How AI's Appetite is Rewiring the World's Energy Equation

The Bottom Line Up Front: Silicon Valley has discovered its Achilles' heel, and it glows in the dark. The same companies that promised to organise the world's information are now scrambling to power it—with atoms, not bits. We're witnessing the most dramatic reversal in energy strategy since the 1979 Three Mile Island accident froze nuclear development. Today, that very same plant is being resurrected to feed Microsoft's AI ambitions.

The irony would be delicious if the implications weren't so profound.

The Exponential Energy Trap

Numbers don't lie, but they do shock. A generative AI query involving ChatGPT needs nearly 10 times as much electricity to process compared to a Google search, according to Goldman Sachs. This isn't merely an incremental increase—it's a fundamental phase transition in how civilisation consumes energy.

The projections read like science fiction: Global electricity demand from data centres is set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today. The International Energy Agency's latest analysis reveals that data centres will gulp down 945 terawatt-hours by 2030, a staggering doubling from current consumption levels.

But here's where the numbers become existential: Data centre energy consumption in the US from 2014 to 2028 by type shows AI-related servers surging from 2 TWh in 2017 to 40 TWh in 2023. This twenty-fold increase in six years represents the steepest energy consumption curve in modern industrial history. Servers for AI accounted for 24% of server electricity demand and 15% of total data centre energy demand in 2024, yet this is only the beginning of the curve.

The mathematical inevitability is stark: if AI adoption follows typical technology adoption curves, and if current energy intensities persist, we're looking at an energy demand that could challenge the fundamental assumptions underlying our power infrastructure. This isn't about upgrading the grid—it's about rebuilding civilisation's energy foundation.

The Nuclear Pivot: When Silicon Valley Embraces Atoms

The response from Big Tech represents perhaps the most dramatic energy strategy reversal in corporate history. Companies that built empires on Moore's Law and cloud computing are now betting their futures on uranium and fission. The scale of commitment is breathtaking.

Microsoft's atomic awakening began with a 20-year, 835MW agreement to restart Three Mile Island Unit 1, with 100 per cent of the power going to Microsoft data centres. The symbolism is profound: the very site that epitomised nuclear failure is being resurrected as the poster child for AI's energy future. The reactor could be running again by 2028, with Constellation investing $1.6 billion to restore the plant.

Amazon's nuclear shopping spree has been even more aggressive. Amazon bought a 960-megawatt data centre campus from Talen Energy for $650 million, directly connected to the Susquehanna nuclear plant. But that was just the appetiser. Amazon announced it will spend $20 billion on two data centre complexes in Pennsylvania, including one next to a nuclear power plant. The company has also signed a deal with Energy Northwest for a planned X-energy small modular reactor project that could generate 320 megawatts of electricity and be expanded to generate as much as 960 megawatts.

Google's bet on next-generation nuclear represents perhaps the most forward-looking gamble. Google signed the world's first corporate agreement to purchase nuclear energy from multiple small modular reactors (SMRs) to be developed by Kairos Power, with the first reactor planned for 2030 and up to 500 MW of capacity by 2035.

Meta's nuclear embrace completed the hyperscaler quartet with a 20-year agreement with Constellation Energy for approximately 1.1 gigawatts of emissions-free nuclear energy from the Clinton Clean Energy Centre in Illinois.

The SMR Revolution: Nuclear's Second Act

The most fascinating aspect of this nuclear renaissance isn't the resurrection of old plants—it's the birth of an entirely new nuclear paradigm. Small Modular Reactors represent a fundamental reimagining of nuclear power, designed specifically for the digital age.

SMRs are broadly defined as nuclear reactors with a capacity of up to 300 MWe equivalent, designed with modular technology using module factory fabrication, pursuing economies of series production and short construction times. Unlike traditional gigawatt-scale nuclear plants that require custom design and decade-long construction timelines, SMRs promise plug-and-play solutions for on-site power generation that can be placed near data centres.

The technology represents a convergence of nuclear engineering and Silicon Valley thinking. The Aalo Pod uses sodium as a coolant, eliminating the need for water and enabling deployment in arid regions or locations closer to digital infrastructure. Kairos Power's design uses a molten-salt cooling system, combined with a ceramic, pebble-type fuel, to efficiently transport heat to a steam turbine to generate power.

But the most crucial advantage of SMRs isn't technical—it's temporal. The project will help meet energy needs “beginning in the early 2030s,” which aligns with when current AI energy projections suggest the crisis will peak. Traditional nuclear plants require 10-15 years from conception to operation; SMRs promise deployment in 5-7 years.

The Regulatory Battleground

The nuclear-AI convergence has triggered the most complex regulatory battle in modern energy history. The Federal Energy Regulatory Commission's rejection of Amazon's expanded Susquehanna deal represents more than bureaucratic friction—it's a fundamental clash over how America will power its digital future.

