On the surface, our digital lives feel almost weightless. Every question asked, image uploaded, or movie queued up flickers across a screen and disappears just as quickly. It’s easy to forget that behind each AI tool lies a physical footprint. Somewhere, far from the glow of your device, something hums, something cools, something burns energy so that your online life can feel effortless.
Artificial Intelligence (AI), especially generative AI models like ChatGPT or Gemini, run on sprawling networks of computers called data centers—warehouses packed with servers that run 24/7, storing and processing mountains of data. All this computing takes an enormous amount of power. According to researchers at MIT, training a single large language model can emit as much carbon dioxide as the equivalent of 300 round-trip flights between New York and San Francisco, or nearly 5 times the lifetime emissions of the average car [1][3]. Once the model is trained, the usage doesn’t just disappear; every user interaction spins the system back up again.
On top of electricity use, AI demands enormous amounts of water. Data centers often rely on water-based cooling systems that use millions of gallons daily to prevent servers from overheating. One study estimated that training GPT-3 used roughly 700,000 liters of clean water, mostly for cooling [2]. In some areas, like Arizona and Iowa, where many major AI data centers are located, this has raised concerns about water scarcity [7].
The electricity that powers AI mostly comes from fossil fuels, meaning that more AI activity often equals more greenhouse gas emissions. The U.S. Government Accountability Office (GAO) warns that the energy consumption of AI data centers could more than double by 2030 if trends continue [5]. Meanwhile, the International Energy Agency estimates that data centers and AI combined could soon use as much electricity as the entirety of Japan [8]. Even when these facilities tout “renewable energy commitments”, most still depend on local grids that rely on coal or natural gas during peak hours [6].
Hardware is another piece of the puzzle. Training and running AI requires specialized chips called GPUs, which are built using rare metals and complex manufacturing. Producing these chips generates emissions long before they even reach a data center [4]. As demand for AI accelerates, so does the demand for new chips, servers, and network hardware, which all contribute to electronic waste when outdated parts are thrown away [3].
But the story isn’t entirely negative. Some researchers argue that AI can actually help fight climate change by improving renewable energy systems and making manufacturing more efficient [8]. For example, AI models can optimize power grids, predict energy demand, and even monitor deforestation using satellite images [12]. That tension—AI as both a tool for progress and a driver of harm—is one of the central contradictions of our time.
That’s where the concept of “green AI” comes in. Green AI focuses on designing models that use fewer resources without sacrificing performance. MIT researchers say this involves making models smaller and more efficient; a shift from “bigger is better” to “smarter is cleaner” [1]. Organizations like the United Nations Environment Programme are calling for transparency from tech companies, asking them to report exactly how much energy and water their AI systems consume [2][9]. Without open data, it’s nearly impossible to measure whether AI is becoming more sustainable or simply scaling faster than we can manage.
Solutions do exist. Some are technical: redesign chips to use less power, build data centers in cooler climates, switch to on-site renewable energy, or reuse server heat to warm nearby buildings [10][11]. Others are policy-driven: carbon labeling for AI products or stricter reporting requirements [13].
The question, then, is not whether we should keep developing AI—it will. It’s whether we can do it responsibly. Just as rare earth elements power green technologies but damage the planet when mined, AI carries its own trade-offs. The challenge is balancing the incredible potential of AI with its hidden environmental costs; making sure our smartest technology doesn’t make the Earth any dumber.
[4]https://www.nea.org/professional-excellence/student-engagement/tools-tips/environmental-impact-ai
