The AI Energy Crisis: When Data Centers Eat the Grid

Key Takeaways

• The Scale: US data centers already consume 4.4% of the country’s electricity, and federal researchers project 6.7–12% by 2028. The individual query is cheap; the fleet is not.
• The Bottleneck: It’s not chips; it’s transformers. Utilities cannot build transmission lines fast enough to power new gigawatt-scale data centers.
• The Solution: Tech giants are buying their own generation. Amazon and Google are investing billions in Small Modular Reactors (SMRs)—mini nuclear plants—while Microsoft is restarting Three Mile Island.

We thought the limit to AI would be intelligence. It turns out, it’s electricity.

In late 2025, the “AI Energy Crisis” is the biggest topic in Silicon Valley. Training a frontier model consumes tens of gigawatt-hours. Running it at scale (inference) requires even more.

Data centers consumed about 1.5% of global electricity in 2024 (415 TWh, per the IEA) — but in the United States the share is already 4.4%, after US data-center demand more than doubled between 2017 and 2023. By 2030, the IEA projects global data-center consumption will more than double to roughly 945 TWh — slightly more than Japan’s entire electricity use today. In the US, federal researchers at Berkeley Lab see data centers taking 6.7–12% of national electricity by 2028.

The Scale of the Problem

To understand the crisis, you have to look at the math.

• Training vs. Inference: Training GPT-3 emitted an estimated 552 tonnes of CO₂, and frontier models since are far larger (the labs no longer disclose figures). But training is a one-time cost. The real drain is inference—every time you ask an agent to “plan my vacation,” thousands of GPUs spin up. The per-query cost is small (OpenAI puts an average ChatGPT query around 0.3 Wh, roughly on par with a classic Google search); the billions of queries are not.
• Water Usage: It’s not just power. Data centers drink water to stay cool. Microsoft’s water consumption jumped 34% in fiscal 2022 as the AI buildout began, sparking conflicts with local communities from West Des Moines, Iowa to drought-prone Goodyear, Arizona.

The Nuclear Renaissance: Big Tech’s New Bet

Renewables (solar/wind) are great, but they are intermittent. AI needs 24/7 baseload power with 99.999% uptime. Batteries can smooth peaks, but covering multi-day, gigawatt-scale baseload with them remains uneconomical.

Enter Nuclear.

• Microsoft: Signed a historic 20-year PPA to restart Three Mile Island (Unit 1) — now the Crane Clean Energy Center — to feed its AI cloud with carbon-free power through the PJM grid.
• Amazon: Paid $650 million for a data center campus next to the Susquehanna nuclear plant. Regulators rejected the original “behind the meter” setup, so in 2025 the deal was restructured into a 17-year, grid-delivered PPA for up to 1,920 MW — proof that even Big Tech can’t fully bypass the public grid.
• Google: Investing heavily in fusion startups and SMRs (Small Modular Reactors). These “nuclear batteries” can be manufactured in factories and shipped to data center sites, bypassing the decade-long construction delays of traditional plants.

Big Tech has effectively become the new utility sector. They are funding the next generation of clean energy infrastructure because they have no choice. The grid cannot keep up with them.

The Irony: Green AI, Dirty Power

There is a bitter irony here. We are building AI to help us solve climate change (optimizing grids, designing new materials), but the creation of that AI is currently driving up emissions.

• The “Jevons Paradox”: As AI becomes more efficient, we use it more, leading to higher total consumption.
• The Reality Check: Google’s carbon emissions are up 51% versus 2019, largely due to AI infrastructure, per its own 2025 environmental report. The company retired its longstanding “carbon neutral” claim back in 2023; its 2030 net-zero target still stands on paper, but Google now frames it as an ambition rather than a commitment.

The Future: A Bifurcated Grid?

We are heading toward a bifurcated grid: one for the people, and one for the machines.

• The Public Grid: Struggling with aging infrastructure, intermittent renewables, and rising costs for consumers.
• The Private Grid: Powered by dedicated nuclear and geothermal assets, owned by trillion-dollar tech companies, ensuring their AI never sleeps.

The question is no longer “Can we build it?” but “Who pays for the upgrade?” If history is any guide, the consumer will foot the bill for the grid, while Big Tech builds its own fortress of power.

Sources (8)

• iea.org Energy and AI
• energy.gov DOE/Berkeley Lab: US Data Center Energy Usage Report
• sustainability.google Google Environmental Report 2025
• epoch.ai How Much Energy Does ChatGPT Use?
• utilitydive.com Constellation-Microsoft Three Mile Island PPA
• ir.talenenergy.com Talen Energy Expands Nuclear Relationship with Amazon
• arxiv.org Carbon Emissions and Large Neural Network Training (Patterson et al.)
• fortune.com AP: AI Fuels Spike in Microsoft Water Consumption