FERC's 2-1 ruling said the parties did not make a strong enough case to prove why a special contract allowing for expanded “behind-the-meter” power sales should be allowed. The opposition from utilities was fierce: American Electric Power and Exelon argued the deal could shift as much as $140 million each year to ratepayers.

FERC Chairman Willie Phillips' dissent revealed the national security implications: “There is a clear, bipartisan consensus that maintaining U.S. leadership in artificial intelligence (AI) is necessary to maintaining our national security”. The regulatory friction represents a deeper tension between traditional utility models and the unprecedented demands of the AI economy.

Yet the nuclear industry isn't deterred. Constellation CEO Joe Dominguez characterised FERC's rejection as “not the final word on data centre colocation” at existing nuclear power plants. The companies are adapting, shifting from behind-the-meter arrangements to front-of-meter power purchase agreements that navigate regulatory concerns whilst achieving the same goal.

The Economics of Atomic Power

The financial mathematics of nuclear-powered AI reveal both the opportunity and the challenge. The global market for SMRs for data centres is projected to be $278 million by 2033, growing at a CAGR of 48.72%—one of the fastest-growing energy markets in history.

Traditional nuclear economics have been brutal: plants regularly face cost overruns that double or triple initial estimates. But SMRs promise to change this equation through manufacturing scale and modular design. The final investment decision in 2025 to proceed with the build of a BWRX-300 SMR in Canada was based on a forecast cost of CA$7.7 billion (US$5.6 billion), with an estimated cost of CA$13.2 billion (US$9.6 billion) for three further units.

However, cost remains nuclear's Achilles' heel. Australian scientific research body CSIRO estimated that electricity produced by an SMR constructed from 2023 would cost roughly 2.5 times that produced by a traditional large nuclear plant, falling to about 1.6 times by 2030. The premium is substantial, but tech companies appear willing to pay it for reliable, carbon-free baseload power.

The real economics driver isn't cost comparison with alternatives—it's the cost of not having sufficient power at all. Utilities in places like California and Virginia can't help data centre developers who want a lot of power right now. When your entire business model depends on computational capacity, energy becomes an input cost rather than an operational expense.

The Geopolitical Dimension

Nuclear's resurgence isn't happening in a vacuum—it's occurring against the backdrop of intensifying technological competition between superpowers. The AI arms race has become fundamentally about energy access and control.

The Trump administration's proposed FY26 budget request includes a 21% cut to the DOE's Office of Nuclear Energy and a 51% funding cut to its Advanced Reactor Demonstration Programme, creating tension between the political rhetoric supporting nuclear power and actual funding commitments. Meanwhile, China is accelerating its own nuclear development, with multiple SMR designs in various stages of development.

The strategic implications are profound: nations that can deploy clean, reliable energy at scale will dominate the AI economy. Those that cannot will become digital dependencies. Nuclear power isn't just about electricity—it's about technological sovereignty in the AI age.

The Infrastructure Reality Check

The nuclear renaissance faces a brutal reality: These early projects won't be enough to make a dent in demand. To provide a significant fraction of the terawatt-hours of electricity large tech companies use each year, nuclear companies will likely need to build dozens of new plants, not just a couple of reactors.

The scale mismatch is staggering. The US alone has roughly 3,000 data centres, and current projections say the AI boom could add thousands more by the end of the decade. Even the most aggressive SMR deployment scenarios fall short of meeting projected demand.

This isn't just about nuclear—it's about the fundamental mismatch between digital ambitions and physical constraints. 20% of planned data centres could face delays being connected to the grid, according to the IEA. The bottleneck isn't just generation—it's transmission, distribution, and the basic physics of moving electricity.

The interim solution is uncomfortable: Even as tech companies tout plans for nuclear power, they'll actually be relying largely on fossil fuels, keeping coal plants open, and even building new natural gas plants that could stay open for decades. The nuclear transition will take a decade; AI's energy demands are growing today.

The Climate Paradox

The nuclear-AI convergence creates a fascinating climate paradox. On one hand, nuclear energy has almost zero carbon dioxide emissions—although it does create nuclear waste that needs to be managed carefully. The technology offers a path to massive computational expansion without proportional carbon emissions.

Yet the timeline creates tension. This timing mismatch means that even as tech companies tout plans for nuclear power, they'll actually be relying largely on fossil fuels during the critical next decade when AI deployment accelerates. The clean energy transition may arrive too late to offset the immediate carbon impact of AI's growth.

The broader question is whether AI applications will ultimately reduce global emissions enough to justify their energy consumption. Whilst the increase in electricity demand for data centres is set to drive up emissions, this increase will be small in the context of the overall energy sector and could potentially be offset by emissions reductions enabled by AI if adoption of the technology is widespread.

