Second part of my three-act reflection on the little-known challenges of these technologies. This week I explore the environmental costs and material challenges of generative AI.
The paradox of the immaterial
I am struck by how most people perceive generative artificial intelligence. They imagine it as a kind of ethereal, almost mystical entity, floating somewhere among digital clouds, freed from earthly constraints. This disembodied conception is reminiscent of our ancient spiritual representations, as if we had simply moved our gods from metaphysical heavens to Californian data centers.
How ironic this vision is, when I observe that perhaps no technology has ever been so deeply anchored in the raw materiality of the world! I witness this paradox daily: the more we believe we are freeing ourselves from matter through digital technology, the more dependent we become on monumental and energy-hungry physical structures.
Behind every text generated in seconds by ChatGPT, behind every portrait created by Midjourney, lies a colossal infrastructure consuming earthly resources. The data centers that host these AI models are technological cathedrals of pharaonic dimensions, whose ecological footprint perhaps constitutes the most troubling cost of this silent revolution.
This forgotten materiality of the virtual confronts us with an ironic paradox: even as we seek to transcend the limits of our condition through technology, we run up against the very real boundaries of our planet. Generative artificial intelligence, in its Promethean quest for creative capacity similar to that of humans, is thus constrained by the same physical limitations as biological organisms: the availability of resources and the balance of ecosystems.
The climate bill that no one receives
The environmental footprint of generative AI constitutes a major cost, often relegated to the appendices of corporate reports, when not simply concealed. This dimension takes on dizzying importance in the current context of climate crisis.
The data centers that host our generative AI models devour astonishing amounts of electricity, both to power thousands of processors and to maintain cooling systems worthy of the largest industrial installations. This consumption occurs in a context where, in the United States for example, about 60% of electricity production still comes from fossil fuels like gas and coal, an irony for technologies supposed to embody our post-carbon future.
The paradox is striking: these technologies that promise to propel us toward a radiant future rely largely on the same carbon energies we are desperately trying to wean ourselves from. Each conversation with ChatGPT, each image created by Midjourney, thus indirectly contributes to the greenhouse gas emissions that threaten the very habitability of our planet. A form of technological schizophrenia where the left hand deliberately ignores what the right hand is doing.
The figures are eloquent, according to the work of Sasha Luccioni (https://fr.wikipedia.org/wiki/Sasha_Luccioni) generating 1000 texts with an AI like ChatGPT consumes about 0.042 kWh. Even more impressively, generating 1000 images with an AI like Stable Diffusion consumes 60 times more energy, or 2.9 kWh. To make these abstractions more tangible, generating a single image with AI is roughly equivalent to a full smartphone charge.
In a data center located in Oregon, a location favored by Amazon for its cloud infrastructure, generating an image releases approximately 2.9 grams of CO2 into the atmosphere, the equivalent of a 23-meter car journey. Producing 1000 images per year is therefore similar to a journey of 7 to 23 km by car, or between 1 and 2 kg of carbon released into our common atmosphere.
Even more striking: a simple query on ChatGPT consumes about 5 times more electricity than a traditional search on Google (at least in the non-AI version of Google, because with the arrival of “AI Overviews” things are likely to change). An intensive user of ChatGPT, say 50 daily queries for a year, thus generates about 283 grams of CO2, still considering a data center in Oregon. A modest individual footprint, certainly, but one that, multiplied by hundreds of millions of users, outlines the contours of a considerable systemic impact.
The situation becomes all the more concerning when we observe that the energy consumption of the most advanced models follows a vertiginous exponential curve. A model like GPT-4 requires computational power and therefore electricity consumption several times higher than its predecessor, illustrating a technological trajectory whose long-term unsustainability seems inscribed in its algorithmic DNA.
Water and rare metals are the great forgotten ones
The environmental impact of generative AI unfortunately doesn’t stop there! The extraction of materials necessary for manufacturing electronic components, their industrial transformation, and finally their recycling (or disposal), when they don’t end up in open-air dumps somewhere in Africa, constitute as many stages generating a monstrous ecological footprint that is rarely accounted for.
Added to this is the intensive use of water necessary for both mining extraction and semiconductor manufacturing and data center cooling. For each kilowatt-hour consumed, a data center swallows about two liters of water for cooling. As an illustration, 50 queries to ChatGPT equate to consuming a half-liter bottle of water, an invisible luxury we afford ourselves without even being aware of it.
