Ethics & Governance

The Power Reckoning: How AI's Hunger for Electricity Is Rewriting the Rules of the Global Energy System

AI's electricity demand has stopped being a projection. Hyperscaler capex guidance for 2026 runs past $700bn, PJM capacity costs hit a record $16.4bn, and Google's footprint is 81% above 2019.

Gabriele Masetti · 2026-09-17

In 2005, Google published a paper noting that its data centres consumed roughly as much electricity as 200,000 average American homes. At the time, this seemed remarkable — a small technology company burning through the electricity budget of a medium-sized city. The paper was a statement of accountability, a pledge to take energy seriously.

Twenty years later, the scale of the problem makes that 2005 figure look quaint.

According to the International Energy Agency, global data centre electricity consumption reached 415 terawatt-hours in 2024 — roughly equivalent to the annual electricity consumption of France. AI-specific workloads accounted for an increasingly large fraction of that total, and their share was growing faster than any other category of demand. In its Energy and AI report the agency put the 2030 figure at around 945 terawatt-hours, more than double the 2024 level and slightly more than the whole of Japan consumes in a year.

The numbers are so large they have become abstract. But they describe a real and unprecedented strain on physical infrastructure: power grids designed for a different era, transmission systems that cannot quickly expand to meet surging demand, communities where data centres now compete with households and factories for limited grid capacity, and a global climate system that does not adjust its mathematics based on the strategic importance of the technology consuming the energy.

The power reckoning has arrived. How the AI industry navigates it — through efficiency improvements, new energy sources, regulatory accommodations, or some combination — will shape not just the economics of AI but the politics of energy, the pace of the energy transition, and the geographic distribution of who controls the world's most powerful computational infrastructure.

The Scale of the Demand

To understand the magnitude of what AI is demanding from the global energy system, it helps to start with the hardware.

A modern AI training cluster — the infrastructure used to train large frontier models — consists of tens of thousands of high-end AI accelerator chips. Take the generation that built the current crop of frontier models. Nvidia's H100, the dominant training chip through 2024, draws 700 watts. A cluster of 32,000 of them — a size typical of major foundation model training runs that year — would draw approximately 22 megawatts continuously, and a run lasting 90 days would consume roughly 48 gigawatt-hours of electricity.

For reference, 48 gigawatt-hours is approximately the annual electricity consumption of 4,300 average American households. A single AI training run consumes, in three months, what thousands of families use in a year.

Those are now historical numbers, and they have not shrunk. Nvidia announced on 31 May 2026 that its Vera Rubin platform had entered full production, with shipments due from the autumn and instances promised at AWS, Google Cloud, Microsoft, Oracle and a set of specialist cloud providers. The company's headline claim is up to ten times lower cost per token than Blackwell, the generation before it. Cheaper tokens have never yet meant a smaller electricity bill, because the racks that deliver them are denser and the campuses built around them are larger.

And training is only part of the picture. Inference — running trained models to respond to user queries — consumes significantly more electricity in aggregate than training, because it happens continuously at massive scale. ChatGPT passed 900 million weekly active users in February 2026, and OpenAI said on 31 July 2026 that its models reached more than a billion active users. Claude, Gemini, Copilot, and dozens of other AI services collectively handle billions of queries per day. Each query consumes energy. The cumulative inference demand of deployed AI systems dwarfs the training demand, and it grows every time a new user adopts an AI service or a new AI-powered application goes live.

Per query, the numbers look small. Google published a measurement in August 2025 putting the median Gemini text prompt at 0.24 watt-hours of energy, 0.03 grams of CO2 equivalent and 0.26 millilitres of water — a figure that covers accelerators, host machines, idle capacity and data centre overhead, and that has not been through peer review. The argument does not turn on whether a prompt costs a quarter of a watt-hour or three. A cheap prompt, taken a billion times a week, is a power station.

The Data Centre Surge

Data centres are the physical infrastructure where all of this computation happens, and the surge in AI demand has triggered a data centre construction boom unlike anything the industry has seen.

In 2024, Microsoft announced plans to spend 80 billion dollars on new data centres over the following year. Amazon Web Services, Google Cloud and Meta each announced comparable investments. Amazon, Microsoft, Alphabet and Meta between them spent about 410 billion dollars in 2025, and their guidance for 2026 — raised repeatedly through the year — totals somewhere between 700 and 760 billion, depending on which tracker is counting and how late the revisions are caught. Either way it is roughly a three-quarters increase in twelve months, and the largest infrastructure investment surge in the history of the technology industry.

