The technology industry spent thirty years building a reputation for immateriality — bits, not atoms. That reputation is now an obstacle to understanding what AI actually requires.
There is a number that appears in almost no mainstream coverage of the artificial intelligence boom, and that number is 500 gigawatt-hours. That is the approximate electricity consumption of training a single frontier large language model at the scale of GPT-4 or its successors. For reference, 500 gigawatt-hours is roughly the annual electricity consumption of 45,000 average American households. It is the energy required to run a mid-sized city for a month. And it is the cost of training a model once — before deployment, before inference, before the continuous compute load of serving responses to millions of users simultaneously.
The AI industry’s energy consumption is not a side effect of the technology. It is the technology. AI is, at its core, a process of transforming electricity into mathematical operations at sufficient scale and speed to produce outputs that approximate intelligence. The quality of that approximation improves as the scale of the computation increases. The scale of the computation is bounded by the availability of electricity. This is not a temporary constraint that better algorithms will eventually dissolve. It is a physical constraint that compounds as AI capability improves, because more capable AI requires more parameters, more training data, and more inference compute to serve at scale.
The Scale of What Is Being Built
The numbers that the major technology companies have publicly committed to data center investment over the next five years are large enough that they require context to process. Microsoft has announced $80 billion in data center capital expenditure for fiscal year 2025 alone. Google has committed to $75 billion. Amazon’s AWS capital expenditure guidance has moved steadily upward, with $100 billion committed for 2025. Meta has announced plans for data centers covering millions of square feet. The aggregate capital commitment from the four largest hyperscalers exceeds $300 billion for a single year — and these commitments are accelerating, not moderating.
Each dollar of that capital expenditure requires electricity to operate the hardware it purchases. The relationship between capital expenditure and power consumption is not linear, because hardware efficiency improves over time, but it is persistent and substantial. Industry analysts estimate that the AI data center buildout underway will add 50 to 80 gigawatts of new power demand to the global grid by 2030. To put that in context: the entire installed electricity generation capacity of Germany — one of the largest economies in the world — is approximately 250 gigawatts. The incremental power demand from AI data centers over the next five years represents a meaningful fraction of an entire major economy’s generation capacity.
Grid operators and utilities in the regions where data centers are being built — Northern Virginia, the Dallas-Fort Worth metroplex, Iowa, Oregon, Ireland, Singapore, the Netherlands — are reporting interconnection queues that stretch years into the future. Power purchase agreements are being signed at prices that would have been considered uneconomic five years ago, because the demand from data center operators is strong enough relative to available supply that the supply side has pricing power it has rarely had in the history of the electricity market. New transmission infrastructure is being accelerated. Permitting timelines are being compressed through emergency procedures that bypass the normal years-long process. The electricity system is being asked to grow faster than it has been required to grow since the post-war industrial expansion of the 1950s and 1960s.
Why Renewables Alone Cannot Solve This
The technology industry’s public commitment to renewable energy is genuine in the sense that the capital being deployed into renewable procurement is real and substantial. Google, Microsoft, Amazon, and Meta have all made commitments to matching their electricity consumption with renewable energy certificates, signing long-term power purchase agreements with wind and solar developers, and investing directly in new renewable capacity. These commitments have driven significant development of renewable energy projects that would not have been financed without the corporate demand signal that technology companies provide.
The operational reality of those commitments is more complicated than the press releases suggest. Renewable energy certificates are an accounting mechanism, not a physical guarantee of clean power. A data center that purchases RECs while drawing power from a grid that is 70% fossil fuel powered is not, in any operational sense, running on renewable energy — it is running on the grid mix and purchasing financial instruments that fund renewable development elsewhere. This distinction matters because the grid impact of data center power consumption is determined by the actual physical grid mix at the time and location of consumption, not by the financial instruments associated with that consumption.
