Big Tech Is Buying Nuclear Plants. The AI Power Crisis Is That Bad.

In September 2024, Microsoft announced it would restart Unit 1 of the Three Mile Island nuclear power plant — the reactor adjacent to the one that suffered a partial meltdown in 1979. The deal: a 20-year power purchase agreement for 835 megawatts of carbon-free baseload electricity, delivered directly to Microsoft's data center operations.

The symbolism was not subtle. The most infamous name in American nuclear history, resurrected to power artificial intelligence. And Microsoft was just the beginning.

Within six months, Google signed a deal with Kairos Power for small modular reactors. Amazon bought a $650 million data center campus tethered to a nuclear plant. Oracle announced plans for nuclear-powered data centers. Meta issued requests for proposals for 1-4 gigawatts of new nuclear capacity.

Something fundamental shifted. The companies that spent a decade branding themselves as renewable energy champions are now the most aggressive corporate buyers of nuclear power in history. The reason is simple: AI broke the energy math, and renewables alone cannot fix it.

The Numbers That Changed Everything

AI's electricity consumption was supposed to be a manageable line item. In 2022, global data center electricity use was approximately 460 TWh — about 2% of global electricity consumption. Growth projections estimated a modest increase to 500-550 TWh by 2026.

Those projections were wrong. Generative AI training and inference workloads consumed electricity at a rate nobody anticipated. A single training run for a frontier model like GPT-5 or Gemini Ultra consumes 80-120 GWh — more than many small countries use in a year. Inference at scale is worse in aggregate: serving billions of AI queries daily across ChatGPT, Claude, Gemini, and enterprise deployments adds hundreds of TWh of annual demand.

The revised projections are staggering:

By 2030, AI data centers alone could consume more electricity than the entire country of France. This is not a theoretical projection — it is the direct extrapolation of current model scaling trends and deployment growth rates.

The tech industry's problem is not that it uses a lot of electricity. The problem is that it needs this electricity to be available 24/7, at massive scale, in specific locations, starting now. And there is exactly one proven energy technology that meets all four requirements: nuclear fission.

Why Renewables Are Not Enough

This is the part of the argument that makes clean-energy advocates uncomfortable, so let us be precise about what the data says and does not say.

Solar and wind are the cheapest sources of new electricity generation in most markets. Their levelized cost of energy (LCOE) has fallen 90%+ over two decades. They should be — and will be — the foundation of global electricity decarbonization.

But AI data centers have requirements that solar and wind, alone, cannot meet:

Baseload reliability. A hyperscale data center requires 99.99% uptime. Solar generates electricity roughly 25% of the time (capacity factor ~25%). Wind generates roughly 35% of the time. Achieving 99.99% reliability from intermittent sources requires massive battery storage — roughly 4-6x the nameplate solar/wind capacity in lithium-ion batteries. At data center scale (500 MW+), this means billions of dollars in battery infrastructure that currently does not exist in sufficient quantity.

Energy density. A 1 GW data center campus requires approximately 5-7 square miles of solar panels or a comparable wind installation. A nuclear reactor providing the same output occupies less than 1 square mile. In areas near population centers — where data centers are located for latency reasons — land availability is a binding constraint.

Speed of deployment. Big Tech needs hundreds of gigawatts of new electricity capacity within 5-8 years. The permitting, construction, and grid interconnection timeline for utility-scale solar and wind is 3-5 years per project. Nuclear restarts (like Three Mile Island) can deliver power in 18-24 months. New SMR construction, if it meets timelines, could deliver in 4-6 years.

Grid impact. Adding hundreds of gigawatts of intermittent generation to existing grids requires massive transmission upgrades and grid management systems. Nuclear provides dispatchable baseload power that integrates into existing grid infrastructure with minimal modification.

None of this means solar and wind are bad or unnecessary. It means they are insufficient for the specific, enormous, time-constrained demand that AI has created. Nuclear fills the gap that intermittent renewables cannot.

The New Nuclear Lobby

The political dynamics of nuclear energy in America have been static for decades. Environmentalists opposed it. The fossil fuel industry ignored it. The nuclear industry itself was a declining, bureaucratic sector building fewer plants each decade.

Big Tech changed the politics in under two years.

When Microsoft, Google, Amazon, and Meta — companies with massive lobbying budgets, cultural influence, and bipartisan political relationships — decided they needed nuclear power, the policy environment shifted dramatically.

