How Does AI Affect the Environment

How Does AI Affect the Environment? (The Full Picture in 2026)

Everyone is talking about what AI can do — write emails, generate images, diagnose diseases, write code. But very few people are asking what AI costs — not in money, but in water, energy, and carbon emissions.

The honest answer? A lot more than most people realize.

In this article, we break down exactly how AI affects the environment in 2026 — covering energy use, water consumption, carbon emissions, e-waste, and whether there's any hope for a greener AI future.

The Big Picture: AI's Environmental Footprint in 2026

AI doesn't run on magic. Behind every ChatGPT reply, every AI-generated image, and every recommendation algorithm is a massive data center packed with servers — and those servers are hungry. They need enormous amounts of electricity to run, and enormous amounts of water to stay cool.

Here are the numbers that put it all in perspective:

• AI data centers now generate 2.5–3.7% of global greenhouse gas emissions — officially surpassing the aviation industry's 2% contribution
• U.S. data centers consumed 17 billion gallons of water in 2023, with projections that this could reach 68 billion gallons by 2028
• A single ChatGPT query uses roughly 10x more electricity than a standard Google search
• Training GPT-4 alone consumed over 50 gigawatt-hours of electricity — enough to power San Francisco for three consecutive days
• AI workloads are growing 30% annually, compared to just 9% for conventional server workloads

These aren't scare tactics — they're measurements from MIT, Nature, Goldman Sachs, and the International Energy Agency. And they're getting bigger every year.

How Does AI Use Energy?

The Training Phase

Before any AI model can answer a single question, it has to be trained. Training means feeding the model billions of pieces of text, images, or data and letting it learn patterns — a process that can take weeks or months and requires thousands of specialized chips (GPUs) running simultaneously.

MIT researchers estimate that a generative AI training cluster consumes seven or eight times more energy than a typical computing workload. Training GPT-3 alone used approximately 1,287 megawatt-hours of electricity — enough to power 50 households for a full year.

And models are getting bigger every cycle. By 2025–2026, next-generation frontier models are expected to consume over 100 gigawatt-hours per training run.

The Inference Phase (Every Query You Make)

Training is a one-time cost, but inference — actually using the model — happens billions of times every day. ChatGPT alone handles around 200 million daily queries. Each query uses roughly 0.34 watt-hours of electricity. That adds up to around 564 megawatt-hours per day just for ChatGPT — more than many mid-sized data centers use in a day.

The inference phase is where AI's real, ongoing energy cost lives. And as more AI tools get embedded into search engines, smartphones, and apps, this number will only grow.

The Scale Problem

Global data center electricity consumption reached 460 terawatt-hours in 2022. The International Energy Agency projects this will grow to 945 terawatt-hours by 2030 — more than double. AI is the primary driver of that growth.

To put it in national terms: if data centers were a country, they would already rank among the top 15 electricity consumers in the world, sitting between Saudi Arabia and France.

How Does AI Use Water?

This is the part most people don't know about.

Data centers use massive amounts of water for cooling. The servers generate heat — a lot of it — and that heat has to go somewhere. The industry standard is chilled water cooling systems, where water absorbs the heat from the hardware and carries it away.

For every kilowatt-hour of energy a data center uses, it needs approximately 2 liters of water for cooling. Some estimates put it even higher — at 2.4 liters per kWh.

What Does That Mean in Real Terms?

• Training GPT-3 required approximately 700,000 liters of water
• Generating a 100-word email with GPT-4 uses roughly three 500ml bottles of water
• A single ChatGPT query uses about 500ml of water on average
• Microsoft's Iowa data center used 6 billion liters of water in just nine months, largely due to ChatGPT's cooling demands
• An average mid-sized data center now uses approximately 1.4 million liters (370,000 gallons) of water per day

Where Is This Water Coming From?

Here's where it gets serious. Many of the world's largest data centers are built in areas that are already experiencing water scarcity. Phoenix, Arizona — one of the driest cities in the United States — has seen a 32% projected increase in water stress due to data center growth alone.

Studies show that 30% of EU data center locations are in areas facing water stress. The data centers don't just use water — they compete with farms, households, and ecosystems for the same limited supply.

How Does AI Affect Climate Change?

Carbon Emissions

AI's carbon footprint comes from two main sources: the electricity powering the data centers, and the manufacturing of the hardware inside them.

When that electricity comes from fossil fuels — coal, natural gas — every query, every training run, every AI-assisted search adds CO2 to the atmosphere. A 2025 study estimated that AI systems had a carbon footprint equivalent to that of New York City in 2025 — between 32.6 and 79.7 million tons of CO2 equivalent.

AI data center emissions are now growing at 15% annually, and if that trajectory continues, they will surpass aviation's total footprint within this decade.

Reopening Coal Plants

One of the less-discussed consequences of AI's energy demand is what it's doing to the power grid. AI data centers are driving electricity demand so fast that utility companies can't keep up with renewables alone.

In Kansas City and West Virginia, coal-fired power plants that were scheduled to be shut down have had their closures delayed — kept alive specifically to meet the electricity demand from AI data centers. Similar situations have unfolded in Salt Lake City and other regions. AI is, in some places, actively extending the life of fossil fuel infrastructure.

