With big AI getting ever-bigger – and therefore more impactful – in the context of historically huge IPOs (OpenAI, Anthropic and xAI/SpaceX), and close-to-a-trillion dollars being sunk into AI data centres, is anyone actually listening to those warning about the ever-bigger associated planetary harms? What is the work that needs to be done, and the messages that need to be spread, to keep challenging the AI hype steamroller?
This was why Sasha Luccioni, Chief Scientific Officer of the Sustainable AI Group (or SAIG, above), gathered an all-star team of world-leading AI experts in Paris June 9-10, as part of her role as chair for social justice and AI of French university the ENS.
Whilst day one saw said experts speak onstage about topics spanning the entire AI lifecyle, day two consisted of closed-door workshop discussions between them. More on those later. For now, here are the day one conference highlights!
Setting the foundations, with Emma Strubell
A huge treat off the bat: an opening keynote from Emma Strubell, of Carnegie Mellon University; a true pioneer, as the very first scientist to publish globally-recognised work on the environmental impacts of AI.
How can we measure AI’s energy and emissions impacts? Look no further than the above formula, which can also be adapted to work for water consumption too (more in the paper “Holistically Evaluating the Environmental Impact of Creating Language Models“).
Why are LLMs so resource-intensive? According to Emma:
- The scaling of training, or using as much data as possible (the root of AI’s “bigger is better” obsession, since Jared Kaplan’s 2020 white paper)
- The scaling of inference, as demonstrated by reasoning, or ‘thinking’ models, which effectively prompt themselves multiple times before providing an answer. “The Claude prompt you don’t see is around 25k tokens”, said Emma, adding that 17 times more energy is required to post-train a reasoning model than a standard LLM.
This leads to rebound effects, or the infamous Jevons paradox, which states that the more efficient a service gets, the more it is used, not the opposite (Emma, Sasha and Kate Crawford wrote a major paper on that too). This leads to what Emma called “the curse of efficiency“, whereby making AI models ever-more efficient leads to them being ever-more used, and therefore more impactful.
“The AI community thinks improving efficiency will solve the problem. Jevons proves that’s not enough. The real challenge is allocating limited resources” – Emma Strubell
(Remember this essential point, as it will come back a lot!)
In this context, key challenges moving forwards include:
- The lack of clear means of impact measurement: the fact that the few figures provided by companies like Google, OpenAI or Mistral AI today are not comparable is “not OK! A huge problem“, re. Emma, who also shared a graph suggesting AI’s future electricity consumption may equal that of EUROPE…
- We know moving towards more specialised, smaller models is a way forward. But choosing the right ones isn’t easy (yours truly is here to help there!)
- Incentives and governance are both seriously lacking.
Emma also flagged the current massive waste inherent in how GPUs are used in AI today; more on that below, too…
To conclude, said our keynote speaker, it’s crucial to characterise indirect effects (such as Jevons-type consequences of AI use); market incentives are not enough (we can’t just hope AI companies will become more sustainable, as growth is their top priority); and innovation and sustainability are not mutually exclusive, as long as we are “more careful and deliberate” on that front, insisted Emma.
Current trends, with Google, CapGemini and Pruna AI
The day’s first panel united (right to left) Google Sustainability Specialist Vincent Poncet, CapGemini Chief AI Scientist Philippe Cordier and Pruna AI Founder Bertrand Charpentier. SAIG’s Nidhal Jegham moderated.
Perhaps the biggest AI trend of the day: “Clients are seeing the limits. They’re spending a lot, but not seeing the value“, said CapGemini’s Philippe. “We’re at a pivotal moment where the far [wild] west phase is over. It’s now about cost consumption.” In other words, now the major LLM coding platforms, like GitHub Copilot or Claude Code, have (mainly) moved to pay-per token invoicing, companies’ AI mindsets need to change. Will they, though?
“Customers just look at the top model on benchmarks, then take that one”, said Bertrand. “So they tend to take ChatGPT to generate an image; that takes 60 seconds per image, when you can do it in one. You get the real gains when you combine different tactics together”, insisted Pruna AI’s founder, referring to techniques like pruning or distillation.
