10x
ChatGPT query vs Google search
Energy per request
1 GW
Single AI cluster power
= 750,000 homes
$20B+
Tech nuclear investments
Announced 2024-25
3x
Data center power by 2030
IEA projection
AI is an energy monster. Training a single frontier model can consume as much electricity as 100 US homes use in a year. Inference at scale is even hungrier. The result? Software companies are becoming energy companies — signing nuclear deals, buying power plants, and racing to secure electricity before their rivals.
The Energy Reality
Every ChatGPT conversation, every AI-generated image, every model training run requires electricity. Lots of it. The AI boom has collided with an aging grid and created an energy crisis that's reshaping how tech companies think about infrastructure.
Energy Per Query: AI vs Traditional Computing
~7 Wh
AI Image Generation
The Scale Problem: A single H100 GPU draws 700W. A training cluster with 25,000 GPUs draws 17.5 megawatts — enough to power 13,000 homes. And the industry is building clusters with 100,000+ GPUs.
The Demand Explosion
US Data Center Electricity Consumption
2030
~8-10% of US grid (projected)
1,000 TWh
Global data center power by 2030
= Japan
Total electricity consumption
A single frontier AI cluster can require as much power as a small city:
Each 🏠 = 50,000 homes. A 1GW AI data center powers the equivalent of 750,000 homes.
The Nuclear Renaissance
Faced with insatiable demand and climate commitments, Big Tech is going nuclear. 2024-2025 saw an unprecedented wave of nuclear energy deals.
☢️
$20B+
Tech nuclear investments announced
⚡
4 GW+
Nuclear capacity contracted
🏭
5+
Nuclear plants involved
| Company | Deal | Capacity | Details |
|---|
| Microsoft |
Three Mile Island restart |
835 MW |
20-year PPA with Constellation. First nuclear restart for AI. |
| Amazon |
Talen Energy (Susquehanna) |
960 MW |
$650M acquisition. Direct nuclear power for data center. |
| Amazon |
SMR investments |
~500 MW |
Small modular reactors via Energy Northwest, X-energy. |
| Google |
Kairos Power SMRs |
500 MW |
First corporate SMR deal. 7 reactors by 2035. |
| Oracle |
SMR data centers |
1 GW+ |
Planning 3 SMRs for 1GW+ data center (announced). |
Why Nuclear? AI data centers need 24/7 baseload power — not intermittent renewables. Nuclear provides carbon-free, always-on electricity. And unlike building new plants (which takes 10+ years), restarting or buying existing plants is faster.
2030 Projections
AI's Share of Global Electricity (IEA Scenarios)
The Grid Problem: US electricity demand grew just 0.5% annually for a decade. Now it's projected to grow 2-4% annually through 2030 — mostly from AI. The grid wasn't built for this. Utilities are scrambling.
📈
+38 GW
New US data center capacity by 2030
⏱️
7+ Years
Grid interconnection backlog
💰
$150B+
Grid infrastructure needed
The Carbon Tension
Tech companies committed to carbon neutrality are now facing an impossible math problem: AI growth vs. climate goals.
2030
Microsoft carbon negative goal
+30%
Microsoft emissions since 2020
48%
Google emissions increase 2019-2023
The Inconvenient Truth: Google's 2023 emissions were 48% higher than 2019, driven by AI. Microsoft's emissions grew 30% since 2020. Both blame data center growth. Both still claim they'll hit carbon goals. The math is... challenging.
Companies are pursuing multiple strategies simultaneously:
☢️
Nuclear
24/7 carbon-free baseload
🌍
Geothermal
Google + Fervo partnership
🔋
Storage
Battery + renewables combo
✓ Key Takeaways
✓ AI queries use 10x more energy than search
✓ Data centers could hit 10% of US grid by 2030
✓ Big Tech is going nuclear — $20B+ in deals
✓ Microsoft restarting Three Mile Island for AI
✓ Grid infrastructure is years behind demand
✓ Climate goals vs. AI growth is unresolved tension