follow at ~$25/MWh with hydro + nuclear. Oregon's hydro-powered grid delivers ~$28/MWh. Texas ERCOT offers ~$32/MWh for solar/wind PPAs. On the expensive end, Singapore ($110/MWh) and Japan ($85/MWh) have the highest costs.
How fast is datacenter construction growing?
Global datacenter capacity was ~52 GW in 2025, projected to reach 200 GW by 2030. That is a ~4x increase in 5 years. Currently 11.7 GW is under construction. The top 5 hyperscalers are investing $283B in capex in 2025 alone, mostly on AI datacenter infrastructure.
See also
Keep exploring the compute graph