“Most clients are looking at token price today; we urge them to look at token value. You should look at the total value of a workflow, not on a case by case basis. If it’s not working, redeploy. Lots of big orgs are bad at that” – Philippe Cordier
Google’s Vincent accepted that the shift to agentic AI means we’ve jumped from human users prompting LLMs to “machines querying machines 24/7… like people coming to an all-you-can-eat buffet with big boxes!” But, he insisted, this doesn’t necessarily mean exponential impacts.
With agentic AI, you don’t necessarily have to recompute all the tokens, you can use caching (to the tune of 80-90%). So we’re seeing larger and larger contexts, but more and more cache possibilities. 300 times more tokens have been processed between April 2024 & April 2026. [Energy consumption has not increased by as much because] we’ve seen massive efficiency gains” – Vincent Poncet
That said, Vincent insisted Google increasingly tells its clients “you don’t need big models for a lot of your needs. For example, to ask a customer chatbot ‘where is my order?’, you don’t need a big model.”
So what are some potential solutions for organisations faced with the above challenges? Philippe had three:
- Deploying AI on premise is “way cheaper than in the cloud”; but it completely depends on your workloads. “For less critical workloads (e.g. a decision that has to be taken in 30 seconds vs. five), you can put them in the cloud. The key is making sure mission-critical needs can run at an SLA that the client is looking for”
- “To make this tech more affordable for companies and society, we need to improve efficiency 100x… BUT also watch out for information corruption in agentic workflows: some papers show 95% corruption after just five turns“
- “Companies tend to forget that if they don’t have data, they don’t have fuel. They are too used to legacy systems. They also tend to forget an agent needs to understand your business goals. This is why so far we don’t see a huge ROI [from AI]…
Indirect effects, with Adora Foundation, ENS and VU Amsterdam
Next up was a panel on “Conceptual definitions and Indirect impacts”, with (right to left) Théophile (Théo) Lenoir, Coordinator of the Observatory on the Environmental Footprint of AI at EMS Paris; Adora Foundation Founder Ismael Velasco; and Vrie Universitet Amsterdam Researcher Alex de Vries. Yours truly moderated.
Firstly what do we mean by indirect impacts? For Alex, in the context of data centres, indirect impacts are the offsite ones, and direct ones are onsite (i.e. like scope 1 emissions are).
As for rebound effects, Alex came back to Emma’s “curse of efficiency” point:
“If you have a Jevons paradox, you’re losing more resources than what you started out with. We’re using a lot more power now than a couple of years ago. Data centres are responsible for 1.5% of global electricity consumption. AI as a part of that would be at least 20%. We often hear from companies like NVIDIA ‘it’s OK, we’re getting more and more efficient’. But we should be careful of messages like that, because the the consumption is far outpacing the gains in efficiency. If you have mass adoption, that totally offsets those gains. And you’re still reinvesting your efficiency gains in bigger models, as well” – Alex de Vries
Indirect effects are also more difficult to measure, said Théo, “mostly because it’s very hard to have a causal relationship between an AI solution and the effects it has“. This also makes it hard to quantify positive impacts, he added, because “you need to create a baseline scenario against which you compare a world in which there is AI compared to a world in which there isn’t AI. And so you have to do a lot of assumptions that are always open to criticisms, and that can be very specific to a certain context.”
What about agentic AI? Isn’t that the biggest contributor to AI’s indirect (and direct) impacts right now? This is indeed precisely why, according to Ismael, “Google processed a few trillion tokens a year ago. Then in October, 1.4 quadrillion. Now, 3.2 quadrillion. All the bets that people are making on products based on available tokens? There’s not enough.“
But tokens aren’t the problem, Ismael insisted: the software and the processing is. “GitHub processed about 1 billion code commits in 2025. Since agentic AI, that’s about a billion per month.” And yet, similarly to Vincent’s previous point, AI’s energy consumption hasn’t increased twelvefold (we hope!)
So what’s the problem?
“The main mediator of AI impact is the software,” said Ismael. “If you have a really efficient model in a really inefficient harness, the impact will be way worse than vice versa. There is a difference between AI impacts and IT impacts. Imagine a UI, with every token being able to press a button. 10,000 buttons trigger each 20 API calls. And those API calls each triggers 20 different system calls. Suddenly, you’ve got just those 10,000 tokens engaging with so many different bits of software. And that’s how you you suddenly get that enabled effect. It’s not the AI. It’s everything else the AI is touching.”