More fundamentally, solar and wind energy are intermittent. They generate power when the sun shines and the wind blows, which is not the same as when the data center needs power. Data centers operate at high utilization rates continuously — the compute load of inference doesn’t pause for cloud cover or calm weather. The gap between renewable generation availability and data center demand is filled by whatever the grid has available at that moment, which in most of the world’s major grid regions is still predominantly fossil fuel generation. Battery storage can buffer intermittency, but at the scale required for a 500-megawatt data center running at constant utilization, the storage cost is prohibitive by the economics of current technology, and the manufacturing supply chain for grid-scale batteries cannot scale fast enough to meet the data center demand timeline.
The backup generation problem is the most underappreciated aspect of the renewable-data center relationship. When renewable generation underperforms — during extended periods of low wind across a broad geographic area, during winter months when solar irradiance is reduced, during the kind of demand spikes that cold snaps and heat waves produce — the grid falls back on dispatchable generation. In most major grid regions, dispatchable generation means natural gas or coal. The AI industry’s actual carbon footprint, calculated on a marginal generation basis rather than a certificate-matching basis, is substantially higher than the headline commitments suggest.
The Nuclear Renaissance Is Not a Metaphor
Nuclear power solves the problems that renewable energy cannot solve for AI infrastructure, and the technology industry has reached that conclusion faster than most energy policy analysts expected. A nuclear plant generates electricity continuously, at capacity factors of 90% or above, regardless of weather, season, or time of day. It requires no natural gas backup to maintain reliable output. It produces no direct carbon emissions during operation. The energy density of nuclear fuel — one kilogram of uranium produces approximately as much energy as 3,000 tonnes of coal — means that the physical supply chain for nuclear fuel is manageable at scale in a way that coal or gas supply chains are not. For a technology company that has committed to carbon neutrality and needs reliable, high-density, continuous power for its AI infrastructure, nuclear is not an ideological preference. It is the engineering answer to a physical problem.
The response from the technology industry has been rapid and concrete. Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Unit 1 of Three Mile Island in Pennsylvania — a plant that was permanently shut down in 2019 following the closure of its second unit, and that most energy analysts had written off as a historical artifact. The agreement required Constellation to invest approximately $1.6 billion in the plant’s restart and recommissioning. It committed Microsoft to purchasing the plant’s entire output for 20 years at a price that made the restart economically viable. The announcement that a technology company had contracted to restart a permanently decommissioned nuclear plant was treated as surprising. It should not have been. It was the logical outcome of the AI energy equation reaching a scale where the economics of nuclear power, which had been uncompetitive against cheap natural gas for most of the past decade, became competitive when the alternative was rationing data center expansion.
Google followed with a partnership with Kairos Power, a developer of small modular fluoride salt reactors, for a series of SMR deployments beginning before 2030 and scaling through the 2030s. Amazon Web Services signed agreements with multiple SMR developers and with Dominion Energy for nuclear power in Virginia. Meta announced plans to issue a request for proposals for nuclear power developers, explicitly citing the need for reliable, carbon-free baseload power for its AI infrastructure. The constellation of these announcements, across multiple companies, multiple reactor technologies, and multiple geographies, represents a demand signal for nuclear power that the industry has not received since the original nuclear buildout of the 1970s.
The SMR Story: Promise and Timeline Reality
Small Modular Reactors — defined broadly as nuclear plants with output typically between 50 and 300 megawatts, compared to 1,000 to 1,700 megawatts for a conventional large reactor — have attracted disproportionate attention in the AI energy discussion because they offer a combination of characteristics that appear well-suited to the data center use case: siting flexibility that allows placement closer to demand, factory fabrication that potentially reduces construction cost and timeline compared to conventional large nuclear, and output scale that matches individual large data center campuses rather than requiring grid-level integration.
