Could the success of AI make us consciously choose greener instead of bigger?

We live in an era in which artificial intelligence is no longer a promise but a reality that permeates our daily lives. From generative image creation on social media to language models that turn essays into podcasts, and agents that can already write sophisticated code, AI is woven into many of our everyday activities. Yet while we marvel at the impressive capabilities of generative AI, an uncomfortable question is emerging behind the scenes: what is the real cost of all this in terms of energy?

Generating a single image consumes roughly the same amount of energy as fully charging an iPhone. One minute of AI-generated video equals dozens of charging cycles. And that is only the energy required to use an AI model. Training large language models has already required energy volumes comparable to the peak output of a nuclear power plant. More recent models such as Google’s Gemini Ultra are around 100 times more complex than ChatGPT-3 and therefore roughly 100 times more energy-hungry.

Today, the global ICT infrastructure including data centers for AI training, networks and devices used for AI applications consumes about 10 percent of all electricity produced worldwide. Forecasts suggest this could rise to 20 percent by 2030, mainly due to the exponential growth of AI applications. Add to that the enormous water consumption required to cool data centers, and the picture becomes even more worrying.

AI is successful and therefore here to stay. The question is whether we continue to organise its growth in a way that requires ever more energy and pushes us toward an energy crisis, or whether we dare to choose a different direction.

The paradox of progress

AI is often presented as part of the solution to climate and energy challenges. Smart software agents, predictive maintenance and optimised logistics are just a few examples. But paradoxically, AI risks becoming a significant part of the problem.

We are witnessing an exponential scaling in model size and computing power. Where computing requirements once doubled every two years, the cloud revolution has brought a hundredfold increase in required computing power every two years. The credo “bigger is better” has resulted in models with hundreds of billions or even trillions of parameters.

Yet bigger does not automatically mean more efficient. Using a generative model to classify film reviews, for example, can consume up to thirty times more energy than a specialised model designed specifically for that recommendation task. In other words, we often use a cannon to shoot a mosquito.

On top of that comes the rebound effect. When AI makes processes more efficient, it does not automatically reduce energy consumption. It may instead lead to more production, more consumption and therefore more total energy use. Efficiency without systemic change is no guarantee of sustainability.

AI can therefore be both an energy guzzler and a climate solution. Which of the two will prevail depends on the choices we make today. 

Time for an AI “Nutri-Score”

What is largely missing today is transparency. We have energy labels for household appliances, CO₂ scores for cars and nutrition scores for food. Why not for algorithms?

Imagine an independent AI Nutri-Score: a standardised sustainability index that evaluates algorithms based on energy efficiency, water use, hardware requirements, reusability and openness. In my book Sustainable AI (sustainable-ai.be), I define a conceptual framework for this based on five criteria: efficiency, energy awareness, modularity, circularity and openness.

Such a score would allow companies and governments to make informed choices. Do you really need a generative model for this task? Is there a more optimised alternative? How much energy does training require compared to daily use?

In short, what we measure, we can improve.

An alternative path exists

The good news is that solutions are not hypothetical. They already exist, but they are not yet widely applied.

For many applications, smarter solutions exist than simply building larger models. Models can be designed more efficiently and deployed more specifically. Not every task requires a giant generative model running on massive datasets and capable of doing everything. Why use millions of examples if a model can generalise from just a few?

We can also rethink where and when computations take place. Does everything have to run in distant, energy-hungry data centers? Or can intelligence move closer to the user, onto local devices and at times we deliberately choose, significantly reducing both data transport and energy consumption?

AI systems can also be built in a modular way. Smaller components can be activated only when needed, avoiding oversized systems. There are also choices at the hardware level. Chips can be designed to consume less energy, and components can be made circular so they can be reused or at least recycled. Sustainability should therefore be considered from the design stage onward.

This is not utopian thinking. It is a strategic choice. The question is not whether it is possible, but whether we want it.

Europe at a crossroad

Bigger models, bigger data centers, bigger investments. Europe does not have to blindly follow Big Tech’s path of “AI for world domination”. If AI pushes us into an energy crisis, it will be the result of our choices. And choices can be reconsidered.

Europe has a strong position in AI and chip research. Two of the most important players shaping the future of AI are in our own backyard: imec and ASML. Without them there will be no next generation of AI superchips. “Wanna play without them, Donald? Try us.”

Europe should use this unique position to write a different story.

Instead of joining Big Tech’s race for scale, Europe could focus on quality and sustainability. On AI systems that consume less energy. On technology that integrates circularity from the design stage. On clear standards that enforce transparency.

Digital sovereignty means more than storing data within the EU. It means helping to shape how the next generation of AI is designed. Not only faster and bigger, but smarter and more efficient. It means investing in research on low-power AI, neuromorphic computing, federated learning and chiplet design.

Europe has a choice: follow blindly or set the direction.

A call to action

The question is not whether AI is system-disrupting. It already is. The real question is whether AI can push us to choose greener rather than bigger.

That is possible, but only if we redefine how we measure the success of AI. Do we measure success in trillions of parameters and market capitalisation, or in social value per kilowatt-hour consumed?

AI can be a lever for sustainability. Its success can force us to design more intelligently and invest more consciously. But that will not happen automatically. It requires deliberate choices.

It is time to abandon the dogma that “bigger is better” and move toward “smarter is better”. Time to label algorithms according to their energy impact. Time to redirect research and innovation toward energy-efficient models. Time to put Europe back on the map as a pioneer of sustainable and sovereign AI.

Because if we do not change course now, the success of AI could become the fuel for the next global energy crisis. And that would be the greatest irony of the new digital century.

Erik Mannens is professor of Semantic Intelligence at Ghent University and a professor of Sustainable AI at the University of Antwerp. His research focuses on how artificial intelligence can be deployed more sustainably, transparently, and socially responsibly. He has previously led large research teams on data science and AI and is committed to open-source, ethical AI and smart digital applications. He is the author of the book "Sustainable AI."