Ecology and Artificial Intelligence

The rise of AI is generating unprecedented ecological pressure on water and electricity resources. Facing this challenge, decentralization and reclaiming digital sovereignty appear as pathways to sustainable solutions.

The Dialectics of Centralization-Decentralization in Computing

One of the lesser-documented challenges of artificial intelligence is that AI provides so many services that it becomes difficult to examine the “unpleasant” aspects—the “hidden side” of AI. I believe it is important to remain clear-eyed to better understand the world we live in. Then, how can we take action and change things? That is much harder to say and is up to each individual. I propose an introduction to the dialectic between Artificial Intelligence and ecology, technically up-to-date as of 2025 when this article was written, and I conclude with concrete proposals.

Artificial intelligences consume natural resources at a much higher rate than traditional data centers, which host websites and other databases—banking, photographic, video, textual, etc. Today, AIs primarily operate in data centers: the computations that produce the results we receive on our computers and phones are not performed on our devices. We send a query from our terminal to a data center, which generates the result using immense computing power, and we receive it on our phone or computer.

Yes, there are also embedded software applications in devices, and it is possible to run a generative AI on one’s own computer without using a data center—though a data center was still required for the initial training of that AI. Afterward, one can potentially install it on their own computer. However, what becomes immediately apparent is the extreme slowness of the response. You can hear your computer straining, the fan spinning at maximum speed, and yet the answers take several minutes to generate—and they are of rather poor quality. This is nothing like the instant results obtained when querying an AI via the internet. Through this experiment, we can gauge the computing power required to generate relevant results.

Moreover, our personal computers cannot run large language models, only small ones, because they lack sufficient RAM to load the entire dataset of a language model into memory to produce a response. So not only are they extremely sluggish in delivering even the most basic, imprecise answers (due to using a small language model), but they simply lack the technical capacity to run large language models—the ones we find useful in daily life.

Thus, we return to a highly centralized system, similar to the Minitel. The Minitel, introduced in France in 1982—over a decade before the arrival of the internet—was a terminal that performed no calculations itself, unlike our modern personal computers and smartphones. It was a terminal that accessed large computers in data centers, which processed requests and sent the results back to the Minitel screen. Then came personal computers, which allowed us to create documents, edit images, and more. The Minitel quickly became obsolete because it had no inherent capabilities—it could do nothing without a network connection and was entirely dependent on external services. Additionally, it tied up the landline (the only one in most households at the time), and billing was per minute of connection. Mobile phones did not yet exist; they became widely available in France around 1997.

After the Minitel, personal computing arrived, marking Microsoft’s immense victory through its monopoly on personal computers with its Windows operating system and many other products. The next step was the arrival of the internet. Bill Gates, Microsoft’s founder and CEO at the time, did not believe in the internet at all and even sought to slow its development because Microsoft’s business model at the time revolved around selling software for the personal (and professional) computer ecosystem. But the internet arrived and revolutionized the world.