The promise is real. The timeline is not. Most of the SMR developers that technology companies have signed agreements with are, as of 2026, pre-commercial. They have demonstrated reactor designs. They have received or are in the process of receiving regulatory approval. They have not built and operated commercial reactors that have generated power on the grid. The transition from regulatory approval to commercial operation typically takes five to seven years for conventional large nuclear plants in jurisdictions with efficient approval processes, and longer in jurisdictions without them. For SMR technology that does not have a comparable commercial operational track record, the timeline to first commercial power is genuinely uncertain, and the cost projections that developers are offering — which form the basis of the economics in the power purchase agreements technology companies are signing — have not been validated by commercial construction experience.
This does not mean the SMR thesis is wrong. It means the SMR contribution to the AI energy problem is a 2030s story rather than a near-term solution. The Three Mile Island restart, which involves a known reactor design in an existing facility with a trained workforce, is a nearer-term contribution — commercial power delivery expected by 2028. The SMR deployments announced for the late 2020s and early 2030s will contribute meaningfully to the decade’s power needs if they execute on schedule, which is a significant conditional.
The Geopolitical Dimension: Who Controls the Power Controls the AI
The energy-AI nexus has a geopolitical dimension that is frequently underweighted in market analysis because it operates on a longer timescale than investment quarterly cycles but that will prove more consequential than near-term price movements in any of the affected assets.
Nations with established nuclear infrastructure, regulatory frameworks that can process new nuclear construction efficiently, and domestic uranium supply chains are acquiring a structural advantage in the AI economy that compounds over decades. France generates approximately 70% of its electricity from nuclear power — a legacy of a 1970s energy policy decision that was controversial at the time and that looks prescient in the current environment. France’s electricity costs are among the lowest in Europe, and French data center operators have access to reliable, low-carbon baseload power that operators in Germany — which phased out nuclear in 2023 and is now managing the consequences — do not. South Korea’s nuclear construction industry is among the most efficient in the world, with per-unit construction costs and timelines that substantially undercut Western equivalents. Canada has both uranium mining capacity and established CANDU reactor technology. These national positions are not accidental. They are the product of sustained policy commitment over decades, and they translate directly into competitive advantage in the infrastructure of the AI economy.
The United States has the largest installed nuclear generation capacity in the world by absolute terms — approximately 100 gigawatts — but has built almost no new nuclear capacity in the past thirty years, has allowed its nuclear construction industrial base to atrophy, and faces regulatory approval timelines that are among the longest and most expensive in the world. The Bipartisan Infrastructure Law and the Inflation Reduction Act contain provisions designed to address this, including production tax credits for existing nuclear plants, investment credits for new nuclear construction, and funding for advanced reactor demonstrations. Whether those provisions are sufficient to reverse thirty years of underinvestment in the nuclear industrial base is a question whose answer will determine the extent to which the United States can capture the AI energy advantage that its technology companies’ lead in AI development would otherwise provide.
China is building nuclear capacity at a pace that no Western country is matching. Approximately 20 to 25 new reactors are under construction in China at any given time, using a combination of domestic designs and technologies licensed from Western and Russian suppliers. Chinese reactor construction timelines and costs are substantially below Western equivalents, reflecting the combination of a trained construction workforce, streamlined regulatory processes, and the economies of scale that come from building many reactors sequentially rather than one-off. China’s domestic AI industry is developing in parallel with its energy infrastructure buildout, and the combination — domestic AI capability built on reliable, low-cost, low-carbon nuclear power — is a strategic position that the United States and its allies will need to reckon with over the next decade.
— HANPRO Says
The nuclear renaissance story is one of the clearest examples of AI creating investment opportunities in industries that had nothing to do with software. Uranium miners, SMR developers, grid infrastructure companies, industrial cooling specialists — these are all beneficiaries of a dynamic that didn’t exist three years ago. The countries that figure out the energy-compute equation first won’t just lead in AI. They’ll have a structural advantage across every industry AI touches.
Related Reading:
Disclaimer: This article is for informational purposes only and does not constitute financial advice.
Author: Han Pro (gusungstar@gmail.com)
Copyright © GusungStar. All rights reserved. Unauthorized reproduction prohibited.
Leave a Reply