Datacenters now consume 4% of US electricity — and Microsoft is restarting a nuclear plant
AI’s energy appetite is resurrecting technologies we thought were buried
The number that should concern us
Data centers already account for 4% of US electricity consumption. And that’s just the beginning.
According to Georgia Tech projections, if we continue the current pace of AI development, we’ll need to triple the energy supply for datacenters by 2030.
To put this in perspective: in regions like Northern Virginia, home to the world’s largest server clusters, datacenters already consume more than a quarter of the regional power supply.
AI is hungry. And we’re not ready to feed it.
The most unexpected solution: Three Mile Island
If there’s one name that symbolizes nuclear fear in America, it’s Three Mile Island.
On March 28, 1979, the Unit 2 reactor at this Pennsylvania plant suffered a partial core meltdown — the worst nuclear accident in US history. The incident paralyzed the American nuclear industry for decades.
Unit 1, which wasn’t involved in the accident, continued operating until 2019, when Constellation Energy determined it was no longer economically viable against cheap natural gas.
That seemed like the end of the story.
But in September 2024, Microsoft and Constellation Energy announced a 20-year power purchase agreement: Three Mile Island will resume operations in 2028, renamed the Crane Clean Energy Center, to power Microsoft’s datacenters.
Why nuclear, why now
The logic is simple but brutal.
AI datacenters require:
- Constant power — 24/7, no interruptions
- Clean energy — Microsoft has committed to being carbon negative by 2030
- Abundant energy — the largest AI datacenters require hundreds of megawatts of continuous power
Renewables (solar, wind) are intermittent. Natural gas emits CO2. The only source that meets all three requirements is nuclear.
Three Mile Island Unit 1 has 837 megawatts of carbon-free capacity, available 24 hours a day, 365 days a year.
Constellation’s CEO Joe Dominguez put it bluntly in June 2025: “We made a mistake closing this plant, but we’re not here to lament that mistake.”
It’s not just Microsoft
Big Tech’s nuclear race has accelerated dramatically:
Microsoft: Agreement with Constellation to reopen Three Mile Island (837 MW)
Meta: 20-year agreement with Constellation to extend the Clinton Clean Energy Center’s life in Illinois (1.1 GW). Also announced seeking 1-4 GW of new nuclear capacity.
Amazon: Agreement with Talen Energy for 1.9 GW from Susquehanna Nuclear Plant in Pennsylvania, plus investments in small modular reactors (SMR).
Google: First partnership with Kairos Power to develop small modular reactors near their datacenters.
According to Deloitte analysis, nuclear energy could meet up to 10% of datacenter electricity demand by 2035.
The challenge of resurrecting a reactor
Reopening Three Mile Island isn’t as simple as flipping a switch.
A former Federal Energy Regulatory Commission (FERC) chairman warned in January 2026 that no fully closed nuclear reactor has ever been restarted in the United States. The reasons:
-
Regulatory barriers: The reactor needs approval from the NRC (Nuclear Regulatory Commission), EPA, FERC, and state and local authorities. Even with a pro-nuclear administration, existing legal processes are extensive.
-
Equipment condition: After 5 years closed, critical systems must be verified — turbines, generators, transformers, cooling and control systems.
-
Supply chain: Nuclear fuel and specialized components require a supply chain that has atrophied.
-
Talent: Highly qualified personnel must be hired and trained. Constellation reports they’re ahead on hiring because many former employees have returned.
-
Cultural stigma: Three Mile Island remains synonymous with nuclear disaster for many Americans, even though the accident was at reactor 2, not reactor 1.
Constellation initially said the reactor would be operational in 2028. In January 2025, they moved the date up to 2027. But nuclear projects have a history of delays and cost overruns — the Vogtle plant reactors in Georgia were completed 7 years late and $17 billion over budget.
The second nuclear renaissance
What we’re witnessing is a paradigm shift.
For decades, the narrative was clear: nuclear energy is dangerous, expensive, and being replaced by renewables and natural gas.
Now, the narrative is inverting:
- The Trump administration has ordered the Department of Energy to end support for solar and wind and facilitate nuclear development
- Energy Secretary Christopher Wright described the Three Mile Island project as the “poster child” of the government’s energy agenda
- State governments are competing to attract nuclear investment as an economic development driver
Pennsylvania Governor Josh Shapiro put it this way: “We’re going to make sure they never have to leave again. This will be one of those places where jobs are passed down from generation to generation.”
The elephant in the room: cost
None of this is cheap.
- The Microsoft-Constellation deal will add $16 billion to Pennsylvania’s GDP according to estimates
- It will create 3,400 jobs and generate over $3 billion in taxes
- But electricity costs for consumers are already rising due to datacenter demand
PJM Interconnection, the electricity market operator for Pennsylvania and 12 other states, announced in December 2025 that its latest capacity auction failed to secure enough generation to cover the 20% reserve for peak demand in 2027-2028.
In other words: demand is growing faster than supply. And someone has to pay.
The coming debate
The next few years will see an intense debate over who bears the cost of AI energy infrastructure:
AI users? Through higher prices for cloud services and subscriptions.
Electricity consumers? Through higher rates for everyone.
Taxpayers? Through government subsidies to the nuclear industry.
The environment? If demand exceeds clean energy capacity and we resort to fossil fuels.
There are no easy answers. But what’s clear is that AI has a physical cost we’ve conveniently ignored until now.
My perspective
As someone who works with data and uses AI tools daily, it’s uncomfortable but necessary to acknowledge this:
Every ChatGPT query, every DALL-E image, every model I train has a real energy cost.
I’m not saying we should stop using AI. I’m saying we need to be honest about what it costs — and start asking if we’re willing to pay that cost.
This connects directly to why 95% of companies aren’t seeing results with their AI copilots: if we’re going to assume these energy costs, the benefits better be real and measurable.
Microsoft restarting Three Mile Island is a solution. But it’s also a symptom of a bigger problem: we’ve built an industry seeking $7 trillion in investment without figuring out how to power it.
Conclusion
The irony is hard to ignore.
The most “futuristic” technology of our era — generative artificial intelligence — is being powered by the resurrection of infrastructure we thought was obsolete: 50-year-old nuclear plants.
This tells us something important about the present: the future always has physical costs. Bits need atoms. Models need electricity. And electricity has to come from somewhere.
Three Mile Island coming back to life is a perfect metaphor for this moment: past and future colliding in the present, with all the contradictions that implies.
Sources
- IEEE Spectrum, NPR, Pennsylvania Capital-Star (2024-2025) — Microsoft-Constellation deal
- Georgia Tech News (January 2025) — Energy demand projections
- EIA (U.S. Energy Information Administration) — Datacenter consumption data
- Trellis, Deloitte (2025) — Amazon, Google, Meta nuclear deals
- The Hill (January 2026) — Regulatory obstacles analysis
You might also like
The $25B Patagonia datacenter: historic opportunity or digital extractivism?
While OpenAI promises to turn Argentina into an AI hub, families in Querétaro have been without water for months. Analysis of the Stargate project and its implications.
The copilot lie: why 95% of companies aren't seeing results
Studies from METR, MIT and California Management Review reveal AI isn't delivering the promised productivity gains. Data-driven analysis and what successful companies are doing differently.
You have 3-5 years before AI agents become normal
78% of executives believe digital ecosystems will be built for humans AND AI agents. The window to position yourself is closing.