The investment is creating a physical geography of AI. The largest data centre clusters are concentrating in locations with access to reliable power, cheap land, and cooling resources. Northern Virginia — already the largest data centre market in the world — is expanding at a pace that has local utilities warning about capacity constraints. Phoenix, Arizona, is grappling with both electricity demand and water use for cooling, in a desert region already under severe water stress. Ireland, where several major technology companies have built large facilities, has watched data centres take a quarter of its power: the Central Statistics Office put their share of metered electricity consumption at 23 percent in 2025, against 5 percent in 2015.

The geographic concentration is not random. It reflects the availability of existing grid infrastructure, proximity to fiber optic backbone networks, land costs, and access to renewable energy. But the concentration creates its own problems: grid infrastructure designed for distributed residential and industrial loads is not well-suited to serving a small number of enormous, power-hungry facilities. The transmission and distribution systems that bring power to the largest data centre clusters require expensive upgrades, and those upgrades take years to complete.

The Grid Is Not Ready

The United States electrical grid was largely designed and built in the mid-twentieth century. It consists of a patchwork of regional grids, operated by different utilities under a complex mix of federal and state regulation, connected by transmission lines that carry power from generators to load centres.

The system worked well for the demand patterns it was designed to serve: residential demand peaking in the morning and evening, industrial demand concentrated in specific zones, commercial demand distributed across many buildings. It was not designed for the demand pattern that AI data centres represent: massive, continuous, highly concentrated load growth in specific locations, on timescales that compress decades of historical demand growth into months.

In Northern Virginia — the so-called Data Centre Alley that hosts more data centre capacity than any other region in the world — utilities began issuing capacity warnings in 2023. The clearest evidence that the warnings were real arrived as a price. PJM, the regional transmission organisation serving the mid-Atlantic states, runs an auction to secure future generating capacity. Capacity for the 2024/25 delivery year cleared at $28.92 per megawatt-day; for 2026/27 it cleared at $329.17, more than ten times as much. In the auction held in December 2025, total capacity costs reached $16.4 billion and PJM came up 6,625 megawatts short of its own reliability requirement — the first shortfall in the market's eighteen-year history. PJM's independent market monitor attributed $6.5 billion of that bill, 40 percent of it, to data centre load.

Those costs do not stay in the wholesale market. They arrive on household bills, which is why data centre siting has stopped being a planning question and become an electoral one.

The constraints are not just in the United States. In Ireland, the grid operator Eirgrid placed a moratorium on new data centre connections in parts of the country, citing concerns about the impact of data centre demand on grid stability and the risk of exceeding the country's energy security limits. In Germany, grid operators warned that the combination of data centre growth, electric vehicle adoption, and industrial electrification was pushing transmission infrastructure toward its limits.

The fundamental problem is a mismatch of timescales. Building a data centre takes 18 to 24 months. Building the grid infrastructure to power it — new transmission lines, substations, generation capacity — takes 5 to 10 years, and in some jurisdictions significantly longer due to permitting and regulatory processes. Technology companies that want to build data centres today are competing for grid capacity that was committed years ago, and the new capacity they need will not be available for years to come.

The Nuclear Bet

The electricity demands of AI have done something that climate advocates had struggled to achieve for decades: they have made nuclear power economically attractive again to major technology companies.

The appeal of nuclear is straightforward from an AI company's perspective. Nuclear power plants generate electricity continuously, regardless of weather conditions, at high capacity factors. They produce no carbon emissions during operation. They are located on specific sites that can be purpose-connected to data centre facilities. And their output is predictable and reliable in a way that wind and solar, however fast-growing, are not.

Microsoft signed a deal in 2024 to purchase all the power output from a restarted reactor at Three Mile Island in Pennsylvania — the site of the 1979 accident that defined public anxiety about nuclear power for a generation. The unit that melted down was the other one; Unit 1 ran until 2019 and shut because cheap gas had made it uneconomic. Constellation, its owner, has renamed it the Crane Clean Energy Center, committed about 1.6 billion dollars to the restart and contracted its full 835 megawatts to Microsoft for 20 years. In November 2025 the Department of Energy's loan office added a 1 billion dollar loan, and the target restart date moved forward to 2027.

The symbolism was not lost on anyone: the facility whose meltdown became a cultural touchstone for nuclear fear is being brought back, with federal credit, to run inference.

Google signed long-term agreements to purchase power from small modular reactors (SMRs) being developed by Kairos Power, with expected delivery beginning in 2030. Amazon Web Services announced an investment in X-energy, another SMR developer. Meta, OpenAI, and several other AI companies have engaged with nuclear energy providers and developers.

