Why This Matters
The AI industry’s constraint has shifted from chips to power. A single frontier model training run now consumes as much electricity as a small city for months. NVIDIA’s next-generation GB200 NVL72 racks draw over 120 kW each — a fully loaded AI data center with 10,000 racks requires over a gigawatt of continuous power, equivalent to a nuclear power plant running at full capacity. The total power demand from announced AI data center projects exceeds 50 GW globally, rivaling the electricity consumption of countries like South Korea or the United Kingdom.
This has created a land rush for power that is reshaping energy markets worldwide. Hyperscalers are restarting mothballed nuclear plants, signing 20-year power purchase agreements worth billions of dollars, building dedicated natural gas generation facilities, and competing with each other for grid interconnection rights that can take years to secure. The AI energy buildout is the largest single source of new electricity demand in the United States since the post-World War II industrialization boom.
The geopolitical dimension is equally significant. The United States has announced roughly 38 GW of AI-dedicated power capacity compared to China’s approximately 8 GW. This 5:1 advantage reflects both the scale of US private capital deployment and China’s constraints from US chip export restrictions that limit the GPU density of Chinese data centers. Whether this gap persists will be determined as much by power availability and grid infrastructure as by chip supply.
This tracker monitors the global buildout: who’s building what, how many gigawatts are coming online, and how the US and China compare.
The GW Scoreboard
Current estimates of AI-dedicated power capacity (announced + under construction):
| Country | Announced GW | Under Construction | Online | Notes |
|---|---|---|---|---|
| United States | ~38 GW | ~12 GW | ~5 GW | Stargate, Colossus, hyperscaler expansions |
| China | ~8 GW | ~3 GW | ~2 GW | State-backed, reporting opacity |
| UAE / Saudi | ~5 GW | ~1.5 GW | <0.5 GW | NEOM, Abu Dhabi clusters |
| Europe | ~3 GW | ~1 GW | <1 GW | Nordics, Ireland, UK |
| Other | ~2 GW | <1 GW | <0.5 GW | Japan, India, Singapore |
Global total: ~56 GW announced, with roughly 18 GW actively under construction as of mid-2026.
Featured Projects
xAI — Colossus 1 (Memphis, TN)
The original Colossus supercomputer cluster. Built in under 122 days on the site of a former Electrolux factory. Currently the world’s largest single AI training facility.
- Address: 3231 Riverport Rd, Memphis, TN 38109
- Power: ~150 MW operational, expanding to 300 MW
- GPUs: 100,000 NVIDIA H100s (Phase 1), expanding to 200,000
- Status: Operational, Phase 2 expansion underway
- What makes it notable: Speed of construction. Musk’s team went from empty lot to operational in 122 days, a pace that shocked the industry.
xAI — Macrohard / Colossus 2 (Memphis, TN)
The massive expansion facility in Whitehaven, Memphis. “MACROHARD” is painted on the roof in letters visible from space — a dig at Microsoft. Designed to house nearly a million NVIDIA chips.
- Address: 5420 Tulane Rd, Memphis, TN 38109
- Power: Targeting 1+ GW
- GPUs: Scaling to ~1,000,000 next-gen GPUs (B200/B300)
- Land: ~100 acres across three parcels
- Status: Active, powering up
- What makes it notable: If fully built out at announced scale, it would be the single largest AI facility on Earth. The roof logo is visible on satellite imagery.
Stargate (Abilene, TX)
The OpenAI–SoftBank–Oracle–MGX joint venture. The largest announced AI infrastructure project by dollar value. A 4-million-square-foot complex on over 1,000 acres on the outskirts of Abilene.
- Location: Outskirts of Abilene, TX (~180 miles west of Dallas)
- Investment: $500B over 4 years (announced Jan 2025)
- Power: Targeting 5+ GW across multiple phases; first campus scaling past 1 GW
- Status: First building operational (June 2025), second nearly complete
- Partners: OpenAI, SoftBank, Oracle, MGX, NVIDIA
Microsoft (Multiple US Sites)
Microsoft’s massive global data center expansion to support Azure AI and OpenAI workloads.
- Capex: $80B+ committed for FY2025
- Power: 3+ GW in new commitments
- Nuclear deals: Three Mile Island restart (Constellation Energy), 20-year PPA
- Key sites: Northern Virginia, Phoenix, Wisconsin, Sweden, UK
Meta (Multiple US Sites)
Meta’s AI infrastructure buildout for Llama training and inference at scale.
- Capex: $60-65B committed for 2025
- Power: 2+ GW in new commitments
- Key sites: Richland Parish (Louisiana), Temple (TX), expansion across existing fleet
- Status: Richland Parish 4 GW campus in planning
Google (Multiple Global Sites)
Google’s data center expansion for Gemini training and Cloud AI inference.