A Synthetic History of Data Centers

• In its first phase, starting in 1995, the Internet was a space for data consultation, providing access to documents—whether textual, audio, video, or visual. It was not about offloaded computations but rather accessing external data, with calculations still being performed on our own computers.
• Then came the smartphone in 2007 with the introduction of the iPhone, which completely changed the game. The phone began to mimic computers, meaning its functions could be modified by adding software, known as applications. This was unlike pre-iPhone phones, where functions were hardwired into the device’s hardware. From the iPhone onward, functions largely depended on the software installed.
• With the success of smartphones, cloud computing emerged—enabling seamless access to documents from anywhere.
• From this point onward, centralized computing and personal document management services were revitalized, unlike the early Internet, which only dealt with public documents (websites). With the rise of smartphones and the need to access personal data, cloud computing expanded. For example, Google Docs or Windows 365 allow access to the same data whether using a computer, phone, or even another device by logging into an account. Thus, data centers, initially used for collective administrative and heritage data as well as web hosting, gained a new purpose: the increasingly massive hosting of personal data.
• This storage, management, and access service for personal data also grew significantly for businesses and organizations. For instance, a municipality, instead of handling data storage, managing servers, networks, maintenance, and backups itself, could subscribe to a third-party data center that handles all of this—securing data, regularly offering new services for the same subscription cost, etc. In a way, we are relinquishing control of our data. This has short-term economic benefits, as economies of scale allow these large data centers to offer attractive subscription prices, reduce staffing needs, physical infrastructure management, and liability. However, this is neither beneficial, economical, nor sustainable in the long run—as we will explore later in this article.
• Then, in late November 2022, ChatGPT arrived—the first generative text-based AI—ushering in widespread personal and collective awareness and use of artificial intelligence in daily life (unlike, say, GPS navigation, which isn’t perceived as AI usage). These generative AIs are not just for “conversational” purposes. For example, Windows Copilot provides suggestions embedded within software itself, enhancing efficiency—a core promise of computing since its inception. This promise is continually renewed by tech companies to stay competitive in a rapidly evolving market where usage and technology can shift quickly. We stick with the manufacturer or software provider that offers the best efficiency. This is why AI tools that simplify our lives are being developed—and why we’ll prefer services that make life easiest.
• For these generative AIs to function, as mentioned earlier, immense storage and computing resources are required. Thus, a fourth category of data centers (even if physically located in the same facilities) has emerged: massive decentralized computing. We are, in a way, returning to the Minitel logic—today, our devices perform very little computation themselves, instead offloading calculations and delivering only the results to us.

How a Data Center Works

Now, let’s explore, based on these historical reminders, how these computing data centers function, in order to introduce the ecological issues tied to artificial intelligence. When these data centers are established in a given area, they exploit natural resources on a massive scale. It’s important to note that there are several types of data centers, various machine cooling technologies, and different kinds of components, processors, hard drives, etc. Therefore, to avoid an oversimplified view, it’s crucial to understand that water and electricity consumption also depend on how the data centers themselves are designed.

Data centers primarily use three material resources: two local resources and one external resource.

• The two local resources are electricity to power the machines and water to cool them.
• The external resource includes all the technical hardware they are made of, such as hard drives, processors, and related components.

I will focus here on the local resources—electricity and water—keeping in mind that the manufacturing of processors, which involves the consumption of rare earth elements among other things, is also a highly non-ecological process.

The Growing Demand for Electrical Resources

Electricity is required to power the processors for computations and the hard drives to store the results of these calculations, which generates heat that must be cooled, most often using water. Let’s focus on electricity and project into the coming years. The computing needs for artificial intelligence are such that the future data centers planned by the three major industrial players—Google, Microsoft, and Amazon, the primary providers of data centers for the next three years—will each consume 10 gigawatts to meet the growing demand. Indeed, the efficiency brought by AI to users generates savings, new services, loyal customers, etc.

10 gigawatts per data center provider—meaning 30 gigawatts in three years—might seem like just an electricity consumption figure. But 10 gigawatts means 10 gigawatts of annual electricity consumption. How much electricity does an average nuclear power plant produce? 1 gigawatt per year.

This means that within a three-year horizon, to produce the electricity needed for these future data centers—which will inevitably be built because they meet the demands of humans who consume AI extensively—we would need to construct 30 new nuclear power plants! A nuclear plant isn’t built in three years but in 5 to 15 years. This level of exponential increase in electricity consumption is unprecedented. Estimates suggest that for the following three years, around 2030–2031, an additional 30 gigawatts would be needed, meaning 30 more nuclear plants… This information comes from Eric Schmidt, former CEO of Google, in early 2025.

What does this mean in concrete terms? How can such a demand be met? It’s impossible to build so many nuclear plants so quickly—and not necessarily desirable from an ecological standpoint! Today, the computing and data industry consumes 3% of the world’s total electricity production. And if we follow these projections—which are realistic given the increasing use of AI—within a few years, data centers could consume 99% of the world’s total electricity production! (Again, according to Eric Schmidt).

The development of these AI services cannot be halted due to this constraint. The rapid evolution of demand, driven by profit-seeking industries, will force these same industries to find solutions very quickly—solutions that may not be very sustainable from an ecological perspective.