What are some ways out of this maze? Could legislation help, given that right now, as Yoshua Bengio famously put it, “a sandwich has more legislation than AI”?
“I don’t know anything about legislation on sandwiches”, joked Alex. What he does know: “Microsoft made a suggestion to not publicly make this [data centre energy consumption] information available. The European Union Commission apparently just word for word copied that suggestion. So now even the policymakers that are supposed to help us are working against us. As long as you’re not getting proper information [from AI providers], it’s really hard to make proper policies.” Even in the cases where providers are transparent, they don’t have to. We know, said Alex, that Meta discloses a ratio of 24:1 for their water consumption, as they include the water used to make their electricity. “But they’re the only ones who declare that”, i.e. because it’s not obligatory.
“I don’t know if legislators need to understand much more about the environmental footprint of AI,” added Théo. “We should prioritize setting emission targets in existing sectors, and make sure that the companies and organisations actually reach these targets, with or without AI. There are obligations now to produce accounting mechanisms for carbon for all the sectors with the CSRD at the European level. What matters is just reducing emissions, right? Businesses are great at having a target and trying to optimize to reach that target. So it depends on how much reducing emissions is actually a target.”
Working within planetary boundaries, as initially suggested by Emma, seems to be way better a way forwards than going all-in on efficiency, said Ismael… somewhat more radically, however!
I am almost at the point of saying I am anti-energy efficiency. The green compute community has been a major contributor to the crisis that we have now. Because we have been optimising the wrong thing. We’ve treated efficiency as the end. [Whereas] the greener we are per process, the dirtier we are per system. We need to think in terms of ceilings. We’ve now hit the planetary limit. What is the maximum that we have available? And then you work backwards.
So if we were able to say with agentic AI, with models, not so much how do we improve the per operation efficiency, but how do we measure and improve the net demand. There’s a lot of room for improvement there. I think if we really climb that hill as a community, if we spend five, ten years just working on that, you’d be able to get business growth and innovation, but dramatically reduce the planetary impacts. The focus on efficiency is accelerating the impacts perversely even with good intentions. Whereas if we focused on demand, the rest will follow.
Yes, we still managed to find a positive note to end with! And that was what Ismael called “societal consent“, as demonstrated by the €152bn in data centre projects currently blocked by local communities in the USA. “The reality is that we have the power, the work that you’re doing around frugal AI (thanks for the Frugal AI Switchboard plug, Ismael!) public opinion, the protest around data centres, the regulatory movements. The big labs are petrified of it.”
“In order to make this more fair, we should make sure that [tech companies] start paying the price for this”, said Alex. “So this could be – I’m just gonna say a dirty word – a carbon tax. Or maybe a token tax. Things like that make sure that those external costs that are now paid for by society become more paid for by the tech companies. And that will potentially also make them think twice before they expand.”
Théo agreed in the importance of “involving citizens in deliberation processes to try to understand when you have a data center next to your city, what do you want it to be used for. And do you are you allowed to have a say in this? Obviously, today, not at all. Is it reasonable to think that maybe you should have a say in this? This is really a priority, to think about usage and and how to prioritise that.”
Hardware impacts, with Sophia Falk
Then it was on to Sophia Falk of the University of Bonn, my go-to researcher when it comes to (mis)use of GPUs for AI. She presented the two main (must-read) white papers whose work she has led of late, namely “More than Carbon: Cradle-to-Grave environmental impacts of GenAI training on the Nvidia A100 GPU“, and “From computation to environmental cost the resource burden of artificial intelligence“. I’d also recommend “From FLOPs to Footprints: The Resource Cost of Artificial Intelligence“, which hammers home the point of how hugely wasteful AI GPU usage currently is (TL;DR: if they were used 3x better and 3x longer, we’d only need half of many of them in total…)
All of the above are based on the first ever 16-PEF-criteria life cycle analyses of AI GPUs (cf. above image); analyses during which Sophia et al constantly ran into massive obstacles caused by big tech opacity. For example, they had to grind a collection of NVIDIA A100 GPUs into dust to find out what they were made of (because NVIDIA won’t say, as they don’t have to), or make a lot of assumptions about the sizes of models like GPT-4 (cf. above), again, because these figures aren’t publicly disclosed. This is notably why the team made an impact calculator aligned with this work, where you can enter your own values.
Most recently, in the “From computation to environmental cost” paper, Sophia established that, despite GPT-4’s training having 3000% more material impacts than that of GPT-3.5, it only scores 40-60% better on two major benchmarks out of five. So are we sure bigger is better?
Whence Sophia’s recommendations:
- Popularise beyond-carbon metrics, then act on them (emissions are just one of any product’s full range of 16 environmental impacts, or PEFs)
- Ensure environmental impact frameworks are truly multicriteria (this naturally would require more industry transparency)
- Algorithmic innovations could improve model performance better than just scaling up. “We will fail if we only look at efficiency”, said Sophia; “you can’t make copper more efficient”
- Carbon-only accounting also tends to ignore localised impacts, she concluded.
Governments and Grassroots activists, with ADEME, AI Planetary Justice Alliance and (the French government’s) Ecolab
Speaking of local communities’ opposition to AI data centres, what role can such activism have in making AI more sustainable? And what role should governments play? That was the topic of the next panel, with (right to left) Matthieu Wellhoff, who leads the digital team at French national environmental agency ADEME; Charline Meyer, in charge of AI initiatives at Ecolab, part of the French Environment Ministry; and Sara Marcucci, Founder of the AI Planetary Justice Alliance. Théo Alves da Costa, of Ekimetrics and Data for Good, moderated.
Sara is actually one of the few people putting boots on the ground in parts of the world where AI’s physical impacts are the greatest. She just got back from Zambia, where a lot of the copper used by all sorts of tech companies is mined. “It’s not looking good” there, she said. “It’s hard to get information about what’s going on in places like these. You have to go there and ask the local people. If 50 of them have been displaced, that may not sound that bad, but it is. This is why we talk to communities, to build a bridge between them and policymakers. In Zambia, there are many effective civil societies. But the government is under Chinese pressure (that’s why RightsCon was cancelled). If you apply enough pressure, communities have power and organising does work.”
Traceability and transparency is another challenge, said Sara: “it’s really hard to trace, for example with this one copper mine in Zambia, who is using its metal? Similarly to waste, it’s a very opaque sector, there’s lots of organised crime.”
“A key way to fight is to promote more transparency from AI providers“, re. Charline; something she said is one of Ecolab’s priorities.
It’s one of ADEME’s too! This is why they asked French model maker Mistral AI for what turned out to be the most complete life cycle analysis (LCA) ever produced by an AI provider. But it wasn’t easy, said Matthieu:
It took us 9 months to contact them! And three more months to do the LCA. The results are not that interesting. But it proved it is possible to do an LCA of a LLM. This proves the problem is not methodology. It’s more a question of whether they’ll open their data. So we had to sign really hard NDAs. It was a perfect LCA. But they just published one page. Because they didn’t want to take too much flak.
Indeed, the fact that a one-page LCA that says nothing about energy is the industry’s most complete says a lot about AI’s overall transparency. But at least this experience proves, as Matthieu concluded, that all “we need is transparency and methodology.”
This is why Ecolab has actively promoted initiatives like French national standards agency AFNOR’s Frugal AI Framework, which lists 31 best practices for more sustainable AI – the first of which being “the most frugal AI is the one you don’t use“, as Charline put it – and is in the process of becoming an EU standard. It has also been pivotal in the creation of the Coalition for Sustainable AI, which unites 220+ companies and organisations – including NVIDIA! – to align on an alternative approach to Silicon Valley’s “bigger is better” AI dogma.
Should we wait, however, for standards, laws and frameworks to be absolutely perfect? Au contraire! “If we wait to determine a clean, quantitative approach [to more sustainable AI], it’ll be too late“, alerted Matthieu. “It’s still time to act now. We have [French sustainable IT guidelines] RGESN. It’s easy to adapt that to AI, so we can say ‘this model is more ecodesigned than others’.”
“Standards are just a little piece of the puzzle,” concluded Sara. “This doesn’t mean they’re not useful. We do need obligatory mechanisms to hold companies accountable. And we need participation, for example from local communities. Plus: this didn’t start with AI. It’s an opportunity to shed light on problems that have been around for a long time.”
The latest AI + energy stats, with the IEA
Then it was time for an essential data dump, with Siddharth Singh, World Energy Outlook for the IEA (International Energy Association). Some of his key points:
- Capex by hyperscalers has already surpassed that of the oil and gas expenditure – over $700bn
- Data centre electricity demand grew 17% in 2025 (vs. 3% for all sectors combined). Data centres are driving 10% of growth in electricity demand right now
- Bottlenecks to that growth are also developing quickly: even gas turbines can now take four years to obtain, vs. six years for grid connections
- The IEA still considers that data centres will go from 1.5% to 3% of global electricity consumption by 2030, because of AI. But it’s important to note that this proportion is considerably higher where data centres cluster. “in some parts of the US, or in Ireland, that share can be 20-30%“, said Siddharth; “that’s causing lots of challenges to grids.”
On a per query basis, impact doesn’t come from text, said Siddharth; one agentic query can use around 1000x more energy than a text prompt, he said (cf. graph above left).
He also detailed one of the key findings from the IEA’s last AI-focused report: the electricity consumption of AI-optimised data centres will triple by 2030 (above left), putting them on a par with that of conventional facilities. This tripling is aligned with that predicted in other reports (cf. here). AI racks, such as those pictured above right, can consume as much electricity as 65 households – the equivalent of “a neighbourhood”, as Siddharth put it.
One of the most worrying trends: data centres are the single biggest source of new orders for gas turbines right now, at around 30% of total orders. These turbines – such as the methane monstrosities installed by Elon Musk in Memphis last year – will as such be one of the biggest contributors to new emissions in coming years.
Finally, a positive note, Siddharth outlined how, at least in terms if domain-specific scientific models, smaller is most definitely better when it comes to energy consumption.
This and many more examples in the IEA’s latest AI report, “Key Questions on Energy and AI“.
AI and Environmental Regulation, with (the other) ENS
We were then treated to a keynote by Philipp Hacker, Professor of Law & Technology at Berlin’s European New School of Digital Studies (the other ENS), who achieved the rare feat of making law entertaining!
Beginning with a figure even more worrying than Siddarth’s – Greenpeace’s assertion that data centres consume 80% of Dublin’s energy – Philipp shared the simple equation he tries to convey to EU policymakers:
Sustainability = Competitive = Sovereign
This then leads to his “Legal Toolkit for Sustainable AI” (above). But that’s just a framework proposal.
What current laws can we lean on to make AI more sustainable?
- ETS, the EU directive which imposes a carbon tax on heavy industry and the transport sector could in theory apply to data centres, if they use gas turbines with over 20MW of fuel power. This, however, is not enough, as it only indirectly applies to AI models’ impacts; doesn’t cover embodied emissions (hardware); and “the price signal is too weak compared with the size of VC investments” currently going into AI, said Philipp
- EED, the other EU directive that demands data centre operators declare their impacts, is useful for transparency, but that usefulness is limited, as it only provides aggregate efficiency, not specifics per facility
- …there’s also that “copy paste” scandal (cf. above), which shows how far lobbying power can reach. Although Philipp pointed out that the EC copy-pasting text from Microsoft’s lobbyists most likely contravenes the Aarhus Convention, as it states that “information which is relevant for the protection of the environment shall be disclosed”.
- The AI Act could also be useful, notably in that its GPAI section, for General Purpose AI models, demands that providers declare the energy consumption of training their models. However, 1/. none do, yet; 2/. open source GPAI models, such as those of Meta or DeepSeek, are exempted; and 3/. inference impacts are missing from the AI Act…
So, what proposals or recommendations could we make? Philipp has some ideas…
Transparency: Data centres should make their emissions, energy and water figures public again, as indeed international law requires; and models should be aligned by public and comparable metrics, with standardised benchmarks. And open source models shouldn’t be exempted from the AI Act, insisted Philipp.
Empowerment: Data centre operators should engage with local communities upfront (cf. the social contract idea, mentioned previously), since “studies show that greener data centre projects have higher acceptance rates“, said Philipp (whence the importance of transparency with emissions/energy/water etc figures). Model empowerment, on the other hand, could notably come from a right to opt out of AI being imposed on users, as it is for example in Google AI Overviews (this opt out possibility is notably something Sasha is pushing for right now). Similarly, users could be given a right to a greener model, whereby platforms default to the smallest and most efficient model for a given task, giving users sufficient information to make those choices. We like that idea!
Substantive obligations: Data centres could be obliged to meet enforceable renewable energy targets – beyond just declaring their PUE & WUE – and operators could accord grid connection linked to sustainability conditions, e.g. data centres must prove that 80% of their energy will come from new renewables (as this is what Ireland is currently doing, so can others, insisted Philipp). Models, for their part, currently benefit from an AI Act loophole. Whilst the act has a clause for models presenting a “systemic risk“, that only applies to new risks. I.e. if the new version of Claude has a systemic risk that was already in the previous version, it doesn’t count as a systemic risk 😵💫. A loophole to be filled, then…
Not to mention the environmental costs that may not be covered by the GPAI risk mitigation rules of EU AI regulations… more on that in Hacker et al’s recent paper.
Philipp then concluded with a handy resumé of these policy proposals. One to keep for future reference!
He then wrapped by insisting it was still time to act. Europe must hedge its bets, said Philipp, by regulating – but allowing – access to large frontier models; by taking the lead in small, specialised models (yay!); and by developing data centres with high environmental standards.
“If Europe doesn’t set the rules for green AI, who will?“, his final slide asked. Who, indeed?
Where do we go from here? Closing panel
With all of that said today, what’s next? The day’s keynotes were then tasked with wrapping up the conference’s key points, and giving us hope for the future. No pressure! SAIG’s Boris Gamazaychikov (left) moderated this optimistic exchange, between (right to left) Siddarth Singh, IEA; Emma Strubell, Carnegie Mellon University; Sophia Falk, University of Bonn; and Philipp Hacker, ENS.
One glimmer of optimism came from the recurring theme of local communities’ opposition to data centre projects. “Companies are very scared of having bad PR due to environmental impacts, plus they have stock value concerns too“, said Emma. “Maine, my home state, just passed a data centre moratorium; but it’s temporary. They’ll used that time to work out how to build these facilities in a way that’s beneficial for local communities.” The social contract, again.
So, asked Boris, if our speakers imagined a world where, in 2040, AI had finally become sustainable, what would that look like?
Philipp said it’s about “orchestration,” or “putting the adequate green model up to the task. It’s a trend we already see, but hard to integrate with agentic and reasoning. But with right invectives and mechanisms in place… this can be done in ways that are green rather than red.”
Sophia said it would be a world where “all (16) environmental impacts, as well as social ones, are cancelled out by AI’s benefits”. Hear hear!
Emma insisted that “the problem is socioeconomic inequality. AI is being used to further accelerate that inequality (indeed, as previous new technologies have). Stopping AI doesn’t fix that. But what I’d like to see is significant transfer of wealth, from a handful of billionaires to everyone else.”
Siddharth came back to his example of focusing on “domain-specific specialised models”, such as those used in biology, and “fewer large, energy-hungry models that do everything”. It’s about “focusing on niche models”, he said, powered by “a maximum of clean energy.”
And that, my friends, was a wrap for day one!
On day two, the above speakers convened behind closed doors to work on small groups on the following topics:
- How do we achieve consensus on AI sustainability metrics? starting points, challenges and opportunities
- What does a “sustainable data centre” look like, depending on where it’s built?
- How can we empower local communities in the face of data center expansion?
- How do we incentivize transparency in for profit companies?
Watch this space for further updates on the above! We’re still working on them, but publications such as further blogposts will be forthcoming. Thanks for following this far. And let’s keep working towards more sustainable AI!
Above, right to left from the top: Sasha Luccioni, Alex Lutz (Data for Good), Siddharth Singh, Emma Strubell, Philipp Hacker, Sara Marcucci, Sophia Falk, James Martin, Vincent Poncet, Boris Gamazaychikov, Julie Moorad (Arm), Théophile Lenoir, Théo Alves Da Costa, Nidhal Jegham, Alex de Vries, Ismael Velasco #dreamteam 💚
All photos © James Martin, except the above, by Guillaume Macaux, Obvia; and the Indirect impacts panel, by Sasha!
No AI was used to write this article! Except the transcription of the audio of the Indirect impacts panel, courtesy of (frugal AI platform) GreenPT 😇
