Can AI Actually Save the Climate?
In 2026, AI is no longer a conceptual lab experiment; it's in production, analyzing petabytes of satellite data, predicting droughts, and steering wind turbines. At the same time, the electricity consumption of the world's digital giants is climbing to levels not seen in decades. The equation is uncomfortable, and deserves to be confronted head-on.
Yes, artificial intelligence already has concrete tools to fight climate breakdown: optimizing power grids, modeling extreme weather, monitoring forests and oceans in real time. But that same technology consumes staggering amounts of energy and water to run. This paradox isn't a reason to reject AI, or to idealize it. It's precisely the paradox we need to highlight.
In Short
AI is already deployed in climate modeling, renewable energy management, and ecosystem monitoring, with measurable results.
Projects like DeepMind's application to Google's own data centers have delivered significant energy efficiency gains, on the order of 30 to 40%.
In return, global electricity demand tied to AI is surging: the International Energy Agency (IEA) now projects data center electricity consumption could roughly double from 485 TWh in 2025 to 950 TWh by 2030, about 3% of global demand.
AI's climate "energy paradox" is real: using an energy-hungry technology to save energy elsewhere.
Digital sobriety and regulation (notably the EU AI Act) are emerging as necessary conditions for AI actually to deliver on its environmental promises.
What AI Is Actually Doing for the Climate (and It's Not Trivial)
Before diving into the darker angles, credit where it's due. In specific domains, AI is a genuinely useful technology for the ecological transition, not in some hypothetical future, but right now, in active deployment.
Climate Modeling: A Field Accelerated by Algorithms
Climate scientists have worked for decades with general circulation models, simulations that reproduce interactions between the atmosphere, oceans, vegetation, and cryosphere. These models can take days, sometimes weeks, to run on supercomputers. AI, and specifically deep learning, is starting to change that dramatically.
In 2024, the European Centre for Medium-Range Weather Forecasts (ECMWF) rolled out its Artificial Intelligence Forecasting System (AIFS), a machine-learning-based forecasting model that now runs operationally alongside its traditional physics-based system, and can generate forecasts more than ten times faster while cutting energy consumption by roughly 1,000-fold. Google DeepMind, for its part, released GraphCast, a weather model trained on decades of global atmospheric reanalysis data. According to results published in Science in November 2023, GraphCast outperformed the ECMWF's traditional operational model on more than 90% of tested variables, using a fraction of the computing power.
These gains are really exciting. This will enable better weather forecasts, which means better-managed power grids, earlier warnings of extreme heat, smarter irrigation planning, and faster disaster response.
Optimizing Renewable Energy: AI Steers the Wind
Wind and solar power share a structural flaw: intermittency. Wind doesn't blow on demand, and the sun sets every day. To integrate these sources into power grids at scale, their output needs to be forecast with real precision, exactly where machine learning adds measurable value. In 2019, DeepMind published results from applying its models to 700 megawatts of wind farm capacity in the central United States: by predicting output 36 hours in advance, the system boosted the economic value of the electricity generated by roughly 20%, by reducing the mismatch between supply and grid demand.
More recently, AI has also been used to optimize offshore wind farm layouts, spacing turbines to reduce aerodynamic interference (the "wake effect"), and to predict maintenance needs before failures occur. At scale, these efficiency gains add up to millions of tons of un-emitted CO₂.
Forests, Oceans, Biodiversity: AI as Planetary Sentinel
Deforestation often advances in remote areas that are difficult for human teams to monitor. NGOs and governments increasingly use deep learning models paired with satellite imagery to detect illegal logging in near-real time. Global Forest Watch now goes a step further, using AI to identify not just where deforestation is happening, but what's driving it, distinguishing large-scale agriculture from mining, wildfire, or road development across the Amazon, Congo Basin, and Indonesia.
In the ocean, Mercator Ocean International's Digital Twin of the Ocean is emerging as a major milestone in modern oceanography, building a dynamic virtual mirror of the marine environment. Far more than simple mapping, the platform blends classical physics-based equations with machine learning, through its AI forecasting system GLONET, continuously correcting for measurement bias and modeling complex planetary-scale variables in seconds rather than hours, with a roughly 23% improvement in ocean current prediction accuracy over traditional models.
For more on AI's role in unexpected domains, see our article on AI and space exploration, many of the technologies developed to observe Earth from orbit feed directly into climate monitoring.
The Paradox of Using AI for Climate
Even when AI serves climate goals, it consumes energy itself, and not a little. That's not a reason to reject the use cases outlined above. It's a reality that deserves equal scrutiny.
