The Google nuclear power agreement turns AI energy demand from a five-year infrastructure problem into a bet stretching deep into the 2040s.

Google Is Buying Nuclear Power for 20 Years. That’s a Very Long Bet on AI.

AI AI Industry by Julian Vane

Google does not know what Gemini will look like in 2046. It does not know which chips will train its models, how efficient those chips will become, whether today’s transformer architectures will still dominate artificial intelligence, or even whether people will still describe the technology using the same vocabulary.

It is nevertheless buying electricity as though one part of the future is already sufficiently predictable.

On October 6, Google and Constellation Energy announced a 20-year power purchase agreement designed to bring 890 megawatts of additional nuclear capacity onto the PJM grid, which serves 67 million people across parts of the Mid-Atlantic and Midwest. Constellation will invest more than $4.3 billion upgrading 11 reactors at six existing nuclear plants in Illinois, Pennsylvania and New Jersey, extracting more electricity from facilities that are already operating rather than waiting for entirely new reactors to be built.

The number that matters most may not be 890 MW.

It may be 20 years.

Technology companies usually inhabit a world in which almost everything becomes obsolete with extraordinary speed. GPUs can feel old within a few product cycles. AI models are superseded within months. Data-center architectures change, software stacks evolve, inference becomes cheaper and yesterday’s computational bottleneck can disappear behind a new optimization technique.

A nuclear power agreement running for two decades belongs to a completely different conception of time.

By helping justify billions of dollars of investment in generating capacity that will serve the grid for decades, Google is making an implicit forecast that survives all the uncertainty surrounding today’s AI boom: whatever artificial intelligence becomes, the company expects large-scale computation to remain hungry for enormous quantities of electricity for a very long time.

That is a much bigger bet than simply believing Gemini will keep growing.

Table of contents

Google’s nuclear deal is really two enormous power deals

The headline agreement between Google and Constellation concerns 890 MW of additional nuclear generation. The capacity will come from upgrades to 11 existing reactors across six sites, where Constellation plans to modernize equipment including turbines, steam generators and digital control systems to increase the amount of electricity the plants can produce.

The first additional capacity is expected in 2028, with the program completed before the end of 2032. Taken together, the upgrades are roughly equivalent to adding the output of a new large nuclear reactor to the grid.

Google’s commitment provides the long-term revenue certainty required to make those investments economically viable. Constellation says the projects represent more than $4.3 billion of new investment, while Google says the arrangement is structured so that the associated costs are not shifted onto other electricity customers.

But the nuclear agreement sits inside an even larger energy relationship.

Google and Constellation have also signed a separate 15-year agreement covering another 2,700 MW from Constellation’s existing generation fleet in the PJM market.

Combined, the arrangements cover 3,590 MW of power.

For perspective, 3.59 GW is not the sort of electricity requirement traditionally associated with a software company. It belongs to the vocabulary of heavy industry, utilities and national infrastructure.

Yet Google is still, fundamentally, a technology company.

That collision between software economics and industrial-scale energy demand is one of the defining consequences of the AI boom.

Twenty years is an eternity in technology

Go back twenty years from 2026 and the technological world becomes almost unrecognizable.

The first iPhone did not exist. Amazon Web Services was only beginning to turn cloud computing into a commercial industry. Nvidia was primarily associated with graphics cards rather than the computational infrastructure of artificial intelligence. Google’s most important business was internet search advertising, and the modern generative AI industry belonged mostly to science fiction and academic research.

Now move twenty years in the other direction.

A contract signed in 2026 reaches into 2046.

Trying to predict the dominant AI model, semiconductor architecture or consumer computing interface that far ahead would be absurd. Google almost certainly cannot do it, and the nuclear agreement does not imply that it can.

The interesting part is that it does not need to.

The company is effectively separating two forecasts that are often confused.

The first concerns which technology wins. That is almost impossible to know over twenty years.

The second concerns whether computation remains economically important enough to require huge amounts of electricity.

Google appears considerably more comfortable making that bet.

This is what makes the duration of the power purchase agreement so revealing. The company does not need Gemini in 2046 to resemble Gemini in 2026. It does not need today’s GPUs, today’s data centers or today’s model architectures to survive.

It needs the demand for computation to survive.

That is a much broader thesis.