Lifecycle analysis must also take into account the production of electronic waste, particularly problematic in the context of generative AI where equipment obsolescence is accelerated by the frantic evolution of models. Specialized chips designed for one generation of models quickly become outdated, creating a cycle of material renewal with environmental consequences as considerable as they are invisible. According to a study published in Nature, the renewal of processors and other graphics cards could generate 2.5 million tons of electronic waste per year by 2030 if no mitigation measures are implemented. To make this figure more palpable: that’s the equivalent of 13 billion iPhone 15 Pros thrown in the trash.
Another study projects that, depending on adoption scenarios, electronic waste generated by generative AI could reach between 1.2 and 5.0 million tons in total between 2023 and 2030, with annual increases potentially climbing to 2.5 million tons in the absence of reduction measures. This electronic waste, true concentrates of toxic substances, poses major risks to human health and ecosystems.
And yet, despite the manifest urgency of the situation, only about 12.5% of electronic waste is currently recycled, even though it constitutes about 70% of toxic waste produced annually on a planetary scale.
A less radiant horizon than announced
According to a study published by Deloitte, following the current adoption rate, AI could require no less than 3550 terawatt-hours of energy in 2050. To give an idea of what this figure represents, already integrating “the improvement of data centers’ energy efficiency” promised by industry players, this would represent 37% more energy than that consumed by the entirety of France in 2023.
Even in the shorter term, the perspectives are just as vertiginous. According to a study conducted by Morgan Stanley, AI energy costs could grow by 70% per year by the end of the decade. “By 2027, generative AI could consume as much energy as Spain in 2022,” notes the firm, which acknowledges having been “surprised by its own projections.”
There is reason to question the long-term viability of the current trajectory!
The end of an illusion
At the heart of generative AI, there is an inescapable economic reality, that of access to ever-increasing computing power. Once reserved for experts, this subject is now becoming a key issue for the sector’s development, with major consequences for both the economy and the environment.
For several decades, the technology industry has lived in the cozy comfort of Moore’s law effects, with a regular doubling of computing capabilities at constant cost. This almost miraculous progression has allowed spectacular advances in many fields, including artificial intelligence, creating a form of abundance of “computation” that has nourished our collective imagination, fueling techno-optimistic narratives where computing power seemed promised to infinite growth.
But here are the physical limits of traditional semiconductors beginning to make themselves felt with growing insistence, even as the computing needs of generative AI follow a frenzied exponential curve. This decoupling creates a fundamental tension whose implications we are just beginning to measure: on one side, increasingly resource-hungry models, on the other, material efficiency gains that are inexorably slowing down, confronted with the implacable laws of quantum physics.
We are thus approaching a form of “peak computation”, analogous to the well-known concept of peak oil, where increasing computing capacities would become increasingly expensive and complex. This situation could profoundly transform the economics of generative AI, questioning the current trajectory of a frantic race towards ever more massive, ever more energy-consuming models, but with increasingly questionable marginal gains.
The great geopolitical game of silicon chips
This relative scarcity of computing capacities has triggered a veritable arms race in the field of specialized AI chips. NVIDIA, long the undisputed leader with its GPUs, faces growing competition from players like AMD, Intel, or ambitious startups like Cerebras and SambaNova. Tech giants are also developing their own solutions, like Google’s TPU chips or Apple Silicon processors integrating dedicated “Neural Engines.”
This competition goes far beyond the commercial framework to be part of a global geopolitical strategy with colossal stakes. Tensions between the United States and China over access to advanced semiconductors illustrate the fundamental importance of this resource, now considered a non-negotiable element of national sovereignty. American restrictions on the export of AI chips to China, and the latter’s efforts to develop an autonomous local industry, outline the contours of a new technological “Great Game” whose implications far exceed the field of artificial intelligence alone.
In this tense geopolitical context, access to the computing capacities necessary for training and operating generative AI models becomes a major strategic issue. This situation could accentuate inequalities between actors with privileged access to these resources and those who are deprived of them, creating new forms of technological dependence and exacerbating the already gaping digital divides that cross our world.
Digital colonization
Beyond the chips themselves, the manufacture of AI infrastructures requires considerable quantities of raw materials, some relatively rare or concentrated in specific geographical areas. Silicon, certainly abundant, but also metals like cobalt, lithium, tantalum, or rare earths, constitute the indispensable elements of this voracious computational economy.