Data Centers With an Appetite That Won't Quit
In its 2026 report Key Questions on Energy and AI, the International Energy Agency (IEA) laid out a projection worth taking seriously: global data center electricity consumption could roughly double from 485 TWh in 2025 to 950 TWh by 2030, about 3% of global electricity demand. AI-focused data centers are growing even faster, with electricity consumption from AI-specific facilities projected to triple over the same period.
To put that in perspective: total data center demand by 2030 would be in the same range as the annual electricity consumption of a large industrialized economy like Japan.
Training large language models like GPT-4 or its competitors also carries a considerable energy cost. A 2019 study from the University of Massachusetts Amherst, published at the Association for Computational Linguistics, estimated that training a model like BERT could emit roughly as much CO₂ as a transatlantic round-trip flight.
This topic sits at the heart of our deep dive on the environmental cost of AI worldwide, which breaks down the real numbers behind global digital infrastructure's water, energy, and rare-earth footprint.
Water Consumption: The Debate's Blind Spot
Electricity isn't the only resource at stake. Data centers need cooling, and evaporative cooling consumes enormous quantities of water, an increasingly stressed resource in a warming world.
A widely cited 2023 study by researchers at the University of California, Riverside, first released as a preprint and later published in Communications of the ACM, estimated that training GPT-3 consumed roughly 700,000 liters of water. The same research group's later analysis suggests global water demand tied to AI servers could reach the range of several billion cubic meters annually within the next few years. In regions already under water stress, Arizona or Spain, for instance, AI is becoming a genuine competitor with agriculture and household use.
The Rebound Effect: Saving Here, Wasting There
The history of technological efficiency is marked by a well-documented economic phenomenon: the rebound effect (or Jevons paradox). When a technology becomes more efficient, we tend to use more of it, canceling out some or all of the gains. LED bulbs use at least 75% less electricity than incandescent bulbs, according to the U.S. Department of Energy, yet global artificial lighting use has risen overall since their adoption.
The same risk exists with AI. Efficiency gains from predictive energy algorithms could be absorbed by mass adoption of increasingly resource-hungry AI services. There's no scientific consensus yet on the scale of this rebound effect for AI specifically, but the question is becoming more central by the year.
This mechanism echoes broader questions about the real risks of AI and automation for society, a useful companion read for situating these environmental stakes within a wider systemic perspective.
Can AI Decarbonize Itself?
This isn't a rhetorical question. A growing number of players are working to shrink the footprint of the models themselves, and some results are genuinely encouraging.
Model Efficiency: Doing More With Less
One of the most significant advances of recent years is the rise of "compact" models, architectures trained to match the accuracy of their larger predecessors using a fraction of the parameters.
Transfer learning, reusing a model pre-trained on a related problem instead of starting from scratch, is also a powerful lever for efficiency. Training a model on 100 hours of GPU time is fundamentally different from training it on 10,000.
Our dedicated article on transfer learning explains exactly how this technique lets AI learn without relearning everything from zero, with direct implications for its energy footprint.
Renewable-Powered Data Centers: A Partial Reality That's Proving Hard to Reach
Google, Microsoft, Amazon, and Meta have all made ambitious renewable energy commitments. Microsoft is targeting 100% renewable, round-the-clock ("24/7 clean energy") power by 2030, alongside pledges to be carbon negative and water positive by the same date. It's worth treating these numbers with some skepticism: Microsoft's own 2026 Environmental Sustainability Report shows the company's emissions actually rose by roughly 25% year over year, driven largely by the buildout of AI and cloud infrastructure, even as it invests heavily in efficiency measures like direct-to-chip liquid cooling.
Regulation as an Accelerator of Sobriety
The European Union has positioned itself as the pioneering jurisdiction for regulating tech's environmental footprint. The EU AI Act, which entered into force in 2024, now imposes strict transparency obligations. Under Annex IV, providers of high-risk AI systems must document the computational resources used to develop, train, and validate their systems as part of their technical documentation. Separately, Annex XI requires providers of general-purpose AI (GPAI) models specifically to disclose known or estimated energy consumption figures.
A Few Concrete Projects Where AI Is Making a Visible Difference
Beyond the big-picture principles, here are some documented examples of what "AI for climate" actually looks like in practice.
Precision Agriculture: Feeding More While Watering Less
Agriculture accounts for roughly 70% of global freshwater consumption, yet water-use efficiency in many countries remains below 50%, according to the International Atomic Energy Agency's (IAEA) work on agricultural water management. How can AI help close that gap?
In precision viticulture trials in the United States, researchers found that real-time, variable-rate drip irrigation controlled by sensor data and machine learning delivered a 26% yield increase alongside a 16% improvement in water-use efficiency, evidence that AI-driven irrigation can meaningfully reduce water waste without sacrificing output.