- Capex: $75B+ committed for 2025
- Nuclear deals: Kairos Power (small modular reactors), 7 reactors by 2035
- Power: 3+ GW in new commitments
- Key sites: The Dalles (OR), Mayes County (OK), Waukesha (WI), international expansions
Amazon / AWS (Multiple Global Sites)
AWS AI infrastructure buildout, including custom Trainium chips alongside NVIDIA GPUs.
- Capex: $100B+ committed for 2025
- Nuclear deals: Talen Energy (Susquehanna), small modular reactor partnerships
- Power: 4+ GW in new commitments
- Key sites: Northern Virginia, Oregon, Ohio, Mississippi
US vs China: The Power Gap
The AI power race has a stark geographic imbalance:
United States advantages:
- Massive private capital deployment ($300B+ capex from Big Tech in 2025-2026)
- Abundant natural gas for near-term generation
- Nuclear restart and SMR pipeline (5+ deals signed)
- Permitting reform gaining bipartisan support
- Grid capacity headroom in sunbelt states
China advantages:
- State-directed buildout with fewer permitting barriers
- Integrated power-to-chip supply chain ambitions
- Rapid renewable deployment (solar manufacturing dominance)
- Lower construction costs per MW
China constraints:
- US chip export restrictions limit GPU access (H20 banned, older chips restricted)
- Total AI-dedicated power capacity lags US by 4-5x
- Less transparency in reporting makes tracking difficult
- Energy mix still heavily coal-dependent (raises ESG questions for partners)
The gap is widening, not closing. US hyperscalers alone announced more GW of AI power in Q1 2026 than China’s entire announced pipeline. The question is whether permitting and grid infrastructure can keep up with ambition.
Energy Source Mix
How the AI industry plans to power these facilities:
| Source | Share of New AI Power | Trend |
|---|---|---|
| Natural Gas | ~55% | Growing (fastest to build) |
| Nuclear (existing restarts) | ~15% | Growing (3 restart deals signed) |
| Solar + Storage | ~12% | Growing (Meta, Google) |
| Grid (existing mix) | ~10% | Stable |
| Nuclear (new SMRs) | ~5% | Early stage (2028+ online) |
| Wind | ~3% | Modest growth |
The nuclear renaissance in AI is real but slow — existing plant restarts can deliver power in 2-3 years, but new SMRs are still 5-8 years out. Natural gas remains the workhorse for near-term capacity.
All Tracked Facilities
| Facility | Location | Operator | Est. Power |
|---|---|---|---|
| Colossus 1 | 3231 Riverport Rd, Memphis TN | xAI | 300 MW |
| Macrohard / Colossus 2 | 5420 Tulane Rd, Memphis TN | xAI | 1+ GW |
| Stargate Phase 1 | Abilene, TX | OpenAI / Oracle / SoftBank | 1+ GW |
| Meta Richland Parish | Richland Parish, LA | Meta | 4 GW |
| Three Mile Island Restart | Londonderry, PA | Microsoft / Constellation | 837 MW |
| Google Waukesha | Waukesha, WI | ~500 MW | |
| AWS Susquehanna | Berwick, PA | Amazon / Talen Energy | ~960 MW |
| Meta Temple | Temple, TX | Meta | ~1 GW |
| Google Mayes County | Mayes County, OK | ~500 MW | |
| Amazon Mississippi | Madison County, MS | Amazon | ~1 GW |
| CoreWeave | Multiple NJ/TX | CoreWeave | ~2 GW |
Data compiled from SEC filings, FERC dockets, EIA reports, earnings calls, and press releases. Estimates marked with ~ are JustSaid calculations based on publicly available data. Last updated: May 2026.
What We’re Tracking
JustSaid monitors AI energy developments through systematic tracking of primary regulatory and financial sources.
FERC dockets and interconnection queues. Every application to connect new generation or new load to the US power grid must be filed with the Federal Energy Regulatory Commission. These filings reveal data center projects months or years before press announcements. The pipeline monitors FERC’s eQueueing system for applications in the top 20 data center markets, flagging any application over 100 MW that matches AI data center characteristics (high power density, liquid cooling specifications, proximity to fiber routes).
Power purchase agreements. Long-term PPAs between AI companies and power generators are tracked from SEC filings, state public utility commission records, and press releases. The Microsoft-Constellation Three Mile Island deal, the Amazon-Talen Susquehanna agreement, and the Google-Kairos Power SMR contract were all identified through regulatory filings before or concurrent with their public announcement.
EIA generation and consumption data. The Energy Information Administration publishes monthly data on power generation by fuel source and electricity consumption by sector. This data provides ground-truth context for AI power demand estimates and reveals whether grid capacity is keeping pace with data center load growth in key regions.
State and local permitting. Data center projects require local building permits, environmental impact assessments, and often zoning variances. County-level permit tracking in Virginia, Texas, Georgia, and other major data center markets identifies projects in early development stages.
