A Dutch researcher has published new estimates suggesting that artificial intelligence systems consumed extraordinary amounts of resources in 2025, with carbon emissions rivaling those of an entire major metropolitan area and water usage surpassing the global bottled water industry combined.
Alex de Vries-Gao, who established Digiconomist to examine the environmental costs of digital technologies, released the analysis on Wednesday. His work represents an early effort to isolate AI‘s specific environmental footprint as conversational AI tools from companies like OpenAI and Google gained widespread adoption throughout the year.
The calculations paint a striking picture of AI’s resource demands during a period of rapid expansion for the technology sector.
Table Of Contents:
- Why AI’s Carbon Footprint Is Suddenly Everyone’s Problem
- The Infrastructure Behind AI CO2 Emissions 2026
- AI, Water Use, and Local Stress
- How Honest Are Current AI Emissions Numbers?
- Is AI Purely a Problem for The Climate?
- Zooming Out: AI And Global Power Demand
- What You Can Do About AI’s Carbon Footprint
- How Companies Are Trying To Adapt
- How You Can Stay Informed and Push For Better AI
Why AI’s Carbon Footprint Is Suddenly Everyone’s Problem
Until a couple of years ago, most people barely thought about data centers. They were invisible background infrastructure, humming away somewhere far from daily life. Then generative AI landed, and the demand for compute went through the roof almost overnight.
New research on AI’s environmental impact, published in 2025, estimates that AI alone could have emitted tens of millions of tonnes of CO2 in 2025. That puts it in the same rough range as a major global city. The technology is still in its early expansion phase, meaning annual emissions are likely to climb further.
On top of that, the same study found AI-related water consumption already exceeds total bottled water demand worldwide. That is not a rounding error. It means your late-night chatbot session has hidden links to rivers, cooling towers, and local drought risks.
This situation mirrors the climate crisis we see in other industries. As AI systems become more complex, they require more physical resources. This makes the digital cloud heavier than it appears.
The Infrastructure Behind AI CO2 Emissions 2026
Every AI model lives inside a data center. These buildings are basically warehouses of servers that eat electricity and spit out heat. The bigger and smarter the model, the more power-hungry it tends to be.
A single data center is a massive operation. When you cluster data centers together, the regional power load becomes immense. This concentration creates a specific center power challenge for local grids.
Managing center power demand is keeping utility providers awake at night. As data center power requirements spike, the grid struggles to keep up. This increase in data center power demand often forces utilities to bring older fossil fuel plants back online.
The International Energy Agency has warned that AI-focused facilities now pull as much electricity as heavy industries like aluminium smelting. Power consumption is set to more than double by 2030 if current trends hold. This growth is driven in large part by Gen AI adoption.
To bridge the gap between renewable supply and constant demand, many operators are turning to natural gas. While cleaner than coal, natural gas still contributes to the problem. Data center operators are often forced to choose between reliability and sustainability.
This sector could need around 23 gigawatts of power on its own, almost twice the demand of the Netherlands.
| Metric | Estimate | Rough Comparison |
|---|---|---|
| AI-related emissions 2025 | Up to 80 million tonnes CO2 | Similar to a large global city |
| AI share of data center power (2025) | Nearly 50 percent | Half of all data center electricity |
| Projected AI power demand | 23 gigawatts by end of 2025 | Almost twice the Netherlands |
| Comparison to travel | Rivals global aviation | A significant chunk of transport |
Those numbers explain why AI CO2 emissions in 2026 is on the radar for climate researchers and city planners. This is starting to look less like a side issue and more like a whole new chunk of global electricity demand.
AI, Water Use, and Local Stress
Energy is only half the picture. AI is also thirsty. The same 2025 Patterns study that mapped out emissions found that AI-related water use had already topped the demand for all bottled water on Earth.
Facilities cool their servers using air, water, or a mix of both. In dry regions or areas already facing water stress, that cooling load matters. Places like the Middle East, which are pushing for rapid AI development, face distinct challenges regarding water scarcity.
Even tech hubs like San Francisco are grappling with resource constraints as compute density increases. Australia and India are also pushing back as big tech companies scout locations for the next round of server farms. Center operators must now account for local water politics.
In India, roughly $30 billion is going into projects. Because the grid can be unstable, many sites are expected to build large diesel generator farms as backup. A recent review from KPMG called this trend a massive carbon liability, since it ties AI expansion directly to fossil fuel use.
How Honest Are Current AI Emissions Numbers?
To truly understand AI’s impact, we must look at carbon emissions from multiple angles. There are operational emissions, which come from the electricity used to run the servers. Then there are upstream emissions, which come from manufacturing the chips and building the data centers.
Analysis of tech sector climate reports from 2020 to 2022 suggested that emissions from some AI leaders may be several times higher than reported. One investigation found they could be more than seven times undercounted due to creative ways of assigning emissions between “owned” facilities and outside providers.
On the company side, you can see the tension. Google shared that it cut data center energy emissions by about 12 percent in 2024, using more clean power. However, it also warned that meeting its climate goals has become far more challenging as AI loads grow and carbon-free energy does not scale fast enough.
