As we’ve seen throughout this book, artificial intelligence has the power to transform nearly every aspect of our lives. But like all powerful technologies, AI comes with significant environmental costs that we need to understand and address. In this section, we’ll explore the environmental footprint of AI systems—from the energy they consume to the water they use and the electronic waste they generate.
When you ask an AI chatbot a question or use an AI-powered feature on your phone, you’re tapping into a vast network of computers working behind the scenes. These aren’t the laptops or phones in your hands—they’re powerful servers located in massive facilities called data centers.
Data centers are the physical infrastructure that makes modern AI possible. These facilities house thousands of specialized computers that train and run AI models. Think of them as giant warehouses filled with computing power.
To put this in perspective, a large AI-focused data center can consume as much electricity as 100,000 households. The largest data centers under construction today may consume 20 times that amount.
Training is the process where an AI model learns from data. This is incredibly energy-intensive. For example, training GPT-4—one of the most powerful language models—required approximately 42.4 GWh of electricity. That’s enough to power about 28,500 households for a day.
Inference happens when you actually use a trained AI model. Each time you ask ChatGPT a question, the model performs millions of calculations to generate a response. While individual queries use less energy than training, the massive scale of users means inference can dominate total energy consumption over time.
You might be wondering: “If AI is so energy-hungry, why doesn’t efficiency keep pace?” The answer lies in what economists call the rebound effect—when improvements in efficiency actually lead to increased consumption. Here’s how it works:
This pattern is similar to what happened with cars: as engines became more fuel-efficient, people drove more, and total fuel consumption didn’t drop as much as expected.
The computers in data centers generate enormous amounts of heat. If not cooled properly, they would overheat and fail. Many data centers use water-based cooling systems that evaporate water to remove heat.
For a typical AI text prompt (like asking Gemini or ChatGPT a question), the water consumption is about 0.26 mL—roughly five drops of water. While this seems tiny, consider that billions of prompts are processed every day, adding up to significant water usage.
There’s also an indirect water cost: the water used to generate the electricity that powers data centers. Electricity generation—especially from coal, natural gas, and nuclear power—requires water for cooling. When you factor this in, the water footprint of AI grows even larger.
Subsection7.3.2.3Geographic Concentration and Water Stress
Data centers aren’t spread evenly across the world. They tend to cluster in specific regions—Northern Virginia, California, Texas in the United States; Dublin in Ireland; and Singapore in Asia. Many of these regions already face water stress.
About 20% of data center servers in the United States draw their water from moderately to highly stressed watersheds. This means AI infrastructure is competing with agriculture, municipal water supplies, and ecosystems for limited water resources.
The hardware that powers AI has a limited lifespan—typically just 3 to 5 years. As AI models grow more powerful, the equipment that runs them becomes obsolete quickly, generating a growing stream of electronic waste.
Electronic waste (e-waste) refers to discarded electronic devices and components. Between 2020 and 2030, it’s estimated that AI-related servers will generate approximately 16 million tons of e-waste. This is about 11% of all global e-waste during that period.
If properly recycled, the valuable materials in AI server e-waste could be worth approximately 70 billion dollars. If improperly disposed of, the toxic materials can contaminate soil and groundwater, damaging ecosystems and public health.
This means the e-waste problem is also concentrated. However, the supply chains that produce the hardware span the globe, with semiconductor manufacturing concentrated in East Asia and mineral extraction occurring worldwide.
Modern data centers have Power Usage Effectiveness (PUE) ratings as low as 1.1, meaning only 10% of energy goes to non-computing functions like cooling
Where data centers are built matters tremendously. A data center’s environmental impact can vary by 50 times or more depending on its location. Factors that matter:
Artificial intelligence offers tremendous potential benefits—advances in medicine, scientific discovery, climate modeling, and countless other areas. But these benefits come with real environmental costs.
As AI becomes increasingly integrated into our lives, understanding its environmental impacts isn’t just academic—it’s essential for making informed decisions about the kind of future we want to build.
A technology company announces plans to build a new AI-focused data center. They emphasize that the facility will use the most energy-efficient servers available. Which of the following statements best explains why this efficiency alone may not reduce the facility’s overall environmental impact?
Energy-efficient servers require more specialized manufacturing processes that produce higher carbon emissions during production, offsetting operational savings over the servers’ lifespan.
Incorrect. While manufacturing emissions are a real concern, this is not the primary reason efficiency gains fail to reduce overall impact.
The rebound effect means that as AI becomes more efficient and cheaper to run, more people and organizations will use it, potentially increasing total energy consumption.
Correct. The rebound effect occurs when efficiency improvements make AI cheaper to run, which encourages more usage, and more users mean more total energy consumption. This is similar to how fuel-efficient cars led to more driving, not less total fuel consumption.
Energy-efficient servers actually consume more water for cooling because they operate at higher densities, creating a trade-off between energy and water consumption.
Incorrect. Water consumption is a concern, but that does not mean that energy-efficient servers necessarily consume more water.
The energy savings from efficiency improvements are negligible compared to the energy required to train large AI models, which dominates total data center consumption.
Incorrect. While training is energy-intensive, the use of trained models can dominate total energy consumption over time due to the massive scale of users. Efficiency improvements apply to both phases.
