Training Artificial Intelligence and Energy Consumption: An Emerging Problem or ‘Hype’?

It is well known that training artificial intelligence consumes a lot of energy. The infamous chatbots and various other AI models that count hundreds of millions of weekly users worldwide and assist many in performing various tasks, creating realistic visuals, graphics for social media, or videos – all of them generate substantial server bills measured in megawatts per hour.

And it seems that no one, not even the companies behind the technology, can accurately say what the cost of that energy is. However, it is certain that electricity consumption in AI companies is growing exponentially, and some believe this leads to a new energy crisis in the U.S.

To be precise, the U.S. will need approximately 34 new nuclear power plants or an additional 38 gigawatts of electricity over the next five years to meet the demand of data centers, factories, and electric vehicles, estimated the consulting firm Grid Strategies in a report submitted to the energy regulator.

Companies that use artificial intelligence require a lot of electricity to train language models, produce chips, and other hardware. A significant source of demand is also the artificial intelligence embedded in search engines as it consumes at least ten times more than regular search engines.

In 2022, data centers, cryptocurrencies, and artificial intelligence consumed, according to a recent analysis by the International Energy Agency (IEA), about two percent of electricity globally, and that share will increase in the future. However, the U.S. and its Asian competitors are quite liberal when it comes to the development of artificial intelligence and are likely to do everything possible to accelerate its development. And while the titans race, the European Union, as always, cautiously regulates.

– Europe’s cautious approach has its advantages and disadvantages. On one hand, it can ensure that we avoid some of the potential negative impacts on society that the rapid application of any, including this technology, can cause, but of course, in today’s digital globalized world without borders, it is extremely difficult to prevent the entry of that same technology into your society that was developed elsewhere.

On the other hand, if regulation becomes too restrictive and bureaucratic and insufficiently flexible, it can seriously slow down the development of technology within the EU compared to the U.S. because the idea that the regulatory system, which takes years to launch regulatory arrows, will hit just the right target of a high-risk AI model in a herd of AI models quickly running across the field is quite optimistic and not very realistic. Regardless, I believe that regulation is necessary, and I am not the only one who thinks so, as my recent survey on LinkedIn showed that a similar opinion was held by another 57 percent of people – commented Davor Aničić, director of the domestic technology company Velebit.AI, which deals with artificial intelligence.

Is energy cost not a problem?

Predicting that energy consumption is a key limitation for further development is very speculative, says Aničić, as there are simultaneously strong trends in optimizing model architecture and training methods as well as increasingly powerful and efficient GPU processors that work in the opposite direction.

Additionally, models have the ability to upgrade and combine with each other, and it is uncertain whether the best models in the future will be those trained from scratch (which requires the highest energy consumption). Ultimately, once trained models consume several orders of magnitude less energy during use, and the question is whether, even with very wide usage, energy consumption will significantly increase in a short time.

– I do not believe that energy cost is a problem. Language models have indeed dramatically increased in complexity and energy cost in recent years, but this trend should not be extrapolated for several reasons. First, model training is done relatively rarely. I expect that the market will be dominated by a few large ‘user-ready’ already trained models, which will be optimized to run locally, e.g., on mobile phones, with very low energy costs.

Second, the trend in the industry is also the optimization of the creation and training of models themselves, which require less energy for the same quality of output. A dramatic example is chess-playing algorithms, which are now so optimized that both retraining and running require very little energy. The evolution of these algorithms has been such that the ‘brute force‘ neural network that required a lot of energy and parameters has gradually been replaced by an expert algorithmic structure optimized for the given problem. I expect that this will also happen with language models, said Siniša Slijepčević, co-founder and director of the Croatian AI scaleup Cantab PI and a regular professor at the University of Zagreb’s PMF.

He added that the very development of ‘hardware’ such as specialized chips for training and running AI models constantly optimizes energy costs and that he believes this trend will continue. Energy and algorithmic optimization of AI models is a necessity for them to be usable, says Slijepčević, which is why he believes that any discussion about artificial intelligence and energy sustainability is unfounded.

– The sustainability of development is certainly an important topic, and the very development of artificial intelligence and its applications in numerous industries, energy production and consumption, transportation, and almost all spheres of society can, if used correctly, lead to much more sustainable societies and the entire civilization, concludes Aničić.