Small modular reactors are a particularly interesting solution for the AI industry because they can, in principle, be sited adjacent to data centre facilities, eliminating the need to transmit power over long distances from large central generating stations. A 300-megawatt SMR located on a data centre campus would provide roughly the power needed for a major AI computing facility, with minimal grid interaction.

The commercial deployment of SMRs is still several years away for most vendors, but the financial commitments from technology companies are accelerating development timelines. The AI industry's demand signal has done more to catalyse investment in new nuclear power development than decades of climate advocacy and government subsidy programmes.

The Renewable Reality

While nuclear captures attention, the majority of new power capacity being connected to data centres is renewable — primarily solar and wind.

Microsoft, Google, Amazon, and Meta have all made commitments to match their data centre electricity consumption with renewable energy purchases. These commitments have driven massive investment in solar and wind development, and the technology companies are now among the largest corporate purchasers of renewable energy in the world.

But matching electricity consumption with renewable energy purchases is not the same as actually powering data centres with renewable energy in real time. Renewable energy is intermittent. Solar panels produce power only when the sun is shining. Wind turbines produce power only when the wind is blowing. Data centres, by contrast, consume power continuously, 24 hours a day, seven days a week.

The gap between matching and actually running on renewable power 24/7 is the central challenge of the clean AI data centre. Google has been the most ambitious in trying to close this gap, with a stated goal of operating on carbon-free energy hour-by-hour by 2030. Achieving this requires a combination of on-site generation, long-duration energy storage, geographic dispersion of computing load to follow available renewable generation, and flexible load management that shifts workloads to times and places where clean power is available.

The technology to do this at scale does not yet exist at commercially viable cost. Long-duration energy storage — the ability to store electricity for days or weeks during periods of abundant renewable generation, to discharge during periods of shortage — remains expensive and limited in deployment. The grid management systems needed to optimise the geographic distribution of AI workloads in response to renewable availability are in early development. The full 24/7 clean power vision is a real aspiration, but one that is likely a decade or more away from practical realisation at the scale that major AI data centres demand.

The Efficiency Race

The power demands of AI are not a fixed physical constant. They are the product of specific technological choices — hardware architectures, data centre designs, model sizes, inference methods — that can be changed.

Efficiency improvements in AI hardware have historically been rapid, and every generation is sold on the gain over the one before: Hopper over Ampere, Blackwell over Hopper, and now Rubin over Blackwell, where Nvidia's claim is up to ten times lower cost per token. These are vendor numbers, measured on workloads the vendor chose, and they deserve the scepticism that vendor numbers earn. The direction is not in dispute. Performance per watt in AI accelerators has improved faster than the Moore's law curve for general-purpose computing, precisely because AI workloads are more amenable to architectural specialisation.

Model efficiency improvements have been equally significant. The post-training revolution described in the previous article in this series has produced models that match the performance of earlier frontier systems at substantially lower inference cost. Techniques like knowledge distillation — training small models to mimic the behaviour of large ones — have reduced the compute required per query by factors of five to ten for many common use cases without significant quality loss.

Speculative decoding, a technique where a small model generates candidate completions that are then verified or rejected by a larger model, has improved inference throughput substantially. Mixture-of-experts architectures, which route each input to a subset of model parameters rather than activating the full model, reduce the effective compute per query for models that might otherwise be very expensive to serve.

The net result of these efficiency improvements is that the absolute energy consumption per AI query has been declining even as the number of queries grows. But the volume growth has, so far, substantially outpaced the efficiency gains. Every efficiency improvement that lowers the cost of AI also increases the number of use cases that become economically viable, and the resulting demand growth has more than absorbed the efficiency savings. The rebound effect — known as the Jevons paradox — is a persistent feature of technology economies and suggests that efficiency alone is unlikely to solve the energy problem.

The Climate Contradiction

For technology companies that have made ambitious net-zero commitments, the energy demands of AI create a painful public contradiction.

Google's 2026 environmental report, published at the end of June, disclosed that the company's total carbon footprint rose 18 percent in a single year and now sits about 81 percent above its 2019 baseline. Its electricity consumption grew 37 percent year on year, the largest single-year increase it has recorded. Microsoft's 2026 sustainability report, covering fiscal 2025, put its combined scope 1, 2 and 3 emissions up 25 percent year on year, to roughly 20.3 million tonnes of carbon dioxide equivalent — six years after it pledged to be carbon negative by 2030. Meta, Amazon and the rest face the same arithmetic.

The companies respond to this contradiction in several ways. They point to the efficiency improvements being driven into AI hardware and software. They point to the massive renewable energy purchases they are making, which are funding new clean generation capacity. They argue that the economic value created by AI — and its potential role in accelerating solutions to climate change itself — must be weighed against the near-term emissions.