Water usage

Let’s now turn to water usage. Water is essential for cooling servers. There is a massive dissipation of heat due to the operation of these microprocessors, which must be cooled. If they were not cooled, they would simply stop functioning, because without cooling, their temperature would rise uncontrollably, and they would melt. We are therefore forced to cool these systems. Perhaps one day, thanks to future quantum computers, we will be able to build processors that generate less residual heat than current ones. This is precisely the topic I was discussing earlier: the way these data center machines are built is, in my opinion, the real ecological issue—to produce as much computing power as needed without causing such severe environmental side effects. It is by redesigning the production system itself that we can limit its environmental impact, in my view.

Where does this water come from? Well, mostly, it comes from the drinking water network of the area where the data center is located… For the same reasons of speed—how do we quickly find the water we need immediately to cool the machines? Well, by connecting to the drinking water supply. Google, for example, uses 75% drinking water in its water consumption. It boasts that 25% of the water it uses is recycled or non-potable (according to Ophélie Cœhlo in Geopolitics of the Digital, 2023). We know how scarce and precious water is. During heatwaves, we may all face collective water use restrictions (just as in winter, there may be collective electricity restrictions due to resource shortages). But how do we arbitrate during heatwaves between the water available for citizens and the water needed for data centers? Can we reasonably, as a local community or state, choose to impose reduced capacity or even temporary shutdowns on data centers? Unfortunately, this is simply impossible. Indeed, as these data centers house the data of the local authorities themselves, because they use data outsourcing, as I explained in the article Digital and rationality, if we were to reduce, at times, the water supply to the data centers, well, we would no longer have access, potentially, to our own data, and so we would block the functioning of the services of a territory, a country, banks, communications, hospitals, etc.

Translated with DeepL.com (free version) There is therefore a hierarchy that is not even political but purely functional: water will be given to data centers before it is given to people.

Is there any virtue left in this world?

A historic Swiss data center (Infomaniak) has chosen to use the heat produced to warm 6,000 homes in Geneva and to keep older machines running its data center, which mixes new and old machines for data requiring less computing power, such as website data, for example. This data center also hosts artificial intelligences. So, it may be more virtuous to use AI services hosted in such places. But the way AI services are integrated into major monopolies—meaning directly into your Microsoft or Google tools, for instance—necessarily limits our personal ability to choose which data center to use.

This is why an introduction based on a historical and technological perspective seems necessary to understand the fundamentals of the ecological impact of artificial intelligence. I believe it is very important to continue educating ourselves on this topic, because our individual choices, given how many of us there are, can have significant impacts. And the decisions made by communities in which we can play a role (as agents, elected officials, or citizens) must be made by people with informed reflections on choices that may seem more economically reasonable at first glance but could, in the relatively near future, prove quite catastrophic ecologically—and in terms of responsibility and the ability to exercise it directly—due to unquestioned dependence on large industrial players who pay little heed to ecological concerns, out of necessity.

Paths for the future

My suggestion is to reclaim control over our data, its storage, its use, and its processing. The tools already exist. The most ambitious artificial intelligence systems are open source, and we can install them wherever we choose. And it seems to me that there is a major ecological imperative in taking back responsibility ourselves for the technical infrastructures that power our systems. This way, their usage can become far more reasoned. I’m not talking about frugality, because artificial intelligence provides many highly useful services, but as soon as we decentralize, people assume we lose economies of scale—and indeed, in purely capitalist terms, it’s less efficient. However, in terms of responsibility, mutual aid, cooperation, collective intelligence, and reclaiming sovereignty over our data and the computations we need to function, we regain agency over our technical and ecological choices.

The cooperative approaches of free software enable skill-building in a decentralized way, across all these small mini-hubs that produce decentralization rather than dependence on a single monopoly. Cooperation is essential. The creation of the commons relies on decentralization and communication among all these decentralized mini-hubs. In other words, cooperation cannot go hand in hand with centralization. We even see this in the management structures and employee practices of large corporations: big companies that want to operate flexibly and agilely must reorganize themselves into micro-units that cooperate and mutually enrich one another. This yields far more compelling results than an overarching centralization, which shows its limits both in terms of contributing to collective intelligence and, moreover, in terms of ecological sustainability.

Let’s take back our data, decentralize, cooperate, and become tangibly and materially responsible for the data that belongs to us—to open up a potential digital future that is less destructive to life than what currently looms ahead.