The Ethical Investor: DeepSeek could finally tame the power-hungry AI beast... or could it?

AI’s energy hunger grows as tech booms
DeepSeek’s low-power AI sparks energy debate
Nuclear, renewables, and more needed to power AI

Artificial intelligence is becoming ubiquitous in our lives, and it’s only getting bigger.

But as this technology surges forward, it’s also turning into a ravenous energy beast.

The UN Environment Programme’s latest report has highlighted that while AI could be a powerful tool for tackling climate issues, it also has a significant environmental footprint.

For example, making just one powerful computer uses 800kg of raw materials, and the microchips that power AI rely on rare earth elements, which are often mined in environmentally damaging ways.

Data centres, where AI runs, also consume huge amounts of resources. These centres generate a lot of electronic waste, including harmful substances like mercury and lead.

On top of that, they use a massive amount of water to cool machines, up to six times more than a country like Denmark.

And at the moment, most of the energy used to power these facilities comes from fossil fuels, which contribute to climate change.

In fact, UNEP’s study has shown that using something like ChatGPT consumes 10 times the electricity of a simple Google search.

Is DeepSeek the saviour we’ve been waiting for?

But recently, there’s been a new twist.

A Chinese company, DeepSeek, might’ve just figured out how to make AI way less power-hungry, and it’s throwing a wrench in our energy predictions.

DeepSeek announced this week that it was able to create a chatbot that rivals some of the US’ best models at 20 to 40 times cheaper, forking out just under US$6million by using outmoded Nvidia GPUs.

Read later: China’s DeepSeek just blew up the AI game. Should investors freak out?

The breakthrough has sent tech and energy experts, as well as the markets, into a bit of a tailspin.

Until now, the energy focus around AI has been all about how much electricity it needs.

OpenAI, Microsoft, Alphabet and others, like Amazon, have been on a quest for more power, hunting down everything from abandoned nuclear plants to renewable energy sources.

But DeepSeek’s new model could change the game by delivering the same AI results using a fraction of the energy. This could mean a much smaller climate footprint for AI than we first thought.

Earlier forecasts predicted AI would drive about 75% of the growth in US power demand. Analysts are now questioning whether those estimates were a bit too optimistic.

Maybe AI won’t be the energy monster we’ve all been bracing for.

AI’s power needs will keep growing, says experts

But hold your horses, there are plenty of experts who reckon power demand from AI is only going to go up.

Take Australia’s largest data centre company, NextDC.

The company’s CEO, Craig Scroggie, argues that while AI models, like DeepSeek’s, might get leaner, the total power consumption won’t drop because more efficient tech means more applications, and therefore, more demand.

It’s a classic case of the “Jevons Paradox”, he told the AFR, as “efficiency doesn’t reduce consumption, it expands it.”

Sasha Luccioni, an AI researcher over at Hugging Face, pointed out that it’s not just training AI that’s the power-hungry problem.

She argues that once those models are out in the world, every time someone uses them (called “inference” in nerd-speak), they’re still burning through energy.

So the more people use apps like ChatGPT or DeepSeek globally, the more energy we will need.

Ayse Coskun, a computer expert at Boston Uni, agrees, saying that as tech gets cheaper, people tend to use more of it, which keeps demand up.

“The amount of AI power demand in the next few years will still be a high number” even if it shifts slightly,” she said.

Nuclear, renewable energy are still the key

Finding sustainable energy solutions to power this AI boom is, therefore, becoming more urgent.

Nuclear energy is seen as a key part of the solution, and big tech companies are already signing contracts for new nuclear capacity to meet their growing energy demands.

For example, Microsoft is in talks to restart the Three Mile Island nuclear plant, famously the site of a nuclear accident though one of its units has operated since, while Meta is looking for up to 4 gigawatts of nuclear power just to keep its operations running.

However, nuclear energy won’t be enough on its own.

While it can provide consistent, low-carbon power, building new nuclear plants is slow and challenging, with issues like waste management and high construction costs.

According to Goldman Sachs, natural gas, renewables, and battery technology will also play important roles in meeting energy needs.

Renewables like wind and solar can indeed help fill some of the energy gap, but they can’t provide the reliable, around-the-clock power required for AI operations.

In the future, advances in energy efficiency and new technologies may help lower the carbon footprint of AI data centres, GS concluded.