Back to Insights

AI Is Scaling Up. Should Its Energy Footprint Scale Down?

AI & Data Centers8 min readUpdated

Energy per AI task is falling 10x a year, yet data center electricity is set to double by 2030. Why "intelligence per watt" may matter more than raw compute.

Compact AI data center at dusk with glass-walled liquid-cooled server halls, rooftop solar, battery storage, a substation, and heat-recovery pipes feeding a greenhouse

Key takeaways

  • Efficiency is improving faster than almost anything in energy history. The IEA says power per AI task is falling by at least 10x a year.
  • Total demand is rising anyway. Global data center electricity is projected to roughly double, from 485 TWh to about 950 TWh, by 2030.
  • That's the Jevons paradox at work: cheaper compute means more compute.
  • Reasoning models can use about 30x more energy per answer, so how we use AI matters as much as the chips.
  • The metric that matters is intelligence per watt: useful work delivered per unit of energy, water and grid capacity.

The paradox at the center of AI

Two things are true about AI and energy at the same time, and most of the debate only talks about one of them.

The first is that AI is getting dramatically more efficient. The International Energy Agency says power use per AI task is falling by at least an order of magnitude every year. It calls that pace unprecedented in energy history. Google reported that the energy for a median Gemini text prompt fell 33x in a single year, to about 0.24 watt-hours. That's roughly what a TV uses in under nine seconds. AMD says its AI energy efficiency is already 4x better than its 2024 baseline, ahead of its own target.

The second is that AI's total energy footprint keeps growing anyway. The IEA found that data center electricity demand rose 17% in 2025, and demand from AI-focused data centers jumped 50%. It projects global data center use will roughly double, from 485 TWh in 2025 to about 950 TWh by 2030. In the U.S., EIA's October 2026 Short-Term Energy Outlook expects record power consumption in both 2026 and 2027, driven largely by data centers.

Both statements are accurate. Understanding why they don't contradict each other is the key to the whole question. Should AI's energy footprint scale down as AI scales up?

Training gets the headlines. Inference pays the bills.

Most early coverage of AI energy focused on training: the months-long runs that build a frontier model on tens of thousands of GPUs. Training is energy-intensive, but it's periodic. A model is trained, then deployed.

Inference is different. It's every question, every summary, every coding request and every AI agent running in the background, all day, every day, for as long as the model is in use. As AI gets built into search, office software, customer service, phones and cars, inference becomes the larger and more permanent share of AI's electricity demand.

That changes what efficiency means. A 10x improvement per query sounds decisive. But if the number of queries grows 20x, total energy still doubles. The real question isn't how much energy one answer takes. It's how many answers the economy ends up asking for.

At the scale some AI campuses are now planned, with gigawatt-scale developments, data centers stop looking like commercial customers. They start looking like steel mills or refineries: industrial loads that shape regional generation and transmission plans. That's the heart of the AI Power Wall. You can see how quickly the numbers stack up with our Data Center Load Calculator.

Efficiency is winning. Demand is winning faster.

This is a textbook case of the Jevons paradox. In the 1860s, economist William Stanley Jevons noticed that more efficient steam engines didn't reduce Britain's coal use. They increased it, because cheaper energy made steam power worth using in far more places.

AI is following the same pattern. When the cost of an answer falls, companies put AI into more products. When inference gets faster, developers build agents that make dozens of model calls for a single task. Lower cost per token turns into more tokens, and usually more total energy.

There's a second twist. The newest reasoning models think before they answer, often generating long internal chains of steps. The AI Energy Score project, led by researchers at Hugging Face and Salesforce, tested 40 open models on the same hardware. Reasoning modes used about 30x more energy on average than the same models with reasoning turned off. For some models the gap was more than 100x.

So the energy footprint of AI depends heavily on how we use it, not just on how efficient the chips are. Sending a simple email summary to a heavy reasoning model is like driving a semi truck to pick up groceries. It works, but it's the wrong tool.

Data Center Electricity Is Set to Double

485 TWh
950 TWhAI-focused data centers: ~465 TWh (3x)
20252030 (projected)
Global data center electricity use. Source: IEA, April 2026.

Is bigger always necessary?

That raises one of the most useful questions in the debate. Does every task need a frontier-scale model?

Summarizing an email, sorting a document or pulling numbers from an invoice doesn't need the same horsepower as drug discovery or advanced math. That opens the door to right-sized AI: matching the model to the task instead of sending every request to the biggest model available.

The evidence is getting stronger. A 2025 Stanford-led study called Intelligence per Watt tested more than 20 small local models on a million real-world chat and reasoning queries. Local models could accurately handle 88.7% of them. The share of queries small models could handle locally rose from about 23% in 2023 to about 71% in 2025, and intelligence per watt improved 5.3x over the same period. Industry research cited by Data Centre Review suggests smaller, specialized models at the edge can cut energy use by 60% to 80% compared with large general-purpose models in the cloud.

That doesn't make hyperscale data centers go away. Frontier reasoning, scientific computing and video generation will still need them, and so will training. But it points to a more distributed AI system. Phones, laptops and on-site servers handle routine work, and the AI factories handle the hard problems.

What should we actually measure?

For years, the data center industry's headline efficiency metric has been PUE (Power Usage Effectiveness). It compares total facility energy with the energy that reaches the IT equipment. A PUE of 1.2 means 20% overhead for cooling, power conversion and lighting. It's a useful number, and you can test it with our PUE / AI Efficiency Calculator. But PUE only tells you how efficient the building is. It says nothing about whether the computing inside is doing anything useful.

That's why a new metric is gaining ground: intelligence per watt (IPW). Instead of counting raw operations or tokens, IPW asks how much useful, accurate work you get per unit of energy. A fast model that gives wrong answers scores poorly. A smaller model that gets the job done on a fraction of the power scores well.