The contract is an implicit forecast about AI

A 20-year power purchase agreement should not be interpreted as Google publishing an official 2046 forecast for AI electricity consumption. Energy contracts have their own economics: long commitments can secure supply, reduce price uncertainty and provide producers with the revenue visibility required to finance expensive infrastructure.

But that is precisely why the agreement matters.

Constellation is prepared to spend billions upgrading nuclear plants because Google is prepared to provide long-term economic support for the resulting generation. Google, in turn, is willing to make that commitment because it expects electricity to remain strategically valuable to its operations.

The contract therefore contains information.

It tells us that one of the world’s largest builders of AI infrastructure is planning its physical resource requirements on a timeline far longer than the current investment cycle.

That matters because the central argument against extrapolating today’s AI electricity boom indefinitely is perfectly reasonable: technology improves.

Chips become more efficient. Models become smaller. Algorithms improve. Data centers become better at cooling equipment. Inference costs fall. The amount of useful computation obtainable from one megawatt should increase enormously over twenty years.

Google knows all of that better than almost anyone.

It is signing long-term power agreements anyway.

What if AI becomes dramatically more efficient?

The apparent contradiction between improving efficiency and rising electricity demand is one of the most important questions surrounding AI infrastructure.

If running an AI model becomes ten times more efficient, it is tempting to assume the industry will eventually need less electricity.

That does not necessarily follow.

Cheaper computation can create more computation.

When the cost of performing a task collapses, businesses frequently discover new reasons to perform much more of it. AI systems that are expensive enough to use selectively today could eventually become cheap enough to operate continuously across software, robotics, scientific research, advertising, logistics, healthcare, entertainment and industrial systems.

The workload can also change.

A simple chatbot answering one question is relatively easy to imagine. An autonomous agent working for hours, searching databases, generating video, writing and executing code, interacting with other agents and repeatedly reasoning through a complicated task represents a radically different computational demand profile.

The International Energy Agency has already seen this tension appearing in the numbers. Electricity consumption from AI-focused data centers increased by roughly 50% in 2025 even as efficiency improved rapidly. The agency expects electricity consumption from AI-focused data centers to triple between 2025 and 2030.

Globally, total data-center electricity consumption is projected to rise from around 485 TWh in 2025 to roughly 950 TWh by 2030.

Those forecasts become much less certain farther into the future, which makes Google’s behavior particularly interesting.

The company is not merely betting on the next five years of AI growth.

Its energy strategy is increasingly being built for decades.

Why nuclear fits the AI infrastructure problem

Nuclear power has several characteristics that make it unusually attractive to hyperscale computing companies.

The most obvious is reliability.

Data centers do not stop needing electricity when the sun sets or the wind weakens. Servers operate continuously, and advanced computing infrastructure represents such an enormous capital investment that leaving it idle because electricity is unavailable is economically painful.

Existing nuclear plants can generate large quantities of electricity around the clock with very high capacity factors and low operational carbon emissions.

That does not make nuclear a complete solution to the data-center energy problem. Google continues to invest in renewables, storage and other technologies, and the electricity system itself depends on a mixture of generation sources.

But AI has changed what Big Tech values in energy.

For years, technology companies could approach electricity primarily as a procurement and sustainability question: buy renewable energy, sign power purchase agreements and offset the consumption of expanding data centers.

AI is turning electricity availability into a constraint on growth itself.

When access to power determines whether billions of dollars of computing equipment can be deployed, reliable generation stops looking like a utility bill and starts looking like strategic infrastructure.

Nuclear fits that new reality unusually well.

Google is not waiting for a fleet of new reactors

There is another important detail in the Constellation agreement: these 890 MW are not expected to come from building 11 new reactors.

They come from extracting more output from existing ones.

A nuclear uprate involves upgrading components such as turbines, steam generators and control systems so that an operating reactor can safely generate more electricity. Because the plant, grid connection and much of the surrounding infrastructure already exist, additional capacity can potentially arrive faster than an entirely new nuclear project.

That difference matters enormously to the AI industry.

A new data center can be planned and built much faster than a conventional nuclear power plant. If technology companies waited for a completely new generation of reactors before addressing their power constraints, the timelines would barely match.

Google and Constellation are instead trying to pull additional megawatts out of infrastructure already connected to the grid.

The first uprates are expected in 2028.

By the end of 2032, Google says the program should have unlocked the full 890 MW.