That last argument deserves scrutiny. There is genuine evidence that AI is accelerating scientific progress on climate-relevant problems: materials discovery for better batteries, climate modelling, optimisation of energy systems, design of more efficient buildings. But the argument that AI will eventually solve the energy problems its training creates is a form of temporal discounting that regulators and climate campaigners are beginning to challenge. The emissions happen now; the benefits are speculative and future.

The more honest position, adopted by some researchers and a few technology company executives, is that the AI industry has a genuine emissions problem that efficiency improvements and renewable energy purchases will not fully solve on the timescales demanded by climate science — and that this is a fact that should inform both the pace of AI deployment and the design of climate policy.

The Geopolitics of Compute

The concentration of AI infrastructure is not just an energy problem. It is a geopolitical one.

The countries and regions that host the largest concentrations of AI data centre capacity are acquiring a form of strategic power that was not anticipated when the infrastructure was being planned. Ireland, which has become a major European data centre hub due to its corporate tax environment and renewable energy resources, now hosts computing infrastructure that processes a significant fraction of European digital activity — creating dependencies and vulnerabilities that Irish and European regulators are beginning to examine carefully.

The United States government has taken an active interest in the geographic distribution of AI compute as a national security matter. Export controls on high-end AI chips are designed to prevent competitors from acquiring the hardware necessary to build comparable AI infrastructure. The Chips and Science Act invested tens of billions of dollars in domestic semiconductor manufacturing to reduce dependence on Taiwanese production. The implicit logic is that AI compute infrastructure has become a strategic resource, like oil refineries or semiconductor fabs, that cannot be allowed to concentrate in geographically vulnerable locations or fall under foreign control.

China, meanwhile, is pursuing a parallel strategy. Despite US chip export controls that have limited its access to the most advanced AI accelerators, Chinese companies have developed domestic alternatives and are deploying AI infrastructure at massive scale. The energy implications within China are significant: major AI data centres in the country's interior are co-located with coal-fired power plants, prioritising energy access over carbon accounting.

The energy geography of AI is thus becoming entangled with the geopolitics of AI in complex and consequential ways. Where AI infrastructure is built, who owns it, what energy it uses, and what sovereignty applies to the data it processes are questions that governments are beginning to answer explicitly rather than leaving to market forces.

The Horizon

The power reckoning for AI will not be resolved quickly or easily. The trajectories are too steep, the infrastructure timescales too long, and the competitive pressures too intense.

But several developments over the next five to ten years will shape how the story ends.

The deployment of small modular reactors beginning in the early 2030s could provide clean, reliable power for AI data centres without the intermittency challenges of renewable energy. The first commercial SMR deployments are expected in the 2028 to 2032 timeframe for the most advanced vendors, and the AI industry financial commitments are pulling forward those timelines.

Continued improvement in AI inference efficiency — driven by hardware advances, model compression, and speculative decoding — will reduce the energy cost per query even as the number of queries grows. Whether the rebound effect absorbs those efficiency gains remains the central uncertainty.

Grid modernisation programmes in the United States, Europe, and other major markets are being accelerated by a combination of AI-driven demand, electric vehicle adoption, and the energy transition. The infrastructure that is being built today will determine the energy geography of AI for decades.

And the regulatory environment has moved from drafting to application. Under Article 53 of the EU AI Act, providers of general-purpose AI models must document the computational resources used in training and the energy that training consumed, estimating where no measurement exists. The obligation has applied to new models since 2 August 2025; models already on the market by that date have until 2 August 2027. It sets no ceiling on consumption — but it produces the first set of comparable, published figures, and comparable figures are what regulation is eventually built from.

Several US states have introduced or are considering legislation requiring data centres to meet specific efficiency standards or renewable energy requirements. Carbon pricing mechanisms, if they strengthen as many climate economists advocate, will internalise some of the environmental costs of AI energy consumption that are currently externalised onto the atmosphere.

What Intelligence Costs

AI is the fastest-growing consumer of electricity in history. It is not a problem that the technology industry can innovate its way out of entirely. It requires confronting some genuinely hard trade-offs: between the speed of AI deployment and the pace of clean energy development; between the strategic imperative to maintain AI leadership and the climate imperative to decarbonise the electricity system; between the economic value created by AI and the environmental costs that the market, left to its own devices, will not adequately price.

Those trade-offs do not have easy technical solutions. They require choices — about what to build, where to build it, how fast to deploy it, and who bears the costs and benefits. They require the AI industry, the energy industry, governments, and communities to negotiate a new relationship between computation and power.

The power reckoning is, in the end, a reckoning about what intelligence costs. The answer, it turns out, is a lot — and the bill is due.