The extraction of these resources raises ethical and environmental questions of a gravity we often prefer to ignore. In certain regions, like the Democratic Republic of Congo for cobalt, mining is accompanied by systematic violations of human rights and significant degradation of local ecosystems. The ecological footprint of AI thus begins well upstream of its energy consumption, from the extraction of raw materials necessary for its existence, in territories sufficiently distant from our decision-making centers that we can conveniently ignore the consequences.
This extractive dimension of the computation economy confronts us with a disturbing reality: the apparent immateriality of generative AI masks a deep dependence on terrestrial resources, with all the geopolitical, social, and environmental problems that this implies. The “dematerialization” promised by digital technology turns out to be, in reality, a different materialization, less visible but just as consuming of finite resources, simply displacing the negative externalities towards populations and ecosystems without a voice in the matter.
Hidden costs or the elephant in the room
Extracting rare metals, synthesizing electronic components, assembling and transporting products: all this generates much more CO2 than the use of these devices themselves, an uncomfortable truth revealed by a 2020 Harvard study. This research shows that it’s not so much the use of electronic devices that emits CO2 as their manufacture, this phase invisible to the end user.
The evolution is striking: between the iPhone 3GS of 2009 and the iPhone 11 of 2019, the share of the carbon footprint due to manufacturing went from 49% to 86%, a trend that is accentuated as our devices become more complex and specialized. Another revealing example: Google and Facebook’s data centers emit 20 times more greenhouse gases through the manufacture of their servers than through their electricity consumption during operation.
Why this disproportion? Because the use of non-greenhouse gas emitting energies is progressing in the operation of data centers, which limits the impact of their use. For manufacturing, on the other hand, we are still at the beginning of a transition that promises to be long and complex.
Thus, when we try to evaluate the ecological impact of AI, we should mainly look at the impact of component manufacturing and calculate the share used specifically to run AI systems (about 25% for servers). We can compare this situation to efforts to make cars more ecological: it’s not enough to replace gasoline with green electricity, we must also fundamentally rethink the construction of the vehicles themselves.
What if digital frugality were a solution
Faced with these material and environmental constraints of unprecedented amplitude, the optimization of the algorithms themselves becomes a priority development axis. This quest for efficiency, long secondary to the gains in raw performance that could seduce investors and media, is emerging today as both an economic and ecological necessity. But will it be enough to reverse the trend?
Among the most promising techniques is model distillation (knowledge distillation), this almost alchemical process by which a massive model “teaches” its capabilities to a more compact model. This approach, inspired by human learning mechanisms, allows capturing the essence of performance while drastically reducing the necessary resources, as if one managed to concentrate the intelligence of an entire library in a simple notebook.
Other techniques, such as weight quantization or sparse attention architectures, significantly reduce the memory and computational footprint of models. These algorithmic optimizations can decrease energy consumption by a factor of 4 to 10, without major impact on performance perceived by the end user. A technical feat reminiscent of the art of bonsai: obtaining the essence of a majestic tree in miniature format, without sacrificing its soul.
This search for efficiency reminds us of an ancient wisdom long forgotten in our race for technological excess: true mastery does not reside in complexity but in elegant simplicity. The experienced craftsman is not the one who uses the most tools, but the one who knows how to accomplish the most with the minimum necessary. Similarly, the future of generative AI might well belong not to the most massive models, but to those that manage to accomplish the most with the fewest resources.
The specialization of models
Another path to efficiency consists of favoring specialized models (SLM: Small Language Model), precisely calibrated for specific domains or tasks, rather than pursuing the chimera of general artificial intelligence with excessive ambitions. These models, more modest in their scope but more precise in their field of expertise, require significantly fewer resources while offering practical utility that is often superior, like a carved scalpel surpasses a Swiss army knife for a delicate surgical operation.
This approach is confirmed by Luccioni’s meticulous research, who discovered that electricity consumption depended greatly on model specialization. The more specialized the AI model, the smaller it is and the less it consumes. She therefore recommends avoiding generalist models (like ChatGPT 4 or Gemini) for specific tasks and favoring more compact models, specialized in a precise task. Or, even more simply, favoring smaller models for simple tasks, like the Llama3 8b model, ten times smaller than ChatGPT 3.5 but just as performant for many applications.