Detecting Leaks in Water and Gas Networks
Water and natural gas distribution networks can have numerous leak points, representing a sometimes considerable loss of resources. In the United States, the Environmental Protection Agency (EPA) estimates that roughly 16% of treated drinking water, around 2.1 trillion gallons a year, is lost through leaks in distribution systems before it ever reaches a tap. AI systems that analyze pressure variations across pipe networks can localize leaks within hours, where manual inspection once took weeks.
Similar approaches could plausibly be extended to natural gas networks, to cut losses of a resource that itself contributes to warming.
What the IPCC and Scientists Really Say About Climate AI
The Intergovernmental Panel on Climate Change (IPCC) is not yet an unconditional cheerleader for AI. Its reports acknowledge the technology's potential while flagging the risks of unchecked development.
In its 2022 Climate Change report, the IPCC names digital technologies, including AI, sensors, the internet of things, and robotics, as tools that can improve energy management and accelerate the adoption of low-emission technologies across sectors including energy, transport, agriculture, and buildings. But the authors are careful to note that these gains can be offset by the ever-expanding digitalization of services and use, partly driven by AI itself.
That nuance matters. AI isn't a magic wand. It can make solutions more efficient, cheaper, faster to deploy. But it doesn't create solutions out of nothing, it accelerates and optimizes what humans decide to do.
Digital Sobriety and AI: Can We Really Have Both?
Digital sobriety isn't a Luddite posture. It's a conceptual framework that demands a simple question before every technological deployment: does this AI service deliver enough environmental or social value to justify its resource cost? It's a question the tech industry still resists asking outright, though some players are beginning to build it in.
"Climate AI" Is a Bet on the Future That Deserves a Clear-Eyed Contract
It would be comfortable to end on a reassuring note. Something like: AI will save the climate, trust it. Or the opposite: AI is an environmental catastrophe disguised as a solution.
The reality is more complex, and more interesting. Artificial intelligence is a multiplier. It amplifies whatever we ask it to amplify. Deployed within power systems run on renewables, built with architectures designed for frugality, embedded in serious climate policy, it can massively accelerate the transition. Left to market forces alone, powered by still-carbon-heavy grids, with no obligation to disclose its footprint, it can become one more aggravating factor.
The paradox, then, isn't technological. It's political and ethical. Humanity has built a tool more powerful than anything that came before it. What's left is deciding, collectively, how to use it. And on that front, the algorithms don't get a vote.
FAQ
Does AI really consume a lot of energy?
Yes. AI is a particularly energy-intensive technology. The IEA projects that global data center electricity consumption could roughly double from 485 TWh in 2025 to 950 TWh by 2030, about 3% of global demand, and in the same range as a large industrialized country's total annual electricity use. Training large language models adds to that cost: training a model like BERT can emit roughly as much CO₂ as a transatlantic round-trip flight, and training GPT-3 was estimated to have consumed around 700,000 liters of water for cooling.
Can AI predict natural disasters?
Yes. Deep learning is transforming weather forecasting by analyzing petabytes of satellite and atmospheric data. Google DeepMind's GraphCast model, for example, outperforms the traditional European forecasting model on more than 90% of tested variables, while using a fraction of the computing power. These algorithms help predict droughts and anticipate extreme weather events, supporting better irrigation planning and faster emergency deployment ahead of disasters.
Can AI replace climate policy?
No. AI can't create solutions out of nothing, it can only accelerate and optimize decisions humans have already made. The IPCC notes that AI's efficiency gains can be offset by the growing digitalization of services more broadly. The final call remains political and ethical: algorithms don't vote on the policies needed to enforce sobriety or regulate markets.
What does digital sobriety mean applied to AI?
It's a framework that requires evaluating, before every technological deployment, whether an AI service delivers enough environmental or social value to justify its resource cost (water, electricity, infrastructure). In practice, that means optimizing models (compact architectures, transfer learning to avoid retraining from scratch) and enforcing regulatory transparency, as with the EU AI Act, which requires providers to document their systems' energy and computational footprint.
Which companies are furthest ahead on climate AI? Google / DeepMind:
applying machine learning to cut its own data center cooling energy by up to 40%, building the GraphCast weather model, and predicting output across 700 MW of U.S. wind farms. Microsoft, Amazon, and Meta: committed to 100% renewable energy sourcing by 2030, though Microsoft's own emissions have risen since setting that target, illustrating how difficult the goal is proving in practice. Mercator Ocean International: developing the Digital Twin of the Ocean with its GLONET AI system to model the marine environment.
Comments