AI ENERGY
US Frontier AI Data Center Power Monitor
Read the full analysis
The Grid War: AI Has Run Out of Power
Current Status
DEFCON 4 - Severe
Palisades Nuclear Restart
800 MW baseload capacity returns
The Depreciation Bomb
H100 chip obsolescence crossover
Transformer Lead Time Apex
Peak supply chain bottleneck
Grid Capacity Catch-Up
Projected equilibrium point
Critical Window: Q2 2026 - Q1 2027
Peak collision between AI deployment pressure and transformer supply bottleneck. H100 generation enters depreciation curve while H200/B200 shipments accelerate without power infrastructure.
Understanding the Deficit
Loading power deficit data...
Loading transformer queue data...
Note: This chart shows AI-specific demand (17 GW gap in 2025). The "Total Grid Gap" in the Famine Clock (47.2 GW) includes additional demand from EVs, industrial electrification, and reserve margin erosion.
Year
2025
17 GW
12.8M homesGap Growth
+20
GW/yr
2025
17
GW
2026
37
GW
2028
67
GW
2030
78
GW
Exponential Gap Growth
AI demand growing at ~40% CAGR while AI-allocated grid capacity grows at ~8% CAGR. By 2030, the 78 GW AI deficit equals 97 nuclear reactors worth of missing power.
Stranded GPUs
2.28M
46%
of shipped
Deployed GPUs
2.72M
54%
utilized
Stranded Power
1.8 GW
1.4M homesIdle Value
$0.0B
Active Power
4.1 GW
3.1M homesData Centers
9
The Bottleneck: NVIDIA shipped ~5M H100-equivalent GPUs, but only 2.7M are deployed in data centers with power. The remaining 2.3M GPUs sit idle, depreciating at 20%/year while awaiting grid capacity.
Power sources for operational AI data centers
Aggregate Energy Mix
Fossil Fuels Dominate: 56% of AI Power
Despite corporate carbon neutral pledges, Natural Gas (35%) is the largest power source for US frontier AI data centers. The grid mix is determined by regional infrastructure, not company choice. Nuclear provides 25%, renewables (wind + solar) provide 15%.
Energy Sources by Data Center
Source: Regional grid mix from EIA Form 861 (2024). Energy mix reflects the regional grid where each data center is located, not direct power purchase agreements. Some facilities (xAI Memphis, Crusoe Texas) use on site gas turbines. IEA Energy and AI Report (2024) for methodology.
Grid reliance vs on-site generation by data center
74% of AI Compute is Grid Dependent
3,054 MW of 4,123 MW total capacity relies entirely on grid infrastructure. These facilities are bottlenecked by the interconnection queue (5+ year wait). Only 667 MW has bypassed the queue via on-site generation.2.3M homes
5
2,511 MW
1
543 MW
2
793 MW
1
276 MW
3,054 MW
Fully queue bottlenecked
2.3M homes667 MW
Bypassed via on-site generation
500K homesData Center Power Profiles
Grid connected via NIPSCO
751 MW
100% grid
Grid connected via AEP Ohio
615 MW
100% grid
100% grid dependent via Georgia Power
615 MW
100% grid
Nuclear PPA with Constellation (still grid delivered)
543 MW
100% grid
Operating unpermitted turbines since 2024, TVA grid as backup
498 MW
40% grid
Grid connected via Entergy Mississippi
341 MW
100% grid
1 GW facility planning mixed on-site generation
295 MW
50% grid
Scaling turbine fleet for 200k GPU expansion
276 MW
20% grid
Grid connected via OPPD
189 MW
100% grid
Source: Company SEC filings, Reuters, CNBC, Shelby County Health Department permits (xAI), EIA regional grid data. On-site generation estimates based on turbine counts and manufacturer specifications.
Loading interconnection data...
Current Power Deficit
47.2 GW
Critical Threshold
Q2 2026
Bypasses 124-week large power transformer queue. Uses readily available distribution transformers (8-16 week lead time).
Projected Outcome
Original Deficit
47.2 GW
Capacity Added
+0.0 GW
New Deficit
47.2 GW
June 2026
No change from baseline
Methodology: Impact estimates based on DOE reactor capacity data, NuScale SMR specifications, LBNL queue analysis, and WoodMac transformer market reports. Actual results depend on execution timelines.
Understanding the US AI energy infrastructure crisis
The Grid War: AI Has Run Out of Power
The unseen fight between hyperscalers, physics, and a grid built for a slower world.