The More Efficient AI Gets, the More It Eats

I built a small AI to cut energy waste. The research sent me down a rabbit hole about why efficiency keeps backfiring.

A while back I built LumiSentry, a little system that watches a room with a camera, works out whether anyone’s in it, and cuts the power to the lights when it’s empty. A small AI model does the looking, an Arduino does the switching. It works. It saves real electricity, and I’m proud of it.

I built it believing a fairly common thing: that AI is one of our better tools for fighting energy waste. Point smart software at a dumb problem, like lights burning in empty rooms, and you claw back energy nobody meant to use.

I still think that’s partly true. But the more I read while building it, the more I bumped into an idea that complicates the whole story, and once you see it you can’t unsee it. It’s called the rebound effect, and it might be the most important thing nobody mentions when they talk about AI saving the planet.

First, the part that’s real

Let me be fair to the optimistic version, because it isn’t nonsense.

AI does cut energy use in measurable ways. Smart grids that forecast demand can shave peak load. AI-tuned heating and cooling in buildings has shown energy reductions in the range of 15 to 30% in studies, including work that came out of Google’s DeepMind. Occupancy-based control, which is basically the grown-up version of what LumiSentry does, has cut HVAC use by a fifth or more in commercial buildings. These are real wins, and I’m not here to wave them away.

So if AI makes this fridge, this building, this grid more efficient, surely it adds up to less energy used overall?

This is exactly where it gets slippery.

The rebound effect, in one annoying example

Here’s the thing efficiency people have known about for over a century: making something more efficient often makes us use more of it, not less.

The classic case is lighting. Light bulbs have gotten staggeringly more efficient over the last hundred years, from candles to incandescents to LEDs. A modern LED sips a tiny fraction of the power an old bulb did for the same brightness. By the logic of efficiency, we should be spending almost nothing to light our lives now.

Instead we light everything. Streets, billboards, screens, empty offices, the undersides of kitchen cabinets. Lighting got cheap, so we used dramatically more of it, and total energy spent on lighting didn’t politely fall the way the efficiency math promised. Economists call this Jevons paradox, after a guy who noticed in the 1860s that more efficient steam engines made Britain burn through more coal, not less, because efficiency made coal-powered everything cheaper and more worth doing.

Efficiency, in other words, is not the same thing as using less. Sometimes it’s the engine of using more.

Which brings me back to AI

Now apply that to AI, and the picture gets uncomfortable.

The big trend in AI over the last couple of years has been making it cheaper and more efficient to run. Models got smaller for the same performance, chips got faster per watt, inference (the everyday running of a model, as opposed to the one-time training) got optimised hard. By the efficiency logic, AI should be getting lighter on the world.

The opposite is happening. As running AI got cheaper, we started running it everywhere, in every app, every search bar, every feature nobody asked for. The day-to-day running of models has overtaken training as the bulk of AI’s energy use, precisely because we now do so much more of it. Data centre electricity use is on track to roughly double by 2030, to around 945 TWh a year by the IEA’s reckoning, which is more electricity than the whole of Japan uses in a year. The more efficient each AI task gets, the more tasks we invent, and the total keeps climbing.

That’s the rebound effect wearing a very modern outfit. Efficiency isn’t braking AI’s energy appetite. It’s feeding it.

The bit that should make us cautious

Here’s what tips me from “interesting” to “we should be careful.”

The savings side of the AI-for-sustainability story is mostly soft. A lot of the efficiency figures come from the companies deploying the systems, not from independent measurement, and the independently verified, scaled-up savings are thinner than the headlines suggest. Researchers have started pointing this out directly. A widely cited 2025 paper put the rebound effect at the centre of AI’s environmental debate, arguing the field talks up the efficiency gains and quietly ignores the usage explosion they trigger.

Meanwhile the consumption side is hard, measured, and steep. We know fairly precisely how fast data centre demand is rising. We know it’s outpacing the efficiency wins. Which is why several analyses land on an awkward conclusion: over the next few years, AI is probably a net negative for energy, with its own growth swamping the savings it enables. Past 2030 it’s anyone’s guess, and depends entirely on choices we haven’t made yet.

So what about my little box?

LumiSentry still saves energy. I’ve watched it do it. At its small scale, the math works fine.

But building it taught me that the danger was never really the hardware. It’s the story we tell with it. “We’ve got AI handling efficiency now” is a deeply comforting sentence, and it’s exactly the kind of sentence that lets the rebound effect win. If smart software makes us feel like waste is a solved problem, we relax, we add more devices, more features, more compute, and the savings evaporate into all that new usage.

Efficiency only reduces energy use if it’s paired with a decision to use less. On its own it just makes using more affordable.

The part of LumiSentry I keep coming back to isn’t the AI that switches the lights. It’s the log, the boring record of how many hours rooms sat lit and empty. Because seeing the waste is what made me want less of it. No model required for that part. Just the willingness to look.

That’s the unglamorous truth under all of this. The thing that cuts energy use is using less of it. There’s no launch event for that, no demo, no neural network to show off. Efficiency is the easy part. Using less is the hard one, and it’s the only part that works.

A note on sources

The data centre and AI energy figures are from the International Energy Agency’s “Energy and AI” analysis (IEA, 2025): data centre electricity use is set to roughly double to about 945 TWh by 2030, close to 3% of global demand and more than Japan’s annual consumption, with inference now the larger share of AI’s energy use. The building and HVAC efficiency ranges (around 15 to 30%) come from recent studies on AI-driven building optimisation, including work associated with Google DeepMind. The rebound argument draws on Luccioni et al. (2025), “From Efficiency Gains to Rebound Effects,” and the wider literature on Jevons paradox. Worth flagging: many AI efficiency-savings figures are self-reported by the companies deploying the systems rather than independently measured, so treat the savings side as softer than the consumption side.

Sources:

• IEA (2025), Energy and AI. International Energy Agency, Paris. (data centre electricity doubling to ~945 TWh by 2030, the Japan comparison, inference now the larger share of AI energy use)
• Luccioni, A., et al. (2025), “From Efficiency Gains to Rebound Effects: The Problem of Jevons’ Paradox in AI’s Polarized Environmental Debate.” Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency. (the central rebound-effect argument)
• Luo, Jerry, et al. (2022), “Controlling Commercial Cooling Systems Using Reinforcement Learning.” ArXiv. (the Google DeepMind building-cooling savings)

Taken from Medhansh Kumar’s substack