The total lifetime emissions of an AI model include its initial training and its daily usage. Model training is an intense burst of energy use. However, the ongoing use of the model for millions of queries can dwarf that initial cost over time.
Is AI Purely a Problem for The Climate?
This is where it gets interesting. AI can raise emissions through heavy computing and data storage, but it can also help cut emissions in other parts of the economy if we use it wisely. AI applications are being designed to optimize energy grids and logistics.
Many experts argue that AI is a tool for climate change mitigation. We need a solid change mitigation strategy that utilizes these advanced tools. A comprehensive mitigation roadmap often includes digital optimization.
Boston Consulting Group estimated that smart use of AI in areas like energy systems, logistics, building management, and industry could reduce global greenhouse gases by five to ten percent by 2030. That is a big prize if the gains outweigh the extra load from AI itself. This climate change mitigation roadmap suggests a dual role for technology.
The report highlights how efficiency gains can scale. A separate analysis confirms that industrial AI can slash waste significantly. The challenge is ensuring the energy efficiency of the AI models themselves does not cancel out these savings.
The problem is that generative AI is far more compute-heavy than older, narrower tools. It handles rich media, long prompts, and billions of daily calls. As research in the ACM Digital Library shows, moving from simple prediction tasks to generative text and video ramps energy per query sharply.
Not all AI usage is equal. Some interactions sip power, while others chug it. The neural network architectures required for complex tasks are massive.
Machine learning has evolved from simple classification to complex content creation. Generative AI‘s ability to create creates a massive energy load. When OpenAI’s ChatGPT launched, it signaled a shift toward these high-intensity models.
Studies on prompt complexity and response style have shown that certain long, intricate prompts can trigger up to fifty times more emissions than simpler ones. One recent communication study underlined this spread, pointing out how prompt design can become a real climate lever at a large scale (Frontiers, 2025).
Image generation is particularly energy-intensive. AI creates pixels through thousands of iterative calculations. Performing these specific tasks requires high-powered GPUs running at full throttle.
Another estimate from Reclaimed Systems, based on reporting in Forbes, suggested that some AI video tools might burn around one kilowatt-hour of energy per short clip. That is the same order as running a washing machine cycle just to spit out a few seconds of content.
Zooming Out: AI And Global Power Demand
By the time we hit 2026, AI will already have shifted parts of the global electricity landscape. Annual electricity consumption for the tech sector is skyrocketing.
The International Energy Agency expects that, by the end of the decade, total AI-related power demand could be similar to the current electricity use of Japan. Japan is one of the largest energy consumers on Earth. The outlook suggests we need massive infrastructure upgrades.
The agency also notes that roughly half of that extra demand is likely to come from non-renewable sources unless policies change quickly. This demand is alarming for sustainability targets.
Goldman Sachs has a similar view. Their analysts project that data center electricity needs, driven largely by AI, could jump by one hundred sixty percent by 2030. That growth path raises serious questions about where the new generation will come from and how clean it will really be
.
Morgan Stanley has also weighed in on the financial implications of this energy crunch. Annual electricity costs are becoming a major line item for tech firms. The technology requires energy reliability that few grids can currently offer without fossil fuel backup.
For many people, the key concern is simple. Will the AI boom slow our climate progress just as we need emissions to drop sharply, or will it pressure leaders to speed up clean power investments instead of delaying them?
What You Can Do About AI’s Carbon Footprint
You may feel tiny in the face of all this, like one chatbot user does not matter. But there are ways that consumer pressure shapes corporate policy.
First, getting familiar with how a carbon footprint is measured helps. Cities have already shown a habit of underreporting their climate impact. AI firms may be tempted to repeat that pattern unless they are pushed toward more complete accounting.
Second, individual and household decisions about digital tools, streaming, and AI use stack up at scale. Guides that explain the role of consumers in reducing their carbon footprint are now extending that logic into digital life, not just physical purchases.
How Companies Are Trying To Adapt
Some firms see the writing on the wall and are adjusting. Others are trying to stretch existing claims as far as possible.
Major tech platforms are investing in more renewable contracts, better power usage effectiveness, and locating data centers in cooler climates. Work in science outlets such as Nature Climate and SPJ has mapped out how aligning AI projects with climate goals requires life cycle thinking.
Yet other financial analyses, including research flagged by Morgan Stanley, warn that generative AI might cause a threefold increase in greenhouse gas emissions from data centers alone by the end of the decade. This will happen if nothing serious changes in infrastructure and policy.
There is also a practical communication gap. Many people still do not realize how their daily apps connect to giant server farms. That gap gives room for very vague sustainability statements that sound comforting but do not answer basic questions about total energy, water use, and local impacts.
How You Can Stay Informed and Push For Better AI
You do not have to become a full-time energy analyst to play a part. But staying informed is easier if you plug into a few solid channels early.
You can track how media and researchers report on AI’s environmental costs through independent outlets. As users search search engines for these topics, the algorithms will prioritize climate news.
The more people ask tough questions about energy sources, local impacts, and reported numbers, the harder it becomes for any company to brush aside AI’s physical footprint. We must prioritize sustainable energy. Energy companies and tech giants alike must be held accountable.
