Energy & Climate Technology
The Smarter AI Gets, the More Power It Wants
Every time researchers made AI cheaper and more efficient, global energy demand for AI went up — and that is not a glitch, it is exactly how efficiency works.
The Idea
There is a 19th-century economic insight that keeps ambushing the AI industry. William Stanley Jevons noticed that as steam engines became more fuel-efficient, Britain did not burn less coal — it burned vastly more, because efficiency made coal-powered production so cheap that demand exploded. The same logic, now called Jevons paradox or the rebound effect, is playing out with AI at a civilisational scale. The intuition most people carry is straightforward: if each AI query costs half as much energy as before, we use half the energy overall. But that is only true if demand stays flat. When a resource gets cheaper, people use it more — often far more. Efficiency improvements lower the effective cost of doing something, which creates new users, new use cases, and new appetites that never existed before. For AI, the effect is unusually strong because inference — actually running a model to answer questions — accounts for roughly 80 to 90 percent of AI's ongoing energy consumption, according to multiple analyses in 2026. Training a frontier model is a large one-time cost, often tens of gigawatt-hours for a GPT-4-class run, with next-generation models estimated to exceed 100 gigawatt-hours. Dramatic as those numbers are, they are dwarfed over time by the cumulative weight of billions of daily queries. And as models get cheaper to run, the query volume grows — so the efficiency gain and the demand surge chase each other upward in lockstep. The uncomfortable implication is that building more efficient AI hardware is not, by itself, a path to lower total energy use. It may be the opposite.
In the World
In January 2025, a Chinese AI lab called DeepSeek released a model that reportedly rivalled the leading frontier systems at a fraction of the training cost. Markets briefly panicked, assuming demand for AI chips would collapse. Microsoft's CEO responded almost immediately by invoking Jevons paradox — the point being that cheaper AI would not reduce compute demand, it would unleash it. The data since then has borne that out. Gartner forecast in June 2026 that global data centre electricity consumption would reach 565 terawatt-hours in 2026, up 26 percent year-on-year, with AI-optimised servers accounting for roughly 31 percent of that total — up from roughly 20 percent the year before. The International Energy Agency confirmed that data centre electricity demand grew by 17 percent in 2025, with AI-focused data centres growing far faster still, surging 50 percent in the same year. The emissions scorecards of the companies running these systems tell the same story. Microsoft's 2026 sustainability report revealed that its greenhouse gas emissions rose 25 percent in fiscal year 2025, driven primarily by data centre expansion for AI. Google's electricity consumption rose 37 percent in 2025 — its largest single-year increase ever, pushing total electricity use up more than 250 percent since 2019. Amazon's carbon footprint climbed 16 percent. All three companies have ambitious climate pledges. All three watched their absolute emissions rise anyway. Meanwhile, the nature of AI queries is shifting in ways that amplify the energy cost further. The IEA's 2026 report shows that agentic tasks — where an AI plans, reasons, and calls on tools across multiple sequential steps — consume energy on a scale that dwarfs simple text generation. Each leap in AI capability tends to arrive bundled with a leap in energy appetite.
Why It Matters
Understanding Jevons paradox reframes the entire conversation about AI and climate. It means that celebrating a more energy-efficient chip or a cheaper model is not the same as celebrating lower energy use — and that efficiency metrics, though genuinely important, can be misleading if read in isolation. This does not mean AI efficiency work is pointless. It means the work is necessary but not sufficient. Getting more intelligence per watt matters enormously; it just does not automatically translate into fewer watts consumed overall, because demand is not a fixed quantity. Demand grows to fill the space that efficiency opens. For anyone making decisions about technology — whether building products, setting policy, or thinking about their own organisation's AI footprint — the honest question is not just 'how efficient is this model?' but 'how much usage will cheaper, faster AI unlock, and where does that energy come from?' Nuclear restarts, long-duration storage, and new grid infrastructure are being planned at least partly in response to this dynamic. Google itself acknowledged in its 2026 environmental report that its AI infrastructure buildout is accelerating faster than the grid is decarbonising. Efficiency buys time; it does not, on its own, solve the equation.
A Question to Ponder
If making AI cheaper reliably increases total AI energy demand, what would it actually take — technically, economically, or politically — to decouple intelligence from energy growth?
One reader's Thinkable for Tech Tuesday — a random draw from the library.
Start learning with Thinkable