The Human Element: Who Keeps the Lights On?

Behind the technological and financial complexity lies a human resource challenge that threatens to derail the entire nuclear renaissance. The agency estimates reaching 200 GW of advanced nuclear capacity in the U.S. by 2050 will require an additional 375,000 workers. The nuclear industry lost much of its workforce during the decades-long construction hiatus following Three Mile Island.

The skills required for SMR deployment and operation differ significantly from traditional nuclear expertise. Software-defined reactors, digital control systems, and automated manufacturing processes require a workforce that bridges nuclear engineering and digital technology. The companies betting billions on nuclear power are also betting on their ability to train an entirely new generation of atomic engineers.

This human dimension may prove more challenging than the technology itself. Whilst SMRs promise simplified operation, nuclear power remains unforgiving of human error. The combination of rapid deployment timelines and workforce constraints creates risks that extend far beyond individual projects.

The Systemic Implications

What we're witnessing isn't just an energy transition—it's a fundamental restructuring of how advanced economies organise power generation and consumption. The nuclear-AI convergence represents the emergence of a new industrial model where computation and electricity become inseparable.

Traditional utilities optimised for distributed consumption across millions of residential and commercial customers now face hyperscale consumers whose individual demand exceeds entire cities. Amazon's data centre would consume 40% of the output of one of the nation's largest nuclear power plants, or enough to power more than a half-million homes. This concentration represents a fundamental shift in how electricity markets function.

The model emerging from Silicon Valley's nuclear embrace resembles 19th-century industrial development more than 21st-century distributed systems. Major manufacturers co-located with power sources, creating industrial ecosystems optimised for energy-intensive production. The difference is that instead of steel or aluminium, these facilities produce intelligence itself.

The Evolutionary Pressure

The nuclear-AI convergence creates evolutionary pressure that will reshape both industries. AI companies that secure reliable, clean power sources will possess fundamental competitive advantages over those dependent on grid electricity. Similarly, nuclear companies that can rapidly deploy SMRs will capture the most valuable customers in the global economy.

This pressure is already driving innovation at unprecedented pace. Kairos Power says it hopes to have the first reactor for the Google deal online in 2030 and the rest completed by 2035. In the world of nuclear power, a decade isn't much time at all. Traditional nuclear development timelines are being compressed by Silicon Valley urgency and venture capital.

The convergence is also driving technological innovation that extends far beyond power generation. Advanced radiation detection, novel sensors, and AI-driven security systems developed for next-generation nuclear plants will have applications across multiple industries. The marriage of AI and nuclear is creating technologies that wouldn't emerge from either field independently.

The Unresolved Questions

As compelling as the nuclear-AI narrative appears, fundamental questions remain unresolved. The first is whether SMR technology can deliver on its promises. Like 'traditional' nuclear, the sector faces potential delays and cost overruns, which could undermine its competitiveness with renewable energy sources. No commercial SMR has operated at scale, making current projections largely theoretical.

The second question involves demand evolution. AI models are becoming more efficient even as their usage expands. The relationship between AI capability growth and energy consumption remains uncertain, with potential for both exponential growth and surprising efficiency breakthroughs.

The third question is geopolitical: will nuclear-powered AI create new forms of technological dependency? Nations without advanced nuclear capabilities may find themselves unable to compete in AI development, creating new hierarchies of technological power.

The Historical Echo

The nuclear-AI convergence echoes previous moments when energy transitions reshaped civilisation. The coal-powered Industrial Revolution, the oil-fuelled automotive age, and the electricity-enabled information society all featured similar patterns: new energy sources enabling previously impossible capabilities, creating winner-take-all dynamics that reshaped global power structures.

What makes the current moment unique is the speed and scale of the transition. Previous energy revolutions unfolded over decades; the nuclear-AI convergence is compressed into years. The stakes are correspondingly higher: early movers may establish insurmountable advantages in the technologies that will define the next century.

The irony is palpable: the same Silicon Valley that proclaimed software would “eat the world” now discovers that algorithms need atoms—specifically, uranium atoms—to function at the scale demanded by AI's ambitions. The digital revolution has circled back to the most elemental form of power: nuclear fission.

Looking Forward: The nuclear renaissance represents more than an energy story—it's a transformation of how human civilisation organises itself around computation and power. Success isn't guaranteed, and the risks extend far beyond quarterly earnings or even company survival. We're conducting a real-time experiment in whether nuclear technology can scale rapidly enough to match Silicon Valley's ambitions.

The next five years will determine whether this nuclear-AI marriage produces the clean, abundant energy that powers humanity's greatest technological leap—or whether the misalignment between digital dreams and atomic realities creates bottlenecks that constrain our algorithmic future.

The atoms are moving. The question is whether they'll move fast enough.

InContrarian