Recent Developments
May 2026: Meta’s Richland Parish campus receives state approval. Louisiana approved Meta’s application for a 4 GW data center campus in Richland Parish, the largest single-site data center ever approved. Construction timeline extends through 2030, with the first phase (approximately 500 MW) expected operational in late 2027. The project includes dedicated natural gas generation and a 500 MW solar installation.
April 2026: Nuclear Regulatory Commission streamlines SMR review. The NRC announced an accelerated review pathway for small modular reactor designs that have already completed preliminary safety assessments, potentially reducing the approval timeline from 7 years to 4-5 years. Google’s Kairos Power reactors and Amazon-backed SMR projects stand to benefit directly.
Q1 2026: Three Mile Island restart on schedule. Constellation Energy confirmed that the restart of Three Mile Island Unit 1 — which will provide 837 MW to Microsoft under a 20-year PPA — remains on track for a 2028 operational date. The project represents the first restart of a decommissioned US nuclear plant and has become a template for similar proposals at other shuttered facilities.
Q4 2025: Grid congestion forces project delays. Multiple AI data center projects in Northern Virginia — the world’s densest data center market — faced delays of 12-18 months due to transmission grid congestion. Dominion Energy’s interconnection queue in the region now extends past 2030 for some applicants, pushing several hyperscalers to redirect projects to less congested markets in the Midwest and South.
Outlook
The AI energy buildout will accelerate through the rest of the decade, but the composition of power sources and the geographic distribution of capacity will evolve significantly.
Natural gas will remain the workhorse through 2028. Despite climate commitments and nuclear enthusiasm, natural gas is the only power source that can deliver gigawatt-scale capacity on 2-3 year timelines. Expect the share of gas-fired generation in the AI power mix to remain above 50% until nuclear restarts and SMRs begin delivering meaningful capacity in 2028-2030.
The nuclear renaissance will be slower than headlines suggest. Restarting existing plants (Three Mile Island, potential restarts at Palisades in Michigan and Iowa’s Duane Arnold) can deliver power in 3-5 years. New SMR construction will take 5-8 years from NRC approval to operational status. The 7 Kairos Power reactors Google contracted will not all be online until 2035. Nuclear will be important for AI’s long-term energy future, but it will not solve the near-term power crunch.
Permitting reform is the single biggest variable. The speed at which new power generation and transmission can be permitted — currently averaging 4-7 years for major projects in the US — will determine whether the announced 38 GW of US AI power capacity can actually be built. Bipartisan support for permitting reform is growing, but legislation has stalled in Congress. If permitting timelines can be halved, the US energy advantage over China in AI will grow dramatically. If they cannot, the advantage will erode as projects stack up in approval queues.
The Middle East is the emerging wildcard. The UAE and Saudi Arabia are aggressively courting AI data center investment, offering cheap land, fast permitting, abundant solar resources, and sovereign wealth fund co-investment. The NEOM AI zone in Saudi Arabia and Abu Dhabi’s data center clusters could capture significant share of global AI compute if they can solve the talent and supply chain challenges of building in the desert.
Frequently Asked Questions
How much power does training a single frontier AI model require? Training a model like GPT-5 or Gemini Ultra requires approximately 50-100 GWh of electricity over 3-6 months — equivalent to powering 5,000-10,000 US homes for a year. The next generation of models, trained on clusters of 100,000+ GPUs, will require proportionally more. Inference (serving the model to users) consumes less power per query but adds up to comparable total consumption due to the volume of queries — ChatGPT alone is estimated to consume over 500 GWh annually.
Why are AI companies interested in nuclear power? Three reasons. First, nuclear provides carbon-free baseload power — continuous, reliable generation that does not depend on weather conditions like solar and wind. Second, nuclear plants can be sited near data centers with dedicated power connections, avoiding grid congestion. Third, the 20-30 year operational life of nuclear plants matches the long-term horizon of AI infrastructure investment. The challenge is timing — nuclear projects take years to approve and build.
Is AI really consuming that much electricity? AI data centers currently consume approximately 2-3% of US electricity generation, up from less than 1% in 2023. By 2030, projections from the International Energy Agency and the US Department of Energy suggest AI could consume 5-8% of US electricity. For context, the entire US residential air conditioning load is approximately 6% of generation. AI is on track to become one of the largest single categories of electricity demand in the US within the next few years.
What happens if power supply cannot keep up with AI demand? AI companies face three options: build their own generation (increasingly common), wait in grid interconnection queues (increasingly long), or build in locations with available power (increasingly competitive). If none of these options delivers sufficient power, training timelines extend, model sizes are constrained, and inference costs remain elevated. Power scarcity becomes a competitive moat — companies that secured power early will have a structural advantage over those that did not.