Here is where we are: the biggest technology companies on Earth have spent so much electricity training AI models that they are now building their own nuclear power plants. Not investing in solar. Not buying wind credits. Nuclear reactors. The kind that split atoms.

Reuters reported this week that Meta, Amazon, and Google are inking deals with modular reactor developers, reshaping the entire funding landscape for next-generation nuclear technology. The reason is brutally simple: AI data centers consume more electricity than some countries, and the grid was not built for this.

The Numbers That Forced Their Hand

Training a single frontier AI model can consume as much electricity as 100,000 U.S. households use in a year. And that is just training. Inference, the act of running the model for users, multiplies that by orders of magnitude as usage scales. When Anthropic signed a 3.5 gigawatt TPU compute deal with Google this week, that single contract locked up more power capacity than most mid-sized cities need.

The math does not work with renewables alone. Solar and wind are intermittent. You cannot pause a $50 billion training run because it is cloudy. AI needs baseload power: reliable, always-on, and massive. That describes exactly one energy source that does not produce carbon emissions: nuclear.

The New Nuclear Customers

Meta has been particularly aggressive. The company that burns 60 trillion AI tokens a month at a cost of $900 million is backing modular reactor projects that promise smaller, cheaper, faster-to-build nuclear plants. Amazon has signed similar deals. Google has gone further, partnering with nuclear developers to power data centers in multiple countries.

For the nuclear industry, this is transformative. Nuclear startups have struggled for decades with a chicken-and-egg problem: they need revenue to prove their technology works, but nobody buys until the technology is proven. Big Tech is solving that problem by writing checks upfront, giving nuclear companies both funding and guaranteed customers.

The Catch

Small modular reactors sound great on paper. In practice, nobody has built one at commercial scale in the United States. The timelines are measured in years, the regulatory approvals are uncertain, and the cost estimates have a habit of doubling. Bernie Sanders and 140 community groups across 24 states are already pushing for a moratorium on new data centers. Adding nuclear reactors to the mix is not going to calm anyone down.

But Big Tech does not have a choice. The alternative is worse. OpenAI plans to burn $85 billion this year. Anthropic just signed deals for 3.5 gigawatts of compute. Meta is spending $900 million a month on tokens. This is an arms race, and the weapon is electricity. Whoever has the most power, wins.

The irony is thick enough to cut: the companies building AI to solve climate change are now building nuclear reactors to power the AI. At some point, someone should ask whether the AI is actually solving more problems than it creates. But that conversation will have to wait. Right now, the nuclear deals are getting signed.