Power grids are cracking as demand for AI increases.

That's a big problem with generative AI, says Sasha Luccioni at machine learning company Hugging Face. Generative AI is an energy hog.

“Every time you query the model, the whole thing is activated, so it's very inefficient from a computational point of view,” she says.

Take large language models (LLMs) at the heart of many generative AI systems. They are trained on vast repositories of written information, which helps them extract text to answer virtually any question.

“When you use generative AI… it's generating content from scratch, it's essentially generating answers,” explains Dr. Lucioni. That means the computer has to work hard.

According to a recent study by Dr. Lucioni and colleagues, generative AI systems can consume about 33 times more energy than machines running task-specific software. The work has been peer-reviewed but has yet to be published in a journal.

It's not your personal computer that uses all this energy. Or your smartphone. The computations we increasingly rely on take place in large data centers that are out of sight and out of mind for most people.

“The cloud,” says Dr. Lucioni. “You don't think about these big metal boxes that get hot and use that much energy.”

The world's data centers are consuming more electricity than ever before. In 2022, they accounted for 460 terawatt-hours of electricity, and the International Energy Agency (IEA) expects that to double in just four years. Data centers could use a total of 1,000 terawatt hours annually by 2026. “This demand is roughly equivalent to Japan's electricity consumption,” the IEA says. Japan has a population of 125 million people.

In data centers, large amounts of information are stored for retrieval anywhere in the world – everything from your emails to Hollywood movies. The computers in these faceless buildings also power AI and cryptocurrency. They underlie life as we know it.

But some countries know all too well how energy-hungry these facilities are. There is currently a moratorium in Dublin to prevent the construction of new data centres. Around a fifth of Ireland's electricity is used by data centres, and this figure is expected to rise significantly over the next few years – all the while Irish households are reducing their consumption.

The National Grid boss said in a speech in March that demand for data center electricity in the UK would increase sixfold in just 10 years, largely driven by the rise of AI. The National Grid expects that the energy required for electricity transport and heating will be much higher overall.

Utilities firms in the U.S. are starting to feel the pressure, says Chris Sepple at Wood Mackenzie, a consultancy.

“They're being hit by data center demands at exactly the time we're having a renaissance in domestic manufacturing – thanks to government policy –,” he explains. According to reports in the US, lawmakers in some states are now reconsidering tax breaks offered to data center developers because the facilities are putting severe strain on local energy infrastructure.

There is an ongoing “land grab” for data center sites near power stations or renewable energy hubs: “Iowa is a hotbed of data center growth, there's a lot of wind production,” says Mr. Seiple.

Few data centers can afford to move to more remote locations these days because latency—the delay, usually measured in milliseconds, between sending information from the data center and receiving it from the user—increasingly popular generative. Not a major concern for AI systems. In the past, data centers handling emergency communications or financial trading algorithms, for example, were located within or very close to large population centers for optimal response times.

There is no doubt that energy demand for data centers will increase in the coming years, but Mr Seipel stressed how much uncertainty there is.

Part of this uncertainty lies in the fact that the hardware behind creative AI is evolving all the time.

Tony Grayson is general manager of Compass Quantum, a data center business, and points to Nvidia's recently launched Grace Blackwell supercomputer chips (named after a computer scientist and a mathematician placed on), which are specifically designed to power advanced processes including generative AI. , quantum computing and computer-aided drug design.

Nvidia says a company can train an AI in 90 days using 8,000 of the previous generation of Nvidia chips, which require a 15 megawatt power supply.

But the same task can only be done by 2,000 Grace Blackwell chips at once, and according to Nvidia, they would need a four-megawatt supply.

(This still ends up using 8.6 gigawatt hours of electricity – more than Northern Ireland's annual electricity demand.)

“Efficiencies are increasing so much that your overall energy savings are huge,” says Mr Grayson. But he agrees that power demands are shaping where data center operators site their facilities: “People are going where the cheapest power is.”

Dr. Lucioni notes that the energy and resources required to make the most advanced computer chips are significant.

Still, it's true that data centers have gotten more energy efficient over time, says Dale Sartor, a consultant and associate at Lawrence Berkeley National Laboratory in the US. Their performance is often measured in terms of Power Usage Effectiveness, or PUE. The lower the number the better. He noted that the PUE of modern data centers is around 1.1.

Mr Sartor says these facilities still generate a lot of waste heat and Europe is ahead of the US in finding ways to use that waste heat – such as to heat swimming pools.

“I still think demand is going to grow from the performance we see,” says Bruce Owen, UK managing director of Equinix, a data center firm. He predicts that more data centers will be built with on-site power generation facilities. Equinix was refused planning permission for a gas-fired data center in Dublin last year.

Mr. Sartor added that costs may ultimately determine whether generative AI is worth it for some applications: “If the old method is cheap and easy, then there won't be much of a market for the new method.”

Dr Lucioni stresses, however, that people will need to clearly understand how the options before them differ in terms of energy efficiency. She is working on a project to develop energy classification for AI.

“Instead of choosing this GPT derivative model which is very complicated and uses a lot of energy, you can choose this A+ Energy Star model which is much lighter and more efficient,” she says. .