In the strange collision between software and energy, this counts as speed.

This is becoming a nuclear portfolio, not a nuclear experiment

The Constellation deal would be significant on its own. It becomes more revealing when placed beside Google’s other nuclear agreements.

In 2024, Google announced an agreement with Kairos Power intended to support multiple advanced small modular reactors, with deployments planned through 2035 and as much as 500 MW of capacity envisioned.

In 2025, Google, Kairos and the Tennessee Valley Authority announced the first project under that broader collaboration. The 50 MW Hermes 2 plant in Oak Ridge is scheduled to begin supplying electricity to TVA in 2030, with Google procuring clean-energy attributes associated with the generation for data centers in Tennessee and Alabama.

Google has also backed the restart of the Duane Arnold nuclear plant in Iowa and nuclear uprates in Georgia.

Then there is Finland.

In September 2026, Google and Fortum announced a 22-year nuclear power purchase agreement tied to the Loviisa nuclear plant. The arrangement begins in 2028 and is designed to reach up to 50% of Loviisa’s generation capacity between 2030 and 2049, providing the economic certainty Fortum says it needs to support the plant’s lifetime extension through 2050.

Google simultaneously announced €13 billion of investment in Finland for 2027 and 2028, including data-center infrastructure.

The pattern is becoming difficult to dismiss as experimentation.

Google is using long-term demand from its computing operations to help finance existing reactors, reactor upgrades, plant restarts and next-generation nuclear technology.

The company says its agreements involving uprates and restarts are now expected to enable more than 1.5 GW of additional US nuclear capacity.

What began as Big Tech shopping for clean electricity is beginning to look much more like Big Tech participating in energy infrastructure planning.

Electricity is becoming part of the technology stack

The traditional mental model of a technology company starts with software and works downward.

There is an application. Underneath it sits a model or software platform. Underneath that sits cloud infrastructure. Then servers. Then chips.

AI has forced investors to keep descending.

Below the chips are cooling systems, transformers, substations and transmission infrastructure.

Below all of them is electricity.

The deeper the industry descends into that stack, the longer its investment horizons become.

A new AI model can appear in months. A new GPU generation might arrive within a few years. A data center can take several years to develop. Transmission projects and large power plants can require much longer.

Google therefore faces a peculiar planning problem.

Its product cycle is measured in months while part of the infrastructure required to sustain those products is measured in decades.

The 20-year Constellation agreement is one solution to that mismatch.

Instead of waiting to discover how much power AI needs once the demand arrives, Google is attempting to make sure some of that power exists beforehand.

That changes the strategic meaning of energy procurement.

Electricity is no longer simply something the data center consumes.

Access to future electricity becomes part of the infrastructure moat.

A 20-year contract is still a bet

None of this guarantees that current forecasts for AI electricity demand will prove correct.

Twenty years creates enormous room for technological surprise.

AI hardware could become vastly more efficient than expected. New model architectures could reduce computational requirements. More inference could move onto phones, computers and other edge devices. Economic returns from enormous AI investments could disappoint. Regulation could slow deployment. Data-center construction could run into financial, political or physical constraints.

The IEA explicitly models this uncertainty.

Its scenarios for future data-center electricity consumption diverge substantially depending on AI adoption, hardware and software efficiency, infrastructure bottlenecks and the evolution of computational workloads.

That uncertainty is exactly why Google’s contracts are worth studying.

Companies reveal priorities not only through what executives say on earnings calls but through the assets, supply chains and contractual obligations they are willing to secure.

Google cannot know how the next twenty years unfold.

It is nevertheless behaving as though insufficient electricity would be a more dangerous strategic problem than securing too much of it.

That is a meaningful choice.

The interesting question is what Google expects to exist in 2046

Imagine trying to sign a 20-year technology contract in 2006.

You would have been committing capital before the iPhone, before modern cloud computing, before the contemporary smartphone economy, before generative AI and before Nvidia became one of the central infrastructure companies of the digital world.

Most specific predictions would have aged terribly.

Electricity would not have.

Every successive computing revolution required it.

Google’s nuclear strategy appears to be built around that distinction. The company does not have to predict which interface replaces the smartphone, which model architecture succeeds transformers or which accelerator dominates the 2040s. It only has to believe that the economic value of machine computation will remain high enough for companies to keep deploying vast quantities of hardware.