The recent emergence of these “small language models” (SLM) or compact language models perhaps marks a turning point in the evolution of generative AI. Companies like DeepSeek with its V3 and R1 models, or Google with Gemma 3, illustrate this trend towards greater efficiency that could reshape the technological landscape in the years to come.
This virtuous specialization reminds us that in biology, evolution has rarely favored the most complex or energy-hungry organisms, but rather those that have managed to adapt precisely to their ecological niche. Similarly, the future of generative AI might orient itself towards a constellation of specialized models, each excelling in its specific field, rather than towards the chimera of a universal model with unsustainable ecological costs.
Creativity born of necessity
Beyond technical optimizations, a more fundamental approach is emerging in certain still marginal circles: frugal innovation, this philosophy that prioritizes efficiency and adequacy to real needs rather than technological one-upmanship. Born in contexts of limited resources, this approach finds particular resonance in the face of environmental challenges of generative AI.
This philosophy of digital frugality is part of a broader reflection on technological sobriety that is timidly beginning to penetrate boardroom discussions. It questions our collective tendency to systematically privilege “more” over “better,” and invites us to rethink our criteria for evaluating technologies, integrating dimensions such as environmental impact, accessibility, or resilience in the face of material constraints.
Frugal innovation also reminds us that constraints, far from being merely obstacles to overcome through a profusion of means, can also be powerful drivers of creativity. Just as the Japanese haiku finds its expressive force in a very strict formal framework, or as the greatest chefs create gustatory wonders from simple and local ingredients, generative AI could find in ecological constraints an invitation to reinvent its approaches, for more elegance and efficiency.
Relocating calculations for distributed intelligence
Another significant evolution, still embryonic but potentially disruptive, concerns the progressive migration of computation from centralized data centers to the edges of the network, what we call edge computing in technical jargon. This transformation of the very architecture of generative AI could have considerable implications both economically and environmentally.
I have been very interested in progress in model optimization that now allows running reduced but functional versions of generative AI directly on users’ devices, personal computers, smartphones, or even specialized connected objects. This approach presents several non-negligible ecological advantages: reduction of needs for centralized infrastructure, decrease in consumption linked to data transmission, and possibility of using resources when they are already active, rather than maintaining armadas of servers in permanent operation.
On the environmental level, this relocation of computation could contribute to a more balanced distribution of the computational load, avoiding excessive concentration of impacts in certain geographic areas already under pressure. It also allows better adaptation to local energy resources, such as preferential use of computing capacities when renewable energy is abundant, a form of virtuous energy opportunism.
This trend towards local intelligence sketches the contours of a more territorialized AI, in phase with the specificities and resources of each geographical context, a form of digital regionalism that contrasts with the globalized uniformity of large centralized models.
Swarm intelligence
Beyond simple relocation, there exists a more radical idea of a fundamentally distributed artificial intelligence, where computational capabilities would be distributed among numerous interconnected nodes. This “swarm” model draws direct inspiration from natural systems like insect colonies, where collective intelligence emerges from the collaboration of individually limited entities, a coordinated ballet that produces intelligence superior to the sum of its parts.
Initiatives like Federated Learning already allow training AI models in a distributed manner, without centralizing sensitive data. This approach could extend to inference itself, creating artificial intelligence networks where each device would contribute to the collective effort according to its available capacities, like an immense decentralized computational symphony.
On the environmental level, this distributed architecture presents the advantage of better resilience and more efficient use of existing resources. Rather than investing billions in the frantic construction of massive data centers, it values computational capabilities already present in our daily environment, billions of electronic devices whose power remains largely underexploited, like a vast ocean of intelligence lying fallow.
This vision of collaborative intelligence, distributed like a neural network on a planetary scale, perhaps offers a path of reconciliation between our technological ambitions and ecological constraints. It suggests that a truly sustainable AI will not be embodied in ever more imposing technological cathedrals, but in a fine and resilient fabric of interconnected intelligences, a form of computational democracy that would contrast with the current oligarchy of technological giants.
AI and humans, a… troubling comparison!
In this complex landscape where technological, ecological, and economic issues intertwine, certain points deserve to be nuanced to avoid reductive simplifications and too easy moralizing postures. Let’s explore some paradoxes that blur the lines of our collective reflection.
If we compare the carbon footprint of AI to that of a human performing the same tasks, the results shake some of our certainties. I recently studied research comparing the time spent writing or producing an image (a few seconds for AI, a few hours for humans) and the average carbon cost of a human and their computer over this duration.
The conclusions struck me: an AI that writes a page of text would emit 130 to 1500 times less CO2 than a human accomplishing the same task! Similarly, an AI that creates an image would emit 310 to 2900 times less carbon than its human counterpart. Currently, the use of AI would thus offer the possibility of performing several creative activities at emission levels drastically lower than those of humans, a statistic that will certainly be highlighted in the next marketing campaigns of AI companies.
This observation, if technically accurate, deserves to be examined with a critical eye. The study only briefly, almost timidly, mentions the risk of rebound effect, this implacable law of our technological societies where each efficiency gain paradoxically translates into an increase in global consumption. In other words, the efficiency gains allowed by AI could in practice be not only canceled, but spectacularly reversed by an explosion in content production.
The authors of the study omitted to quantify this rebound effect, this metamorphosis of efficiency into overconsumption that has characterized the history of our technical innovations since the steam engine.
I already observe it in my professional environment: when it becomes possible to generate a thousand images in a time that previously allowed us to create just one, we do not witness a proportional reduction in emissions, but a vertiginous multiplication of production.
A carbon drop in the ocean of emissions?
To contextualize these figures and avoid sterile and counterproductive moral panics, it is appropriate to place the impact of AI in the more general picture of our emissions. Information and communication technologies currently participate in between 2 and 6% of global greenhouse gas emissions, some more recent and more optimistic studies estimating it at only 1.4%. In this digital landscape, data centers currently contribute to 0.1% of global greenhouse gas emissions, according to an article in Nature. It is in this relatively modest segment that AI is situated (for about 25% of this fraction).
I consider it important to maintain a proportionate perspective: if AI certainly poses significant and growing environmental challenges, it currently represents only a fraction of global emissions, far behind the energy, transport, or industrial agriculture sectors. This perspective does not aim to minimize the urgency of addressing the environmental impact of AI, but rather to avoid the disproportionate media attention it arouses from diverting our collective gaze from the much more fundamental systemic transformations that the climate crisis demands.
This quantitative precision does not mean that we should ignore the exponential growth of this sector and its potential future impact, but rather that it should be placed in the broader context of our climate challenges. An increase in AI energy consumption of 70% per year, as projected by some studies, would quickly transform this carbon drop into a torrent if nothing is done to bend the current trajectory.
AI both poison and remedy?
AI is not only a problem from an environmental point of view; it could also, under certain conditions, be part of the solution, perfectly embodying this notion of pharmakon, both poison and remedy, dear to Greek philosophers.
Google’s report on AI ethics (one will appreciate the irony of a company evaluating its own technologies) cites several cases of AI uses to curb climate change: systems that can predict extreme weather events, reduce energy consumption of industrial cooling systems, help design plastic-eating enzymes, and accelerate research on nuclear fusion.
I was very interested in an article published in 2023 by a team of scientists from the University of Chicago. They demonstrate that AI solutions, like their “CarbonMin” system, can help intelligently direct queries to data centers powered by low-carbon electricity. According to this article, this approach would reduce emissions by 35% today and 56% in 2035. Combining this with technological progress would limit the sector’s emissions increase to 20% in 2035 despite a load 55 times higher, a considerable efficiency gain, but one that nevertheless constitutes a net increase in emissions, perfectly illustrating the limits of purely technological solutions to our ecological challenges.
However, this same report notes with honesty that AI could just as well amplify the problem by increasing the productivity of extractive and carbon-intensive industries such as oil, gas, or industrial livestock farming, helping them optimize and thus intensify their polluting activities. This fundamental ambivalence, this capacity to amplify both our destructive tendencies and our preservation efforts, perhaps constitutes the most troubling characteristic of AI in its relationship to the environment.
Finding the right measure
The environmental footprint of generative AI, far from being a simple negative externality that could be relegated to the appendices of CSR reports, invites us to fundamentally rethink our relationship with technology and innovation. It confronts us with collective choices about the direction we wish to give these powerful tools, in a world with manifestly finite resources.
Faced with the identified ecological challenges, I see the notion of “technological proportionality” gradually emerging, the almost heretical idea in our culture of one-upmanship that the power deployed should be proportional to the real utility generated. This principle invites us to question without complacency: do we really need the most advanced and resource-hungry models for each use, or can we accept reasonable compromises to preserve our common environment?
This reflection on the right technological measure reminds me of those ancient wisdoms that already warned us against the excesses of immoderation, this hubris that the Greeks considered as the source of all human tragedies. It suggests that true sophistication does not reside in perpetual one-upmanship, but in the ability to find the optimal balance between means deployed and ends pursued, a form of technological minimalism that would also be an ethic of responsibility.
In this perspective, I imagine the future of generative AI not as a linear race towards ever more power, but as a diversification of approaches, each adapted to specific usage contexts, with their own constraints and their own criteria of excellence.
Recent studies confirm the emergence of this trend. According to a survey of companies regarding the technologies they are considering adopting to support their artificial intelligence initiatives, GPUs, these specialized processors as powerful as they are energy-hungry, surprisingly only come in fourth place in their priorities. On the other hand, increasing storage capacity and hybrid cloud solutions occupy the top positions. This trend is explained by the pragmatic need of companies to manage and store the massive volumes of data necessary to feed their AI models, rather than launching into a computational arms race.
More significantly, companies no longer seem to blindly prioritize raw computing power at the expense of efficiency when it comes to using generative AI. A recent study thus reveals that concerns related to energy consumption have reversed, with a 10% increase in concerns about energy impact. Organizations are gradually becoming aware that they can take advantage of AI while integrating energy-saving strategies, abandoning the illusion that more is necessarily better.
Environmental ethics
Taking into account environmental costs also transforms the ethical landscape of generative AI. Beyond traditional concerns such as privacy, transparency, or algorithmic biases, ecological impact is now imposing itself as a fundamental dimension of the ethics of these technologies, a pillar hitherto strangely absent from the charters and declarations of principles that flourish in the industry.
This environmental ethics of AI invites us to consider not only the immediate consequences of our technological choices on human users, but also their long-term implications for future generations and for all living beings. It broadens our moral horizon beyond immediate human interests alone, to integrate the health of ecosystems on which, ultimately, our own existence depends, a change in ethical paradigm whose magnitude should not be underestimated.
It also reminds us that the environmental costs of generative AI are not distributed equitably. The consequences of climate changes to which this technology contributes disproportionately affect the most vulnerable populations, the very ones who benefit least from its advantages, thus creating a question of environmental justice that cannot be ignored.
The digital divide is thus doubled by an ecological divide, drawing a global geography of inequality with multiple dimensions.
The responsibility of actors
To effectively respond to the environmental challenges posed by AI, an approach involving all actors, from developers to end users, through companies and regulators, is necessary.
But beyond declarations of intent and solemn commitments that flourish in international conferences, what concrete actions can truly bend the current trajectory?
- The role of developers: Developers and AI researchers have a crucial role to play in designing more efficient models. A few weeks ago, I participated in a workshop where we discussed the integration of sustainability metrics into the development process, on par with traditional performance metrics. The emergence of evaluation frameworks integrating the environmental impact of AI models could orient research towards more sustainable solutions, transforming the very criteria of what we consider a technological “advance.” Algorithmic efficiency must become a priority research area, with a particular effort on techniques to reduce energy consumption and material resources without significantly compromising performance. This includes the development of more efficient training methods, optimization of model architectures, and development of less resource-hungry inference techniques, an ambitious research program that will require a significant reorientation of funding and academic prestige towards these hitherto neglected questions.
- The responsibility of companies: Technology companies, particularly the giants of the sector that dominate the development of generative AI, have a particular responsibility in this transformation. They must not only adopt more sustainable practices in their own operations, but also show radical transparency regarding the environmental footprint of their models and services, a transparency that goes well beyond formatted CSR reports and vague commitments that too often characterize corporate communication on these subjects. The establishment of ambitious but realistic objectives in terms of reducing greenhouse gas emissions and energy consumption is essential. This could include the use of renewable energies to power data centers, improving the efficiency of cooling systems, and adopting circular economy principles for electronic equipment, concrete measures whose impact should be measured and verified by independent third parties. Some companies are already making efforts in this direction, which should be welcomed without naivety. For example, IBM reported that 74% of the energy used by its data centers came from renewable sources in 2023, while Google aims for 100% decarbonized energy supply by 2030. These commitments, if kept and verified, constitute steps in the right direction, even if they are not enough to compensate for the global increase in consumption linked to the continuous expansion of these infrastructures.