IPW still needs a standard method before it can sit next to PUE in sustainability reports. But the direction is right. It shifts the conversation from "how much power can we get?" to "how much value do we get from the power we have?"

How We Measure AI Efficiency, and What Each Metric Misses
MetricWhat it measuresWhat it misses
PUE (Power Usage Effectiveness)Total facility energy ÷ IT energy, i.e. how efficient the building isNothing about whether the computing itself is useful or efficient
Energy per query (Wh/prompt)Energy for one AI response, e.g. Google's 0.24 Wh median Gemini promptTotal volume: billions of queries add up, and reasoning answers cost far more
Tokens per wattRaw text throughput per unit of energyQuality: fast wrong answers still count
Intelligence per watt (IPW)Accurate, useful answers or completed tasks per wattStill lacks a standard method across labs and hardware
Total MWh and peak MWWhat the grid and the community actually feelValue: says nothing about what the energy produced

Electricity isn't the only footprint

AI's physical footprint goes beyond kilowatt-hours. Data centers use water for cooling, either directly through evaporative systems or indirectly through the power plants that serve them. They also consume copper, transformers, chips, concrete and land. All of these have supply chains and local impacts of their own.

As reporting in Transformer put it, AI expansion may not be one big environmental problem so much as lots of smaller environmental problems. Those include a stressed aquifer in one county, a transmission line through another, a gas plant extended in a third. Those local impacts are why the industry now faces a public trust problem in many communities, and why states are rewriting the rules on who pays for grid upgrades.

In the U.S., Lawrence Berkeley National Laboratory estimated that data centers used about 4.4% of the nation's electricity in 2023. It projected that share could reach 6.7% to 12% by 2028. At the top of that range, efficiency becomes a matter of grid reliability and household bills, not just corporate sustainability.

What scaling down could look like

Scaling AI's footprint down doesn't mean scaling AI back. It means getting more useful output from every megawatt, every gallon and every grid connection. In practice, that looks like:

  • Model routing. Send simple tasks to small models and save frontier reasoning for problems that need it.
  • Smarter reasoning budgets. Let models think hard only when the task calls for it, instead of by default.
  • Better hardware and cooling. Liquid cooling, higher-efficiency chips and lower PUE all compound. Run your own scenarios with the GPU & Compute Power Load Calculator and the AI Training Energy Calculator.
  • Flexible operation. Shift training and batch jobs away from grid peaks. Use on-site storage to shave demand. Battery energy storage can make a data center a better grid citizen; test the economics with the Behind-the-Meter BESS ROI Calculator.
  • Heat reuse. Recover waste heat for district heating, greenhouses or industry instead of venting it.
  • Transparency. Report energy, water and intelligence per watt, not just total capacity and PUE.

The real question isn't "does AI use too much energy?"

That framing is a dead end. Energy use alone never decides whether an activity is worth it. Hospitals and chip fabs use enormous amounts of electricity and water, and few people question it, because the value is clear.

The better question is whether the value AI produces justifies the resources it uses, and whether those resources are being used as efficiently as possible. Medical research, grid optimization and materials science will clear that bar easily. Some other uses of the same compute may not. As AI moves from novelty to infrastructure, that distinction will matter more. As we asked in "Is compute the new oil?", the resource only matters as much as what we do with it.

Conclusion: it's about what we choose to scale

AI is going to scale. The more important question is what gets scaled. If progress is measured mainly in parameters, GPUs and gigawatts, infrastructure demand could keep outrunning efficiency gains. If the industry focuses on getting the most useful intelligence from every unit of electricity, water and capital, the AI buildout could look very different a decade from now.

The physical world sets real limits on the digital one. Transmission lines can't be built at software speed. Copper can't be downloaded. Communities aren't empty spots on a site-selection map. The next breakthrough in AI may not be a bigger model. It may be learning to do more with less.

This article was originally inspired by Chris Middleton's 'AI Energy Usage — Is the Industry Scaling Up When It Should Be Scaling Down?' (diginomica) and Amit Katwala's 'AI Expansion May Not Be One Big Environmental Problem. But It'll Be Lots of Smaller Ones.' (Transformer). Updated November 2026 with new IEA, EIA, Stanford, Google, AMD and AI Energy Score data.

Chris Kalowes

Founder, WattThe?!

Energy intelligence. Simplified.

FAQ

How much energy does one AI prompt use?+

It varies widely. Google reported that a median Gemini text prompt used about 0.24 watt-hours in 2025, down 33x in a year. Reasoning models can use far more. The AI Energy Score project found that reasoning modes averaged about 30x more energy than the same models without reasoning.

What is the Jevons paradox, and why does it matter for AI?+

The Jevons paradox is when making a resource more efficient lowers its cost enough that total use goes up. AI energy per task is falling by about 10x a year, but cheaper AI gets used far more, so total data center electricity is still projected to roughly double by 2030.

What is "intelligence per watt"?+

Intelligence per watt (IPW) measures useful, accurate output, such as correct answers or completed tasks, per unit of energy. Unlike PUE, which only measures building efficiency, or tokens per watt, which ignores quality, IPW tries to capture the real value delivered per watt.

Can small AI models really replace large ones?+

For many everyday tasks, yes. A Stanford-led study found small local models could accurately answer 88.7% of a million real-world single-turn chat and reasoning queries. Large frontier models are still needed for complex reasoning, research and training new models.

How much electricity will data centers use by 2030?+

The IEA projects global data center electricity use will roughly double, from about 485 TWh in 2025 to about 950 TWh in 2030, with AI-focused data centers tripling to about 465 TWh. In the U.S., Lawrence Berkeley National Laboratory projected data centers could use 6.7% to 12% of national electricity by 2028.