Under the nave of the Grand Palais, long speeches are lost in an infinite echo. The shock formulas slam under the immense glass roof. “Plug, baby, plug!” – Branch, baby, branch! – dropped Emmanuel Macron, Monday, February 10, on the occasion of the closing of the first day of the summit on artificial intelligence (IA). A benchmark for Donald Trump and his slogan “Drill, Baby, Drill!”, On the white-only oil drilling. But rightly connect? Data centers.
France, which relies a lot on AI, needs these digital infrastructure in order to train and execute the language models that feed Chatgpt, or the cat of Mistral. Because these applications are particularly energy delicious. Is this trajectory sustainable and compatible with the fight against global warming? Sasha Luccioni, AI researcher for Hugging Face, a cutting-edge Franco-American company in the field, in doubt. Not only do the needs are increasing enormously, but large companies are very transparent about their consumption. This high -flying scientist, which is among the 100 most influential personalities in AI according to the magazine Timealso delivers her shock formulas. Disconnect, baby?
https://www.youtube.com/watch?v=ttxauazbhzs
L’Express: a large climate and AI coalition was launched during the summit. What do you expect?
Sasha Luccioni: Gather those who want to get involved on the subject. Once the critical mass has been reached, we may be able to demand figures from major players in the sector. Especially on their energy consumption. At the top, I present a project on the “Energy Scores”, a standard of energy efficiency for AI models. We have tested hundreds of open source models, but the proprietary models remain inaccessible. Because from Chatgpt, it is radio silence among the GAFAMs, which give as little information as possible so as not to be exposed.
The problem is as follows: what is their interest in being transparent on their climate impact? Should we create a “green AI” certification? The discussion must relate to incentives, because a binding international regulation would be difficult and long to set up. But the solution can come from social, community pressure. What reassures me is that the subject is gaining momentum. Originally, it was to be mentioned in a “transverse” way at the AI summit. He finally has his own events and announcements. Europe, in particular, begins to ask questions. Although the regulations, AI Act, does not yet contain specific provisions on the environment, there is an increasing interest in transparency.
A simple request on Chatgpt would require 2.9 wattheures of electricity, ten times more than a Google research, according to the International Energy Agency (AIE). Is this figure reliable?
No ! Because we do not even know, at a very concrete level, which is hidden behind the interface of a chatgpt. Is there one or more models? What architecture, what size? Fashion, now, is very energy-consuming “Mixture of Experts” (MOE), that is to say several models that can turn at the same time. If they are eight, in this case it would be necessary to multiply the figure of the AIE by eight!
Is it possible to create universal indicators to compare the different models of AI?
It is very complex, especially due to the contexts of use that vary. Some models are optimized for smartphones, others for data centers. Apple, for example, will surely not want to test its models in an environment of servers, because it is not its way of deploying them. And not everyone uses the same material. Google thus uses specific processors.
However, we have the feeling that large technological companies are aware of the problem. They display, for some, ambitious carbon neutrality objectives.
But they missed them all! Google and Microsoft, for example, say they do not know precisely the share of AI in their overall energy consumption. That’s wrong. They know how many GPUs (Editor’s note: electronic fleas dedicated to training and execution of AI models) are used in their data centers, their occupancy rate, and associated costs. But no one wants to take the first step and disclose these figures. Again: where is the carrot? Same thing for Nvidia, which sells many of these chips: the company knows full well what they consume. Basically, these GPUs were developed for video games. So you had to know when they overheat to optimize them. In my opinion, all these companies are just afraid to tarnish their brand image.
What is your analysis of the approach to Deepseek Regarding the energy optimization of its AI models?
This approach is interesting, especially on the training of the model. But the deployment raises questions: Deepseek “debits” a lot of text. So it actually consumes much more energy. More generally, inference – the commissioning of models – is the most decisive. In a study published last year, it was calculated that it took only 5 to 20 days for a model with 10 million users to be more energy -consuming than to training. Companies have therefore understood that it was not necessary to constantly reread them. A version can be adjusted, optimized, serve to different applications …
We also mention the water consumption Data centers. Is this subject underestimated?
Water is the next “bomb” in the sector. Large companies have already saturated the areas where electricity and water are quite abundant – I am thinking in particular of Virginia, in the United States. From now on, they are trying to settle elsewhere, in Chile or in India. But they meet resistances within the local population. It will be more and more tense. Rather than building huge centralized data centers, it would be more effective to distribute them, recover the heat released and use lower quality, or reused water. The current trend unfortunately goes in the opposite direction: companies favor the concentration of infrastructure, for cost reasons.
An argument has been widely advanced lately: AI will make us consume energy, but it can make us gain a lot elsewhere. Is it true?
This is Jevons’s paradox! It is recurrent, especially with technologies that have multiple uses, such as coal, electricity or AI: say that energy gains will be erased with use. An example: each year, Nvidia says that her chips are more and more effective. In parallel, the company sells more and more. So yes, per unit, the chips are more effective. Except that this sobriety fades since many more people have access to it.
What is the difference with the rebound effect?
It is a generic term. There are different types of rebound effects. Economic: This is Jevons’s paradox. Material: Our phones are more and more powerful and small, but the data centers – where the calculations are made – increasingly large. Behavioral: we say that AI allows you to make the right choice, but it is used in advertisements, so it ultimately leads to the purchase of more products … or, more simply, the use of chatgpt like a simple calculator .
Companies in the sector are fighting to secure all possible sources of energy, in particular low carbon and renewable: long -term electricity purchase contracts (Power Purchase Agreement, PPA), constructions of their own wind or solar parks … Is it viable?
For ten years already, technological companies have been the biggest PPA buyers. They therefore participated in this economy even before AI explodes in terms of energy. It is quite criticized because they take the place of other actors – factories, farmers … – who cannot “plug” into this or that wind turbine because Google has already taken the place.
More generally, the case is complicated for solar And wind, intermittent energies, because data centers use energy 24 hours a day. So you have to find a way to store energy or have it constantly. It’s a big challenge. And that is why there is no 100 % renewable at present. On the nuclear side, small reactors are still experimental. At best, it will be necessary to wait 7 or 10 years to see it happen. In the midst of all this, there is Microsoft’s solution, which seems the most realistic, with the reopening of the Three Mile Island nuclear power plant. But even here, it will take a few years.
Will we arrive, one day, to prioritize the uses of AI for reasons of access to energy and water resources?
Non -renewable energy, you can always have it. In the United States, for example, in some cases where we approach the limit, we add coal or natural gas to “solve the problem” and avoid shortage. Obviously, climate level, there is better … We will have to make choices the day when we set up constraints in terms of carbon intensity. For water, unfortunately, there is no good mechanism to monitor practices. And we have zero figures. We don’t even know what order of magnitude we are talking about.
Will AI still be able to serve better predict the climate of tomorrow ?
Whether it is a model to generate molecules intended for batteries or another to recognize birds, everything is called AI. Overall, the most expensive models are the least useful, and vice versa. There are many people who have been working on AI for a long time and the fight against climate change. It is annoying to hear that the AI night in the climate or the planet. And then, now everyone wants to use generative AI or Llmmajor language models. So these pioneers of the climate cause benefit from less attention, and less money. Which is a shame, because level LLM, I have not seen anything, to date, which is really useful against warming of the planet. The most energy -consuming language models are strictly useless for climate prediction. Admittedly, there are tools like Climategpt, to whom we can ask questions about the GIEC reports. It’s nice, but it doesn’t fundamentally change the situation.
.