Concepts explained
AI Is a Power Hog — So What Is the Answer?
One data centre, they say, uses as much electricity as a city; in a few years, as much as a whole country. That is broadly true. But once you see why it eats so much, the answer splits three ways: make more, make it eat less, or use less. All three are being built in Daejeon.
How much does it eat?
A large data centre uses about as much electricity as a city of several hundred thousand people. That is not an exaggeration, and such centres are multiplying every year. Data centres worldwide now draw around two per cent of the world's electricity, and most forecasts have that doubling by about 2030.
And the thank-you? One question to an AI costs about as much as running a light bulb for a few minutes. Ten questions a day for a year comes nowhere near what a refrigerator uses.
Why so much?
What a computer does, in the end, is flip switches. Each flip costs a very small amount of electricity. To answer one question an AI flips them trillions of times, and a small thing done a trillion times is a large thing.
But most of the electricity does not go into the switches. Today's computers keep memory in one place and calculation in another. To calculate, a number must be fetched from the memory warehouse, carried to the calculator, and the result carried back. The carrying costs far more than the arithmetic.
AI carries an unusual amount: hundreds of billions of numbers in and out of the warehouse. The road between warehouse and calculator jams; to widen it more chips are added; the added chips run hot, and cooling them costs electricity again. Three or four tenths of a data centre's power goes not into computing but into keeping it cool.
The brain runs on twenty watts
Here an odd comparison appears. The human brain uses about twenty watts of — one light bulb. On twenty watts it understands speech, recognises faces and writes. An AI doing the same work uses tens of thousands of times more.
The brain is cheap because it has no separate warehouse and calculator. Each nerve cell both remembers and computes. There is nothing to carry, so nothing is spent on carrying. The cost that eats most of a computer's power is a cost the brain never had.
So the answer splits three ways
When electricity runs short there are only three things to do: make more, build machines that eat less, or use less.
- Make more. Fusion belongs here: it does not care about location and gives a great deal of power from a small footprint, so the picture is one plant beside each data centre. But it is an answer for the 2030s and beyond.
- Build machines that eat less. Chips that, like the brain, remember and compute in the same place — called neuromorphic. Being built at KAIST in Daejeon.
- Use less. Getting the same answer out of a smaller AI. This is already happening fast: the electricity needed for the same task is falling by a factor of several every year.
No one of the three is the answer. The first is close to twenty years off, and the second and third have to fill that gap.
A little further in
So what is real and what is inflated?
A large one matches a city of several hundred thousand, and their number grows every year.
All data centres together draw around two per cent today. But they are growing faster than any other sector. The problem is not the size but the slope.
A few minutes of a light bulb. Not something individuals fix.
If it arrives, yes. But that is the 2030s or later, and the twenty years before it are where this problem actually lives.
The direction is right and laboratory chips show such numbers. But nobody yet knows how to make those chips do everything today's AI does. The chip exists; the AI that fits it does not.
Where the structure came from
Keeping memory and calculation apart was set out by von Neumann in 1945: the program sits in the memory warehouse and the calculator fetches it piece by piece. It was so convenient that nearly every computer for eighty years was built this way.
That convenience has now become an electricity bill. It caused no trouble for eighty years because the road between warehouse and calculator was never that busy. When AI began crossing it trillions of times, the structure itself became the bottleneck for the first time — and so, after eighty years, a different structure is being sought.
Building the brain's way is not new either. In the late 1980s Carver Mead proposed imitating nerve cells in , and coined the word neuromorphic. A forty-year-old idea, called back because of the price of electricity.
In Daejeon
Two of the three answers live in Daejeon. Making more is KSTAR in Yuseong-gu; eating less is KAIST, in the same district.
KAIST's materials department has built a device imitating the way a nerve cell remembers its past activity and grows more or less sensitive on its own, and an artificial neuron that uses the noise in a semiconductor for signal processing rather than suppressing it. Its electrical engineering department made one transistor serve as both nerve cell and connection. The number these people always quote is that twenty watts.
And there is a curious loop. AI has begun to be used to hold KSTAR's plasma. AI eats electricity; fusion is a candidate to make it; fusion needs AI to be held. In the same neighbourhood, each needs the other.
What you would study to do this
- Someone who designs chips that eat less — semiconductor design and materials. The people building brain-like devices.
- Someone who gets the same answer from a smaller AI — computer science. The seat cutting electricity fastest right now.
- Someone who brings power to the data centre — power engineering. New data centres are often held up not by chips but by a shortage of wires.
- Someone who cools it — refrigeration and air handling. Three or four tenths of the power goes to cooling, so this person is also someone who cuts electricity.
- Making more — the people listed in the fusion piece.
The fourth line is the unexpected seat. Say AI and electricity and everyone thinks of chips, yet three or four tenths of the power goes to cooling hot chips. It is a trade of chillers, pipework and airflow, and the refrigeration and pipefitting certificates from a technical high school lead straight to it. The same kind of seat as the fridge-keeper in the fusion piece. Beside every technology that makes heat stands someone who takes it away.
So
AI eats a great deal of electricity: true. But if the only answer is fusion, we sit idle for twenty years. Those years must be filled by chips that eat less and ways of using less, and both are already moving.
And quantum computers have nothing to do with this. As the previous piece showed, they use almost no power and do not replace AI computation. The three are often named in one breath; in front of an electricity meter they stand in entirely different places.
The question that remainsBetween making more and using less, which will look like the bigger answer twenty years from now?