AI is growing faster than anything we have built. Now we have to choose what computing is for.

The International Energy Agency expects data centres to use 945 terawatt-hours of electricity by 2030, up from 415 in 2024, which is about 15% more every year, compounding, enough to double demand roughly every five years, with AI as the main driver.¹

The computing used to train leading AI models has grown about four to five times a year for over a decade,⁴ and measured from each one's first mass-market launch, people took up generative AI faster than they took up personal computers or the internet.⁵ The campus going up in Abilene, Texas, is designed for about 1.2 gigawatts,³ roughly the output of a large power station, and electricity on that scale is electricity that homes, hospitals and the switch away from fossil fuels are also queuing for.

A 2026 preprint that models the IEA's high-growth case puts the extra operational emissions of AI data centres at 0.8 billion tonnes of CO₂ a year by 2035, and manufacturing the servers adds about 0.2 billion tonnes more at the median estimate.³⁰

A data centre needs a grid connection and a planning permit, so a council can see one coming and say no. Nobody approves the ad auction that runs every time a news page loads,⁹ and if we only argue about the new buildings we will miss most of the system they plug into.

A study of browsing, social media, streaming and video calls put the average internet user at 229 kg of CO₂e a year, about 40% of a personal carbon budget compatible with 1.5 °C, before anyone has asked a chatbot a single question, and much of it comes from manufacturing the devices themselves.⁶

Part of that load exists to sell attention. Devices use a median 34% more energy on websites that run real-time ad auctions,⁹ and tracking adds at least 21% to the data a page transfers,¹⁰ so readers pay for the advertising industry's electricity on their own phones and laptops. The only whole-system estimate of online advertising put its 2016 electricity use somewhere between 20 and 282 TWh,⁷ a fourteen-fold range that says less about advertising than about how little anyone has looked.

About half of all web traffic in 2024 came from bots, according to Imperva, a company that sells bot protection.¹⁴ The same advertising study estimated that about a quarter of ad traffic was fraudulent,⁷ which means computing spent serving ads to machines and then more computing spent catching them.

Spam shows who pays. Economists estimated in 2012 that it cost American firms and consumers almost $20 billion a year while earning spammers and their merchants about $200 million worldwide, a ratio of roughly 100 to 1,²⁸ and a study that traced real spam across the internet found that networks working together could filter up to 70% of it at the origin, before it travels anywhere.²⁹

Some computing creates more computing. An ad campaign sets off an auction on every page it reaches and fraud filtering after that, and nobody has measured how much demand those knock-ons add. The scenarios below use placeholder multipliers to show the shape of the problem until someone does.

The usual answer is efficiency, and the record backs it, with data-centre workloads growing about 550% between 2010 and 2018 while their energy use barely moved.^{33, 11}

Not every early warning was right. In 1999 a much-quoted report claimed the internet was already a heavy drain on US electricity, and researchers at Lawrence Berkeley National Laboratory showed it had overstated the figure by about a factor of eight.³¹ The concern behind it was sound all the same. In August 2007 the US Environmental Protection Agency told Congress that data centres had used about 61 billion kilowatt-hours in 2006, 1.5% of all US electricity and more than double their use in 2000, and that existing technology could already cut a typical server's energy use by a quarter or more.¹⁹ In 2024 data centres used about 1.5% of the whole world's electricity.¹

In 1865 the economist William Stanley Jevons noticed that more efficient steam engines had led Britain to burn more coal, because cheaper energy services invite more use, and economists now call that the rebound effect, or backfire when the extra use outweighs the saving.¹⁶

No one has proved that Jevons' paradox holds for computing. The researchers who reconciled the main estimates of the sector's footprint found no sign that efficiency has ever cut its total emissions, and concluded that only a cap or a carbon price would.¹¹

Governments had adopted two climate treaties by the time the internet's data-centre industry was taking off in the late 1990s.^{24, 25} The people who built it knew enough to take energy seriously, and the choice they made, again and again, was to put growth, cost and speed first. Choose what happens to a 30% efficiency gain.

Computing is the newest curve in what Earth-system scientists call the Great Acceleration, the surge since about 1950 in energy use, fertiliser, transport and almost every other measure of human activity.²⁶ Nobody serious answers that curve by proposing a return to a pre-industrial farm, and nobody needs to switch off the internet to bend this one, because in both cases the real question is what the growth is for.

Computing is finite in practice, since every server needs a grid connection and every grid sits inside a carbon budget, and right now we let price share it out, so a kilowatt-hour goes to whoever earns the most from it and social value never enters the sum.

Elinor Ostrom won a Nobel prize for showing that the people who share a resource can run it well without handing it to either the market or the state, provided they write the rules themselves and can see who breaks them.¹⁸ Computing could be governed that way, and the first fight would be over what counts as essential.

Below, you have 100 units of the world's computing to divide between 18 uses. Count each use by its purpose, so AI helping with a diagnosis is health computing and AI writing 10,000 marketing emails is advertising. A use set to zero still shows as a rounding error, because no budget will ever stop any of them entirely. Pick a governance approach to start from, or set your own.

Most global figures in this field come from models, and direct measurements exist mainly for single devices or single websites, so each card says which kind of figure it is.

Why the numbers disagree

One team measured the extra 6.13 watts a PC draws when ads and trackers load on news sites.⁸ Another modelled the whole system behind online advertising and got tens to hundreds of terawatt-hours.⁷ Setting one beside the other is like comparing a car's fuel bill with the oil industry's.

Where you draw the boundary decides the answer. ICT's share of global emissions in 2020 comes out at 1.4% on industry data¹² and up to 3.9% once the full supply chain is counted,¹¹ and a 2026 study found 42% of the emissions embodied in digital industries sitting in other sectors' supply chains, where standard greenhouse gas accounts never show them.¹³ Allocation moves the numbers too, because network energy barely changes with traffic in the short run, so a per-gigabyte figure overstates what you save by sending less data.¹⁵

The gaps

These gaps are probably widening. Generative AI makes spam, marketing copy and tracking code cheaper to produce at scale, so we expect it to be adding to all three and to the filtering that cleans up after them, and nobody is yet counting what that adds. The synthetic content flood scenario in section 2 shows the shape of it with placeholder numbers.

Seven testable claims

1. Ad tech uses at least as much energy on servers and networks as on users' devices.
2. Real-time bidding adds more energy per page view than any other common web feature.
3. Recipients' systems spend more of the energy in marketing email than senders' systems do.
4. Filtering abuse takes a material share of the energy of the services it protects.
5. Deleting unused data saves little energy in the year it happens and avoids building capacity later.
6. People given a fixed budget spend far less on attention than the market does today.
7. Generative AI is raising the volume of spam, automated marketing and tracking, and the computing spent filtering them.

The full literature review covers 40 sources, separates peer-reviewed work from industry reports, and is available from the author.

Open your organisation's responsible AI principles and look for a sentence about how much computing a team may use, and on what. If there isn't one, nobody has decided, and the default is whatever the tools and their vendors make easy.

That default has a measurable cost. Researchers who measured the energy of 1,000 inferences found general-purpose generative models orders of magnitude more expensive than models built for a single task, even at the same size.³²