- The role of users: I am convinced: users of generative AI, whether individuals or organizations, also have a role to play in reducing the environmental impact of these technologies. More conscious and responsible use, favoring models adapted to their real needs rather than the most powerful by default, can significantly contribute to reducing the overall footprint, a form of digital minimalism that would also be an ethic of individual responsibility. Raising user awareness of the environmental costs of generative AI is therefore crucial. Tools allowing real-time visualization of energy consumption and CO2 emissions associated with the use of these technologies could encourage more responsible behaviors, transforming the invisible into visible, the abstract into concrete. I imagine a meter integrated into ChatGPT that would display in real time the CO2 equivalent of your conversation with AI, a transparency that might radically change our usage behaviors.
- The indispensable regulatory framework: Finally, regulators have a central role to play in framing the development of generative AI from an environmental point of view. Initiatives like the European legislative framework on AI, which underlines the necessity of taking into account the energy consumption of models, show the way towards more comprehensive regulation that could transform industry practices. The establishment of specific environmental standards for AI, including requirements in terms of energy efficiency and sustainability of materials used, could contribute to orienting the sector towards more responsible practices. Fiscal measures, such as carbon taxation adapted to the digital sector, could also incentivize actors to reduce their environmental footprint, as experience shows that economic incentives often remain more effective than appeals to virtue. Initiatives are already emerging in this direction, timidly but significantly. For example, in October 2023, California Governor Gavin Newsom signed a climate disclosure law that requires companies with annual revenue exceeding $1 billion to disclose their greenhouse gas emissions by 2025. At the European level, the directive on corporate sustainability due diligence aims to ensure that organizations conduct due diligence regarding impacts on human rights and the environment. These first regulatory stones, if they are effectively implemented and controlled, could progressively transform the AI landscape.
Beyond technological solutionism
The exploration of the material and environmental dimensions of generative AI confronts us with a reality often obscured by technological enthusiasm and marketing discourse: these apparently ethereal tools are deeply rooted in the materiality of the world, consuming limited resources and generating significant ecological impacts. They do not float in an abstract cloud, but rest on titanic infrastructures, materials extracted from the earth, and massive quantities of energy.
This awareness, far from being only constraining or guilt-inducing, opens new paths of innovation and progress. It invites us to develop more balanced, more efficient approaches, more in phase with the limits and riches of our planet. Algorithmic efficiency, model specialization, computation relocation, or frugal innovation constitute as many promising paths towards an ecologically viable generative AI, paths that require not less ingenuity, but differently oriented ingenuity.
More fundamentally, this exploration reminds us that artificial intelligence, like any human creation, does not escape the fundamental laws that govern the living. Like us, it depends on the finite resources of our planet and is part of cycles of exchange with its environment. Recognizing this deep kinship between artificial intelligence and natural systems could well be the key to a harmonious coevolution between these two forms of intelligence that will shape our common future, a paradigm shift that would invite us to think of technology not as a conquest of nature, but as a particular expression of it.
For ultimately, I do not believe the question is whether we can continue the development of generative AI without consideration for its environmental costs, we manifestly cannot in the long term, but rather how we can orient this development so that it contributes to a viable and desirable future for all living beings. It is in this quest for harmony, more than in the frantic race for raw power, that perhaps lies the true intelligence of our species.
As we find ourselves at a crossroads, with on one side the seductive promise of tools of unprecedented power and on the other the very real threat of unsustainable environmental pressures, we collectively have the responsibility to shape a generative AI that amplifies our creativity and our ability to solve the complex problems of our time, without compromising the ecological foundations of our existence.
This delicate but essential balance could well be one of the greatest challenges, and one of the most beautiful opportunities, of our time. It will require from us not only technological innovations, but also social, economic, and political innovations to create the framework in which a truly sustainable AI can flourish.
For technology alone, however brilliant it may be, will never be able to solve problems that are fundamentally systemic and that require a profound transformation of our modes of production, consumption, and governance.
I am intimately convinced: generative AI will not escape this reality. Its future will be ecological, or will not be.
This second part completed, next week we will explore the third and final part: Ethical issues and future perspectives