The Nuclear Regulatory Commission, which had been processing license applications at a pace of 1-2 per year, suddenly faced a queue of corporate-backed applications. Congressional support for nuclear energy became genuinely bipartisan for the first time since the 1970s. The ADVANCE Act, signed in 2024, streamlined NRC licensing and reduced regulatory costs for new reactor designs.

State-level politics shifted even faster. Pennsylvania, home to Three Mile Island, approved tax incentives for nuclear restart projects within months of Microsoft's announcement. Georgia, Virginia, and Texas — all states competing for data center investment — fast-tracked nuclear-friendly policies.

The tech industry accomplished in 18 months what the nuclear industry failed to do in 40 years: make nuclear power politically mainstream again. Not because the technology changed, but because the demand driver — AI — is something both parties support, and the solution — nuclear — is something the existing grid cannot substitute.

The SMR Bet

The long-term play is not restarting old reactors. It is building new ones — specifically, small modular reactors (SMRs) designed to be factory-manufactured and deployed at data center scale.

The pitch is compelling: instead of a $20-30 billion, 10-year construction project for a traditional nuclear plant, SMRs promise 50-300 MW modules built in factories, shipped to site, and operational within 3-5 years at a cost of $1-3 billion per module.

Google's deal with Kairos Power, Amazon's investments in X-energy, and Bill Gates' TerraPower project are all SMR bets. The first commercial deployments are expected between 2029-2031.

But SMRs carry real risk. NuScale Power, the furthest-ahead SMR developer, cancelled its first commercial project in 2023 when costs escalated from $5.3 billion to $9.3 billion. No SMR has yet been built at commercial scale in the United States. The factory-manufacturing cost savings are theoretical until proven. And nuclear construction has a long history of cost overruns and schedule delays.

The optimistic case: SMRs work as promised, costs decline with factory learning curves, and by 2032-2035, every new hyperscale data center is powered by on-site or adjacent modular nuclear reactors. The pessimistic case: SMRs are delayed by 5+ years, costs do not decline, and the industry falls back on natural gas — the dirtiest but most available baseload option.

The realistic case is probably somewhere between: SMRs arrive late and over budget but eventually work, and the interim is covered by a combination of nuclear restarts, natural gas, and aggressive grid-scale solar-plus-storage deployment.

The Grid Conflict

There is a darker dimension to Big Tech's nuclear shopping spree that utility regulators are increasingly concerned about: who gets the power?

When Microsoft contracts for all 835 MW of Three Mile Island's output, that electricity is no longer available to the residential and industrial consumers who previously used it. When Amazon buys power from the Susquehanna plant, neighboring communities lose access to clean, cheap baseload electricity.

This is not a theoretical concern. In Virginia, home to the largest concentration of data centers in the world, electricity demand growth has forced Dominion Energy to propose new natural gas plants — burning fossil fuels to backfill the clean energy that data centers absorbed. In Texas, ERCOT has warned that data center demand growth could strain grid reliability during peak summer months.

The equity question is real: should the wealthiest corporations in history be allowed to monopolize clean electricity supply, forcing residential consumers onto dirtier, more expensive alternatives? Utility commissions in multiple states are actively debating this question, and the answers will shape energy policy for decades.

What This Means for the AI Industry

The electricity constraint is not a background issue for AI companies. It is becoming the binding constraint on the entire industry's growth trajectory.

Model training requires massive, concentrated power at specific facilities. Inference requires distributed power at data centers worldwide. Both are growing exponentially. And electricity supply — unlike compute chips or model parameters — cannot be scaled by throwing money at manufacturing. Power plants take years to build. Grid infrastructure takes decades to upgrade. Physics does not bend to deployment schedules.

The companies that secure reliable, abundant, clean electricity will be the companies that can train the next generation of models and serve them at global scale. The companies that cannot secure power will hit a physical ceiling that no amount of software optimization can overcome.

This is why Microsoft, Google, and Amazon are spending billions on nuclear power agreements despite the political complexity and construction risk. They are not making an energy bet. They are making an AI bet. And they have concluded — correctly, based on the data — that the future of artificial intelligence runs on nuclear power, whether the rest of the world is ready for that conclusion or not.

Nuclear energy's comeback is not driven by environmentalism, energy security, or climate policy. It is driven by the most powerful economic force in technology: the insatiable, exponentially growing electricity demand of artificial intelligence. The atom is back, not because we chose it, but because AI did.