Google Nuclear Power Deal: The AI Race Has Become a Power-Plant Race

Behind every chatbot answer, every generated image, every autonomous agent, there is a physical fact we rarely see: electricity, in staggering quantities. This week that invisible foundation became impossible to ignore. On 6 October, Google signed a deal for roughly 3,590 megawatts of power from Constellation Energy — about a quarter of it new nuclear capacity, under a twenty-year agreement — purely to feed its AI data centres. It is one of the largest such deals yet, and it is not an outlier. It is the clearest sign to date that the race to build artificial intelligence has quietly become a race to secure raw electrical power. The GPUs are useless when the lights go out.

What Google Signed

The specifics are worth stating, because the scale is the story. Google contracted 3,590 megawatts of power within PJM Interconnection, the largest electricity grid in the United States: 890 megawatts under a twenty-year nuclear power-purchase agreement, delivered through upgrades to six existing Constellation reactors in Illinois, Pennsylvania, and New Jersey, and a further 2,700 megawatts through a separate long-term supply agreement. The deal enables more than 4.3 billion dollars of new investment, and Google says the added nuclear output is roughly equivalent to building another large reactor — but arriving years faster, because no new reactors are built; existing ones are upgraded. Crucially, this was a response to a new PJM proposal that would require data-centre customers to bring their own power or risk being cut off from the grid during peak demand. In other words, the grid operator effectively told the tech giants: your appetite is now so large that you must supply your own electricity. Google's answer was to go and buy a power company's output for two decades.

Not Just Google: A Pattern

The reason this matters beyond one company is that it completes a pattern, and the pattern is the real headline. Over the past two years, every one of the hyperscalers has done a version of this. Microsoft struck the deal that is restarting the Three Mile Island reactor in Pennsylvania to feed its data centres. Amazon partnered with Talen Energy to draw on the Susquehanna nuclear plant for AWS. Meta signed a twenty-year deal for over a gigawatt of nuclear power in Illinois. Google itself had already contracted to restart a plant in Iowa before this week's far larger move. Read together, these are not isolated energy contracts. They are the five giants collectively concluding that the binding constraint on AI is no longer chips, or money, or even talent — it is power, measured in gigawatts, and they are moving to lock up decades of it. PJM power prices have surged more than elevenfold since 2024 as the grid strains to keep up. The race everyone has been watching on benchmark charts is, underneath, a race for the physical capacity to generate electricity.

The Cost We Don't See on the Screen

This is where the story reaches the rest of us, and it deserves to be told plainly. An AI answer feels weightless — a few words appearing on a screen. It is not weightless. It is the end point of a chain that now runs all the way back to upgraded nuclear reactors and a grid under visible strain. And the hunger is reshaping things far from the data centre. On the same day as Google's deal, reports described memory chips being diverted en masse into AI infrastructure, driving a roughly sixty-percent drop in shipments of sub-100-dollar smartphones — the devices on which much of the developing world first comes online. The pattern is consistent: the resources the AI boom consumes, whether electricity or memory, are being pulled toward the richest companies on earth and away from everyone competing for the same finite supply. When a handful of corporations buy decades of a nation's power output and a continent's worth of memory production, the cost does not disappear. It lands on grids, on prices, on the people and places with the least leverage to compete for what is left.

Claim and Counter-Claim

In fairness, there is a genuinely positive reading, and it should be stated clearly. The nuclear turn is, in climate terms, far better than the alternative. If AI's enormous electricity demand must be met, meeting it with carbon-free nuclear power — by upgrading and restarting existing plants rather than burning gas or coal — is close to the best available outcome, and it is channelling billions into a clean-energy source long starved of investment. Google is not hiding its consumption; it is contracting transparently to supply it cleanly, and bringing idle or under-used reactors back to productive life. Some of these deals are reviving nuclear capacity that would otherwise have been lost. That is real, and environmentalists who have made peace with nuclear power can reasonably call it progress.

The counter-case is about concentration and priority, not carbon. Even clean power is finite, and every gigawatt locked into a twenty-year contract for a data centre is a gigawatt not available for homes, hospitals, electrified transport, or industry — and it is being claimed by the entities most able to outbid everyone else. The elevenfold surge in grid prices is paid, in part, by ordinary ratepayers who never ran a single AI query. And the deeper question is one of social choice made without a social decision: no one voted to prioritise AI compute over other uses of a constrained clean-energy supply, yet the market, driven by a handful of trillion-dollar firms, is making that allocation anyway, one twenty-year contract at a time. The honest synthesis: the nuclear path is the right way to power AI if AI must be powered at this scale, and the scale itself — and who gets to command the finite clean energy to feed it — is a question that is being settled by corporate purchasing power rather than by any public reckoning.