The Short Shelf Life Problem

AI models have a surprisingly short shelf life. Companies release new models every few weeks, which means the energy used to train previous versions effectively goes to waste. Each new model is typically larger than its predecessor, requiring even more energy to train. It's a treadmill that never stops, and the energy cost keeps climbing.

AI's E-Waste Problem

The environmental impact of AI doesn't end with electricity and water. The hardware itself — the GPUs, servers, and cooling systems — creates a growing mountain of electronic waste.

• NVIDIA ships approximately 3.5 million GPUs per year for AI applications
• Each GPU generates around 5kg of e-waste at end of life
• AI demand has shortened the server refresh cycle from 5 years to just 3 years, increasing e-waste by 25%
• Annual AI hardware production now generates approximately 50,000 tons of e-waste globally
• Manufacturing these chips also requires rare earth minerals, whose extraction causes its own environmental damage

E-waste is already the fastest-growing waste stream in the world. AI is accelerating it.

Does AI Vary in Environmental Impact by Query Type?

Yes — and by quite a lot. Not all AI interactions are equal.

• Simple text queries (like asking a factual question) use as little as 0.002–0.007 Wh per prompt
• Text generation (writing an essay, drafting an email) uses around 0.05 Wh per prompt
• Image generation is the most intensive, averaging 2.91 Wh per prompt
• Video generation is in another category entirely — a single Sora 2 video reportedly burns 1 kilowatt-hour, uses 4 liters of water, and emits 466 grams of carbon
• Reasoning queries (philosophy, complex math, abstract questions) use significantly more energy than simple factual lookups

One surprising finding: a ChatGPT session run at 3 AM can be up to 67% more carbon-intensive than the same query at noon. This is because the electricity grid draws more heavily on fossil fuel backup power during off-peak hours when solar and wind output drops.

Can AI Help the Environment Too?

It's not all bad news. AI has genuine potential to reduce environmental damage in other sectors — and some researchers argue that these benefits could eventually outweigh the costs.

AI is being used to:

• Map renewable energy potential — identifying optimal locations for solar and wind farms
• Optimize power grid efficiency — reducing waste in electricity distribution
• Monitor deforestation and pollution from satellite imagery in real time
• Accelerate climate research — modeling climate systems faster than traditional computing
• Improve agriculture — reducing water and fertilizer use through precision farming
• Optimize logistics — cutting fuel use in shipping and transportation

The UN Environment Programme has noted that AI is becoming increasingly important in sustainability planning across energy, water, and agriculture. The question is whether these benefits will arrive fast enough and at a large enough scale to offset the emissions from AI's own infrastructure.

What Are Tech Companies Doing About It?

The biggest players have made pledges, though their track record is mixed.

• Microsoft pledged to go carbon negative and water positive by 2030. But since 2020, its total planet-warming emissions have risen around 30% — largely due to AI. Its water use increased by one-third in a single year.
• Google reported that a single Gemini text prompt uses about 0.24 Wh of energy and 0.26ml of water — far lower than many estimates for ChatGPT. But its AI operations still account for 15% of the company's total electricity use.
• OpenAI has not disclosed detailed emissions figures for ChatGPT, but CEO Sam Altman confirmed in 2025 that the average query uses about 0.34 Wh and 0.32ml of water.
• DeepSeek, a Chinese AI model, achieved 95% lower energy use compared to GPT-4-scale models through efficiency improvements — showing that smarter model design can drastically cut environmental impact.

International bodies are also pushing for change. The ISO has developed sustainable AI standards, and multiple governments are drafting regulations requiring companies to disclose the environmental footprint of their AI products.

The 30% Rule for AI Energy

You may have seen references to a "30% rule" in AI energy discussions. This refers to the finding that AI workloads are growing at approximately 30% per year — compared to just 9% for conventional server workloads. This compounding growth is what makes AI's energy trajectory so steep. At 30% annual growth, AI energy consumption doubles roughly every 2.5 years.

Summary: How Bad Is It, Really?

AI's environmental impact in 2026 is real, growing, and underreported. Here's a quick scorecard:

The story isn't that AI is evil. The story is that AI's environmental cost is invisible to most users — and that invisibility is a problem, because it means there's no pressure to use it responsibly.

FAQs: AI and the Environment

Q: Does AI use more energy than Bitcoin? 

AI is approaching and may soon surpass Bitcoin in energy use. Data centers, AI, and cryptocurrency together consumed 460 TWh in 2024. AI workloads are growing 30% annually, while Bitcoin's growth is slower.

Q: How much water does one ChatGPT query use? 

Approximately 500ml — the equivalent of a standard water bottle, according to multiple estimates. OpenAI's own figures put it slightly lower at around 0.32ml per query, though this likely reflects more efficient newer models.

Q: Is AI worse than flying for the environment? 

In terms of total sector emissions, AI data centers now generate 2.5–3.7% of global greenhouse gas emissions — already surpassing aviation's 2% contribution.

Q: Are AI or cows worse for the environment?

Global livestock (including cows) account for roughly 14.5% of global greenhouse gas emissions according to the FAO. AI data centers currently produce 2.5–3.7%. Cows are still significantly worse overall — but AI is growing much faster.

Q: Can I reduce my AI carbon footprint? 

Yes. Use text queries instead of image or video generation when possible. Avoid re-running the same query multiple times. Use AI tools powered by renewable energy where available. And simply being more selective about when you actually need AI vs. a regular Google search helps.