AI Sustainability: How Harmful is Artificial Intelligence to the Environment?

The sustainability of AI is a growing environmental dilemma. Artificial intelligence takes a heavy toll worldwide through immense power and water consumption. In 2024, AI accounted for 11 to 20 percent of global data centre electricity usage. Training heavy models like GPT-4 generates tremendous amounts of CO2, and the daily server infrastructure constantly requires millions of litres of cooling water (an average of 500 millilitres per ChatGPT conversation). Yet, simultaneously, artificial intelligence offers solutions for combating climate change. Ultimately, the true footprint depends on efficient hardware utilisation, the right choice of models, and green data centres.

The explosive rise of generative AI has a complex dark side. While the technology accelerates processes and increases human efficiency, it builds an alarming ecological footprint in the background. The sustainability of AI is no longer a theoretical debate, but an immediate challenge for businesses, governments, and consumers.

The Harsh Reality: AI's Energy Consumption in Numbers

The backbone of artificial intelligence consists of massive data centres filled with thousands, and sometimes tens of thousands, of Graphics Processing Units (GPUs). These servers run around the clock to both train new models and process daily queries (inference).

According to recent research by Vrije Universiteit Amsterdam (Alex de Vries-Gao, 2025), AI now consumes a significant portion of global computing capacity. In 2024, AI accounted for 11% to 20% of the total power consumption of all data centres worldwide. The electricity required to keep the global AI infrastructure running is comparable to the power usage of a major international metropolis or even a small country.

A classic search query uses relatively little energy. Generating a response with a large language model (LLM), however, consumes up to thirty times more power. This happens because generative models must compute complex mathematical calculations across all layers of their neural networks for every single word.

Expectations for the Near Future

The International Energy Agency (IEA) has also raised the alarm. The agency predicts that total data centre energy consumption will double by 2030, primarily driven by the scaling of artificial intelligence. Without intervention, this will place immense pressure on already overloaded electricity grids (grid congestion), jeopardising the transition to sustainable energy.

Task Type	Estimated Energy Consumption	Relative Impact
Google Search Query	0.3 Watt-hours (Wh)	Very Low
Basic ChatGPT Prompt	~2.9 Watt-hours (Wh)	Medium (10x traditional)
AI Image Generation	~4 to 10 Watt-hours (Wh)	High
Short AI Video Generation	> 20 Watt-hours (Wh)	Very High

Hidden Costs: Water Consumption and Cooling Systems

Beside electricity, the data centres hosting AI models consume staggering volumes of water. Computer systems generate excess heat. Without effective cooling, servers crash, hardware degrades, and fire risks emerge.

The most efficient way to cool massive server farms is through water cooling (evaporative cooling). This process consumes fresh drinking water that evaporates out of the local ecosystem. Independent analyses show that an average ChatGPT session (ranging from 10 to 50 related prompts) consumes approximately 500 millilitres of cooling water (equivalent to a small bottle).

When you consider that ChatGPT and similar tools are used by hundreds of millions of people weekly, the water footprint grows exponentially. This causes immediate problems in areas grappling with droughts and 'water stress', such as parts of the United States and Southern Europe, where many large data centres operate. Water is crucial for local agriculture and drinking water reserves, causing regular friction between tech giants and local governments.

CO2 Emissions: The Heavy Burden of Model Training

Within the realm of AI sustainability, impact is based on two fundamental phases: the training phase and the inference phase.

The Training Phase

Building a frontier <i>foundation model</i> is extremely energy-intensive. Supercomputers run at full tilt for months to configure billions of parameters and optimise weights based on massive datasets. Training a model the size of GPT-4 emits an estimated 25,000 tonnes of CO2 equivalent. By comparison, this equals the annual emissions of thousands of fossil-fuel-powered passenger cars.

The Inference Phase (Daily Usage)

Although training requires a burst of massive energy, the true long-term impact occurs during the inference phase: the daily processing of user prompts. Once a model becomes popular, usage emissions quickly surpass the one-off training costs.

These environmental factors fuel the call for regulation and transparency, and they are among all concerns about AI at a glance that critics rightfully highlight. Companies must evaluate not only their direct emissions (Scope 1 and 2) but also account for the CO2 emissions from procured IT services (Scope 3 emissions).

New insights (2025–2026): more transparency, uneven distribution

Two recent publications sharpen the picture further.

UN report on the environmental cost of AI (June 2025). Researchers from a United Nations think tank conclude that AI already accounted for roughly one-fifth of global data centre electricity use last year, and that AI water consumption is substantially higher than earlier estimates by the International Energy Agency. The distribution is also strikingly uneven: about 90% of global AI compute capacity sits in China and the United States, while countries in South America and Africa mainly bear the downsides through raw material extraction and climate damage. Making AI more energy-efficient only helps so much: the Jevons effect lowers prices and pushes usage back up, NOS reports on the UNU study.

