World Environment Day 2025: Is AI Sustainable? The Environmental Cost of Artificial Intelligence Models

From drafting your emails to generating music, artificial intelligence (AI) feels almost magical. With every passing month, tools like ChatGPT, Midjourney, and Gemini seem to get smarter. But as we reflect on the impact of technology this World Environment Day 2025, here's a less-talked-about truth-behind that slick chatbot reply or AI-generated image is a data center somewhere burning through electricity, guzzling water, and leaving a carbon footprint.

The question is simple but urgent: Is AI sustainable? And if it's not yet, can it ever be?

What Powers Artificial Intelligence?

AI might seem like software, but it's powered by vast networks of physical machines-think warehouses full of specialized chips working around the clock. These data centers are the brains behind AI models like GPT-4.5, Gemini 2.5, Llama 4, and more.

Training a large model like GPT-3 consumed over 1,200 megawatt-hours of electricity, roughly the same annual energy use as more than 100 average U.S. homes. According to a study from the University of Massachusetts Amherst, training such models can emit over 500 metric tons of CO₂, equivalent to the lifetime emissions of five cars.

And that's just to train the model. Once it's up and running, serving millions of users every day-a process called inference-consumes even more energy. OpenAI and others have acknowledged that inference is now a major contributor to the overall environmental footprint of large models.

In fact, MIT researchers caution that as generative AI becomes more embedded in our lives, inference could become the dominant source of electricity use. "Generative AI interfaces are so easy to use, but they obscure the enormous compute behind the scenes," notes Noman Bashir, a postdoc at MIT's CSAIL and co-author of a 2024 paper titled The Climate and Sustainability Implications of Generative AI.

The Thirst of AI: Cooling and Water Use

These chips get hot-really hot. To keep them from overheating, data centers rely heavily on cooling systems. Many of those use water. A report by The Associated Press revealed that Microsoft's data centers used over 11 million gallons of water in a single month while training GPT-4 in Iowa.

According to MIT's analysis, a data center can consume about two liters of water for every kilowatt-hour of electricity used-a staggering figure when you consider the massive energy footprints involved. "Just because this is called 'cloud computing' doesn't mean the hardware lives in the cloud," Bashir says. "It has direct, physical implications-especially for biodiversity and municipal water supplies."

The Carbon Cost of Intelligence

Every AI query draws electricity, and if that electricity comes from fossil fuels, it leaves behind a carbon trail. Research by Hugging Face and Climate Change AI showed that a single ChatGPT query can consume up to five times more energy than a traditional Google search.

A 2023 Nature commentary estimated that if generative AI becomes as ubiquitous as internet search, its carbon footprint could surpass 47 million metric tons of CO₂ annually, similar to the total yearly emissions of a small country like Portugal.

Meanwhile, Sam Altman, CEO of OpenAI, tweeted that simply exchanging pleasantries with ChatGPT adds "tens of millions of dollars" in electricity bills due to the backend processing involved, most of which still runs on fossil-fuel-powered grids.

Are Big Tech Companies Doing Anything?

Yes, and they're beginning to take sustainability more seriously:

• Google uses DeepMind AI to optimize cooling in its data centers, reportedly cutting cooling energy use by 40%. It also employs carbon-aware scheduling, shifting workloads to data centers in regions with abundant renewable energy.

• Microsoft has committed to becoming water positive and carbon negative by 2030. It's also developing water-free cooling systems to cut dependency on local water supplies.

• OpenAI, through its Azure partnership, benefits from Microsoft's green infrastructure, but it hasn't yet published detailed carbon accounting for GPT-4 or GPT-4o.

Yet MIT experts remain cautious. "The demand for new data centers cannot be met sustainably at the current pace," Bashir warns. With power densities seven to eight times higher than typical workloads, generative AI is accelerating energy demand faster than clean infrastructure can keep up.

Can AI Go Green?

Yes-and several promising strategies are already in motion:

• Smaller, efficient models: DistilBERT and TinyLLaMA deliver near-par results with a fraction of the energy and parameter count.

• On-device AI: Apple's Neural Engine and Google's Edge TPU bring AI directly to your device, reducing reliance on central servers.

• Federated learning: Used in Gboard and Apple keyboards, this approach distributes training across devices, avoiding massive central compute loads.

• Carbon-aware computing: MIT experiments show that shifting AI jobs to off-peak hours or cleaner energy regions can reduce emissions by up to 80%.

What You Can Do

While big changes must come from industry, users can still play a part:

• Choose lighter tools for simple tasks.

• Prefer apps that run AI locally.

• Support companies that disclose their environmental impact.

• Be mindful when using compute-heavy prompts or generating images/videos repeatedly.

Final Thoughts

AI is undoubtedly one of the most powerful tools humanity has ever built. But power comes at a price. From electricity consumption to water use and mining-intensive chip manufacturing, the true cost of generative AI is vast, and mostly hidden from view.

As Olivetti puts it, "We need a more contextual way of understanding the implications of generative AI. The technology is moving fast, and our ability to measure and respond must catch up."

AI doesn't have to be a zero-sum game. With more transparency, greener infrastructure, and thoughtful development, we can build systems that are not just intelligent, but sustainable too.