AI Devours Energy: 80 Million Tons of CO2 in 2025

You use ChatGPT, Claude, or Gemini every day without a second thought. But every prompt you send consumes the equivalent of a water bottle and generates CO2 at a scale that now exceeds the emissions of entire countries. Spoiler alert: the 2025 figures are frankly alarming, and nobody is really talking about it.

1. The alarming numbers

Honestly, I was floored when compiling this data. A study by VU Amsterdam Institute for Environmental Studies published in Patterns estimates that AI systems will produce between 32.6 and 79.7 million tons of CO2 in 2025. The comparison the study itself draws: the top of that range is in the same range as the carbon footprint of New York City (52.2 Mt in 2023).

Source: de Vries-Gao A., "The carbon and water footprints of data centers and what this could mean for artificial intelligence," Patterns, 2026;7(1):101430 — doi.org/10.1016/j.patter.2025.101430, indexed on PubMed (PMID 41583976).

Another way to picture it: that upper figure also exceeds the annual emissions of Norway (31.5 Mt) and approaches those of Singapore (53 Mt). We are talking about a technology that most people consider "immaterial."

The thing is, it is only accelerating. The projections for 2030 are even more alarming:

To put the 23 gigawatts by end of 2025 into context: the study describes it as approaching the average electricity demand of the United Kingdom, which stood at 30.7 GW on average in 2023. An entire country.

2. The invisible water: AI's hidden thirst

Carbon emissions are already serious. But wait until you see the water figures.

Per the same study in Patterns, AI could consume between 312.5 and 764.6 billion liters of water in 2025. That volume is in the same range as global annual bottled water consumption, which the study puts at 446 billion liters. Yes, you read that right.

And this is where it becomes concrete for you: according to the UC Riverside team that measured this cost, GPT-3 consumes about 500 milliliters of water per 10 to 50 medium-length responses — so not per query, contrary to what you read everywhere, and the figure varies with where and when the model runs. Multiply that by the hundreds of millions of daily users anyway, and you get entire lakes evaporating to run language models.

Some figures that struck me:

• Training GPT-3 evaporated 700,000 liters of clean freshwater on site — and 5.4 million liters in total, once the water consumed to generate its electricity is counted
• Microsoft saw its water consumption increase by 34% in 2022 (6.4 billion liters), mainly due to its AI investments with OpenAI
• Nearly one-third of new data centers are being built in areas that will experience acute water shortages by 2050

Source for the GPT-3 figures: Li P., Yang J., Islam M. A., Ren S., "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models," Communications of the ACM, 2025;68(7):54-61 — doi.org/10.1145/3724499 (open-access preprint: arXiv:2304.03271).

3. Why AI electricity is so polluting

We tend to see digital technology as something clean, dematerialized. This is a complete illusion.

A generative AI search consumes about 30 times more electricity than a simple Google search. And generative AI is expected to represent at least 50% of data center electricity consumption - an acceleration of 5 to 10 times compared to 2022.

But the worst part is where these data centers are built. A study by Harvard and UCLA covering 2,132 US data centers operating between September 2023 and August 2024 reveals that about 95% of them sit in areas where electricity is more carbon-intensive than the national average.

A second figure from that study is often conflated with the first: the average carbon intensity of the electricity these data centers consume reaches 548 gCO2e/kWh, or 48% above the US average (369 gCO2e/kWh). That 48% is a consumption-weighted average across the whole fleet — it is not how much dirtier the 95% mentioned above are.

Source: Guidi G., Dominici F., Gilmour J. et al., "Environmental Burden of United States Data Centers in the Artificial Intelligence Era," arXiv preprint, 2024 — doi.org/10.48550/arXiv.2411.09786. Work from the Harvard and UCLA schools of public health, not yet peer-reviewed.

This is no coincidence. Tech giants seek the cheapest electricity - and often, it is also the dirtiest.

And let us not forget GPU manufacturing. The extraction of necessary rare earths generates systematic ecosystem destruction. In China, areas near extraction sites show significantly reduced biodiversity levels and much slower ecological recovery rates.

4. Pros and cons of the AI race

5. Emerging solutions (finally)

The good news? Serious solutions are beginning to emerge.

On the algorithm side

The Zeus framework from the University of Michigan has demonstrated a reduction of up to 75% in energy consumption for training. Compression techniques (quantization, pruning) can reduce model size without sacrificing accuracy. DistilBERT uses 40% fewer parameters than BERT with nearly identical precision.

On the infrastructure side

Microsoft has reported an 80% improvement in Water Usage Effectiveness through closed-loop cooling systems. Immersion cooling (servers submerged in non-conductive fluids) reduces cooling needs by over 90%.

On the energy side

Sam Altman has invested $20 million in Exowatt to power OpenAI's data centers with solar. Google is targeting 100% carbon-free electricity by 2030. Amazon has already matched 100% of its consumption with renewable electricity.

On the regulatory side

The EU has incorporated mandatory environmental impact assessments into its AI Act. In the United States, 7 states have legislated that data centers must pay for grid expansion costs. Maine has imposed a limit: companies cannot dedicate more than 25% of their electrical capacity to data centers.

Frequently Asked Questions

How much energy does a ChatGPT query consume?

A generative AI search consumes about 30 times more electricity than a standard Google search. On water, the UC Riverside team measured that a model like GPT-3 consumes roughly 500ml per 10 to 50 medium-length responses, not per query (doi.org/10.1145/3724499).

Can AI become environmentally friendly?

Technically yes. New chips and optimized algorithms enable efficiency gains of 50 to 75%. However, the rebound effect (more efficiency = more usage) risks canceling out these gains without strict regulation.

Which AI companies are the most transparent?

Google and recently OpenAI have started publishing their energy consumption data. Meta, on the other hand, does not disclose its AI-related emissions.

What can I do at my level?

Use lighter models when sufficient, limit unnecessary queries, and support regulatory initiatives. Public pressure matters.

Conclusion

AI stands at a crossroads. Its current energy architecture is fundamentally unsustainable: up to 80 million tons of CO2 per year, hundreds of billions of liters of water, mining extraction that devastates ecosystems.

Three levers are essential: mandatory emissions transparency (like existing ESG standards), massive investments in optimization and renewables, and binding regulations before AI consumes unsustainable shares of our resources.

A technology built on the systematic destruction of the environment cannot be the architecture of the future. And to be honest, if we do not act now, technological solutions - even revolutionary ones - will no longer be enough.

---

About the author: The EverydayAITech team has been testing and analyzing artificial intelligence tools since 2023. Our mission: democratizing AI by offering practical and honest guides, without unnecessary technical jargon.