Amazon discloses its water use for the first time (June 2026). According to The Verge, Amazon's global data centres consumed 2.5 billion gallons of water (≈ 9.5 billion litres) in 2025, at a rate of 0.12 litres per kWh — a 2% drop from 2024 despite continued expansion. Amazon says its data centres rely on air cooling about 90% of the time, switching to evaporative water cooling only during the hottest hours. Important caveat from the article: this figure excludes indirect water use at the power plants supplying electricity and at construction sites of new data centres. Amazon's own comparison places Microsoft, Google and Meta higher on water per kWh — though the cited Google number is specific to Gemini data centres rather than its full fleet.

The through-line: more transparency helps, but the absolute environmental load is still growing. For strategic decisions the advice remains the same — pick models deliberately, ask vendors for water and CO2 numbers, and look beyond just the meter inside the data centre.

The Sustainability Paradox: AI as the Saviour of the Climate?

Besides the evident damage that AI development inflicts on ecosystems, the technology has a fundamentally different side. AI has become one of the most powerful tools for sustainability innovation. This paradox makes the debate around AI sustainability twofold: its operations partly fuel the climate crisis, yet its applications could potentially slow it down.

Successes where AI concretely accelerates sustainability:

1. Energy Optimisation and Smart Grids: AI identifies patterns and optimises power grids in real-time. This significantly improves the yield of wind and solar farms by flawlessly matching supply and demand.
2. Weather Forecasting: AI models (such as Google DeepMind's GraphCast) predict extreme weather events much faster and more accurately than traditional meteorological models. This saves lives and limits infrastructure damage.
3. Materials Research: AI is used to simulate molecular structures for advanced battery technology, carbon capture, and biodegradable packaging. Research that typically takes decades now happens in weeks.
4. Precision Agriculture: Models analyse satellite imagery and drone data to drastically reduce the use of water, pesticides, and fertilisers, leading directly to reduced soil pollution.

Practical Steps: How Users Can Reduce AI's Impact

As an individual or professional working with AI daily, you have direct influence over the server capacity (and therefore electricity) your prompts demand. Make conscious choices about which model you deploy for each task.

• Match the right model to the task: Do not use heavy, expensive models (like GPT-4 or the largest variants of Claude or Gemini) for simple summarisation or text analysis. Instead, choose efficient <i>Small Language Models</i> (SLMs) or lightweight variants like Gemini Flash. This significantly reduces the load. For more on this, check out our guide on choosing energy-efficient AI tools.
• Limit image generation: Generating images or video uses many times more energy than generating text. Do so purposefully to avoid pointless iterations.
• Batching and Caching: Perform tasks in clusters when working with APIs. Save results locally (caching) so you do not have to trigger a new API request every time for common queries or data preparations.
• Reuse effective prompts: Work using structured, efficient prompt engineering. The quicker the AI model grasps exactly what you need, the less redundant computing power goes to waste on poor outputs and subsequent corrective prompts.

Corporate Policy: A Sustainable AI Strategy for Businesses

For organisations, integrating AI is no longer just a technical choice; it shares accountability for your ESG (Environmental, Social, and Governance) targets. A company that integrates AI carelessly across every department will quickly discover its cloud costs and carbon footprint expanding exponentially.

1. Demand Transparency from Suppliers

Do not accept that the climate impact of the software your company procures is a black box. Openly ask suppliers for clear reporting on power and water consumption (Water Usage Effectiveness - WUE), and CO2 equivalents. Companies like Microsoft (Azure), Google, and Amazon (AWS) document the extent to which they run on sustainable energy.

2. Data Centre Location

The power mix varies hugely by region. Are you hosting or fine-tuning your own models in the cloud? If so, select server regions with a green energy grid share.

• United States (central): Often reliant on coal or gas.
• Iceland and Norway: Operate almost entirely on hydroelectric or geothermal cooling.
• The Netherlands: Offers local data centres that use a vast array of renewable sources, leaning heavily on offshore wind power.

3. Efficiency in Architecture and Local Models

The more compact the model, the more sustainable it is. Train smaller, domain-specific open-source models locally (on-premise or in an internal cloud). Running lightweight models requires less electricity than constantly pinging data back and forth to massive monolithic APIs on other continents.

A solid IT plan accounts for all these technical environmental variables. This is the crux of drafting a sustainable AI policy where you do not sacrifice speed of innovation, yet act responsibly towards the climate.

The Future: New Hardware and Regulation

The call for energy efficiency is driving new hardware innovations. Chips are being developed with ever-increasing efficiency and specific focus on AI calculations. Conversely, software is evolving too: fundamental new model architectures, such as 'Liquid Neural Networks' or more efficient variants of the dominant 'Transformer' architecture, promise to drastically reduce the required compute power per prompt in the future.

Legislative bodies are not sitting still either. Following the European AI Act, regulators are deliberating over transparency requirements concerning the ecological footprints of foundation models. As a result, AI developers will not only compete on intelligence, but specifically on operational sustainability.

You do not want to miss this playing field or catch up too late. For structural analyses and the broader in-depth economic impact of these technologies, take the time to carefully scan and download the ai.nl reports. These will instantly provide you with the tools to shape your strategy for the coming years.