Artificial Intelligence: what to expect in terms of the environmental footprint in 2025

"I’m not surprised, but if this is really the case, I didn’t think it would happen so quickly," says Vincent Courboulay, associate professor at La Rochelle University and co-founder of the Institut du Numérique Responsable in France, regarding the launch of Deepseek, whose creators boast of its lower energy consumption compared to ChatGPT.

We will therefore have to get used to it - the race to reduce AI’s energy consumption has only just begun. This is primarily an economic benefit for businesses, seen as a way to facilitate the wider deployment of AI. If this reduction proves effective, it could initially lead to a significant decrease in AI’s energy consumption, but then result in a major rebound effect due to the growth in usage and the number of associated services.

"It was written, since the boundaries of American high-tech are closing to China! It became obvious that innovation would come in an ecosystem under constraint" explains Vincent Courboulay. Indeed, for some time now, China has not had access to the latest generation chips developed by Nvidia. This is a good way to develop new innovation methods under constraint.

"We can see this as a positive narrative: it shows us that geopolitical, geographical, and environmental resource constraints can be bypassed!" adds Vincent Courboulay, "it reminds us that we need to move beyond the utopia of a limitless world that has brought us to where we are. Constraints will be an ally."

Economic and geopolitical constraints as allies of ecology

The transformation of constraints into allies is undoubtedly an unintended consequence for the Trump administration. While retaliatory measures were intended to allow the USA to dominate the market by having privileged access to the latest technologies, they are pushing other countries to innovate with fewer resources, thus minimizing environmental impact.

In this context, the potential success of Deepseek might inspire other countries. The potential savings in investment with a solution like Deepseek are certainly a particularly positive promise for the financial world. Indeed, the drop in stock prices following the launch of Deepseek by Nvidia clearly illustrates this. It also shows the volatility of this emerging market and the lack of understanding of AI’s technical challenges by speculators.

"Deepseek may be a bluff, but it could prove to Europeans, Africans, and others that another path is possible and that innovation can still happen!" says Vincent Courboulay. He sees it as the reopening of new paths to offer alternative AI models. This statement echoes some projects already underway to develop more frugal models.

For example, in France, the SHARP project aims to make a significant leap forward in terms of frugality by designing, analyzing, and deploying intrinsically efficient models capable of achieving the versatility and performance of the best models while requiring only a fraction of the resources currently necessary. SHARP is a project created by a consortium of schools and research institutes (Inria, CNRS, CEA, ENS Lyon, etc.) as part of the French government’s France 2030 plan.

"Deepseek shows that the game isn’t over! But I think two movements will coexist for a while: one with ever-increasing resource consumption, and the other with models that have minimized energy consumption!" says Vincent Courboulay. He sees the first trend of "more and more" as a strategy that is bound to run out of steam.

"After the announcement of Deepseek, Mark Zuckerberg reacted by saying he would invest even more massively in capital and infrastructure," he explains. "This manly movement is not sustainable, but in a world that thinks short-term, it will continue to dominate for a few more years."

Period of appropriation: the boom in usage

So, while we can expect a decrease in AI’s unit energy consumption due to the race for cheaper innovation, we must also expect a significant rise in electricity consumption - and thus AI’s carbon footprint - due to the ever-growing number of uses and the new services that will emerge as a result of this energy consumption reduction. This is what is known as the Jevons Paradox.

Moreover, given all the promises of AI, businesses and citizens do not want to be left out of what increasingly resembles a revolution: "Companies think this subject needs to be studied. The time for optimization will come later," adds Vincent Courboulay.

In his view, organizations currently consider energy consumption as primarily the problem of companies that produce language models. "Until companies are forced to, they won’t want to shoot themselves in the foot in terms of competition" he states, "there is a real drive in companies to accelerate processes." He tempers, however: "Many observers say that AI will be everywhere, that jobs will disappear, etc. These are partly just announcements."

And for the end user, the time for appropriation is also here. "People don’t know how generative AI works, but that doesn’t stop them from using it; it’s the same as with the car’s internal combustion engine!" he argues, "people are diving in and adopting the subject of AI. The tipping point is already here!"

Enabling informed choices

For Vincent Courboulay, legislating on usage wouldn’t really make sense. Instead, the key issue is enabling everyone to make informed choices. "For cars, there’s an energy label, there are labels, and soon there will be carbon labels for clothing..." he explains, "we should quickly ask ourselves if we should do the same in a sector like IT, which creates tensions over global electricity consumption."

In this context, educating and informing are key, according to Vincent Courboulay. "Universities and schools must take ownership of the subject of sustainable AI development, and public authorities must demand transparency on the impact of queries and the scope of subscriptions" he explains. This would be a good way to "question the usages that spread like wildfire, namely AI usages."

And this wildfire continues to bring its share of surprises day after day. Deepseek announced just days after the launch of its generative text-response AI the release of its generative image AI: Janus-Pro. And every day, more companies are launching applications that seem promising.

Illuminating usage and showing that other, more frugal development paths are possible becomes essential in a context of exponential AI deployment and adoption on one hand, and accelerating climate change on the other. This is a differentiating role that the European Union could perfectly take on to regain a central place in this technological revolution.

Vincent Courboulay is an associate professor at La Rochelle University, co-founder of the Institut du Numérique Responsable (INR) in France and is part of the GreenTech Forum Brussels 2025 program committee.

‍