If artificial intelligence becomes more capable, the workloads may expand.

If inference becomes cheaper, usage may expand.

If agents become ubiquitous, they may operate continuously.

If robotics becomes a major computing platform, AI could escape the screen while still consuming computation somewhere behind it.

The exact technology is unknowable.

The infrastructure thesis is simpler: the future will compute.

And computation needs power.

What this means for investors

For investors following the AI boom, Google’s nuclear agreement provides a useful way to distinguish short-term excitement from long-term industrial positioning.

The stock market can debate whether the current generation of AI models justifies today’s valuations. That is a legitimate question.

Google has to solve a different problem.

If AI demand continues growing, it cannot manufacture additional electricity with software. Neither can Microsoft, Amazon, Meta, OpenAI, Anthropic or any other company deploying increasingly large computing systems.

Power infrastructure has physical lead times.

That reality is pushing technology companies into markets they once treated mostly as suppliers: electricity generation, nuclear development, grid infrastructure, energy storage and long-duration power contracts.

It also explains why the AI investment story has expanded far beyond semiconductor stocks. Utilities, nuclear operators, uranium suppliers, electrical-equipment manufacturers, grid companies and data-center developers increasingly occupy the same economic chain as GPUs and cloud platforms.

Investors should still resist the temptation to conclude that every company connected to nuclear energy automatically benefits from AI. Project economics differ, regulatory risks remain substantial, valuations matter and long-term electricity contracts can distribute risk in complicated ways.

But Google’s behavior gives the underlying thesis more weight.

The company is not merely talking about AI needing more electricity.

It is helping finance power infrastructure designed to remain useful long after today’s AI products have become technological archaeology.

A twenty-year agreement cannot tell us what artificial intelligence will look like in 2046.

It tells us something arguably more interesting.

Google expects to need the electricity.

Google nuclear power deal: key questions

What did Google agree to buy from Constellation Energy?

Google and Constellation announced a 20-year power purchase agreement supporting 890 MW of additional nuclear generation on the PJM grid. The new capacity will come from upgrades to 11 existing nuclear units across six sites in Illinois, Pennsylvania and New Jersey.

How much is Constellation investing?

Constellation says the nuclear uprate program represents more than $4.3 billion of new investment and will support approximately 7,200 construction jobs while sustaining roughly 4,400 existing jobs.

When will the additional nuclear capacity become available?

The first uprate is expected in 2028. Google says the projects should deliver the full 890 MW before the end of 2032.

Is Google building new nuclear reactors with Constellation?

Not under this specific agreement. The 890 MW will come from increasing the output of existing Constellation reactors through upgrades to equipment such as turbines, steam generators and digital control systems.

Does Google have other nuclear agreements?

Yes. Google has nuclear projects and agreements involving Kairos Power, TVA, NextEra Energy, Georgia Power and Fortum, among others. Its strategy includes existing-reactor uprates, plant restarts and advanced nuclear technologies.

How long is Google’s nuclear agreement in Finland?

Google and Fortum signed a 22-year power purchase agreement linked to the Loviisa nuclear plant. The PPA begins in 2028 and is designed to cover up to 50% of the plant’s capacity between 2030 and 2049 while supporting investments needed to extend operations through 2050.

Does the 20-year agreement prove AI electricity demand will keep rising for 20 years?

No. The contract is not a forecast guaranteeing a particular level of AI electricity consumption. It does, however, demonstrate that Google is willing to make long-duration energy commitments while expanding its data-center infrastructure, suggesting that the company views access to reliable electricity as a strategic long-term requirement.

Could AI efficiency reduce future electricity demand?

Yes. Hardware, models and data centers are becoming more efficient, and future breakthroughs could materially reduce the electricity required for individual AI tasks. However, falling computational costs can also encourage much greater usage and enable more intensive applications. The net effect remains uncertain.

Why is nuclear attractive for AI data centers?

Nuclear plants can provide large quantities of reliable, low-carbon electricity around the clock. That makes nuclear particularly interesting for data centers containing expensive computing equipment that operators want to keep running at high utilization.

Is nuclear power a guaranteed AI investment opportunity?

No. Rising data-center demand can benefit parts of the nuclear and electricity ecosystem, but individual investments remain exposed to valuation, construction costs, regulation, financing, execution, commodity prices and project-specific risks. This article is analysis, not financial advice.

Sources: