Maybe we should talk about the environmental cost of AI

The progressive development of artificial intelligence systems is bringing increasing attention not only to its potential but also to the environmental costs associated with their operation. One of the main concerns is energy consumption, which is particularly high both during the training phase of models and during user interaction. For example, according to some estimates, a single interaction with Google's AI Overviews can consume up to ten times the energy required for a standard online search, or an amount comparable to what is needed for one hour of a landline phone call. Although these figures may vary depending on technological infrastructure and workload, the issue demonstrates that AI models require significant computing power, which translates into a substantial energy impact. Consumption is not limited to users' daily use. The training phase of models, especially large-scale ones, is even more demanding. According to The Verge, the training of GPT-3 (a model developed by OpenAI and now surpassed by more recent versions) required an amount of energy equivalent to the annual average consumption of about 130 American households. Even more energy-intensive are models specialized in visual content generation: creating a single image with an AI system can consume enough energy to recharge a smartphone.

The growing energy demand is reflected in national infrastructures, particularly in electric grids. In Virginia, one of the American states with the highest concentration of data centers, the technology sector consumes about one-fifth of the available energy. More generally, US data centers absorb about 4% of the national electricity supply—a share that could double by the end of the decade. Similar predictions have been made elsewhere: in Sweden, it is estimated that the demand will double by 2030, while in the United Kingdom, a 500% increase is expected over the same period. It is not ruled out that in many regions, the growth in demand will exceed the supply capacity, with potential consequences on rising energy costs and the risk of overloads or blackouts.

Is the future of AI nuclear-powered?

In addition to questioning the actual economic sustainability of the sector, companies operating in this field will sooner or later also have to face the environmental consequences of artificial intelligence. «This is the problem with artificial intelligence,» summarized James Mathes, head of DataBank, a company that builds and manages data processing centers, to Bloomberg. Several major players—including Google, Amazon, and Microsoft—are considering turning to alternative energy sources, particularly nuclear power. But there’s a catch. Nuclear energy will not be used directly to power data centers in the short term, but it represents a strategic resource to compensate for the growing demand while meeting the environmental goals that many companies have set. Indeed, numerous tech companies have committed to achieving carbon neutrality by 2030. In this context, nuclear power is seen as one of the few energy sources capable of offering stable, low-emission supply on a large scale. At the same time, investments are also being made in improving hardware and software systems to reduce the energy requirement per unit of computation. Even with technological improvements, the growing diffusion of AI could still keep overall energy demand high.