The Environmental Impact of Artificial Intelligence: Data Centers and Beyond

While most people think about AI in terms of the apps on their phones or the chatbots they interact with, the real environmental story happens behind the scenes—in massive facilities called data centers. These are the physical buildings filled with thousands of computers that process every AI request. Data centers come in many shapes and sizes, from small server rooms in office buildings to warehouse-sized facilities operated by tech giants like Google, Microsoft, and Amazon.

But the newest generation of data centers, specifically built to handle AI workloads, are fundamentally different from their predecessors. They require far more powerful processors called GPUs (Graphics Processing Units), consume dramatically more electricity, and need significantly more water for cooling. A single AI-focused data center can use as much electricity as a small city and as much water as a large neighborhood.

Generative Artificial Intelligence (GenAI) refers to tools and systems capable of generating text, images, and video. While its use perplexes many people, others react unimpressed by its day-to-day applications or fear its capacity to displace jobs. A recent Pew Research Center report reflects the public and experts’ mixed sentiments, showing that only 17% of people in the United States think AI will have a positive impact within the next 20 years, compared to 56% of experts. In addition, 51% say they are more concerned than excited compared to 17% of experts. One area where both groups agree on is the need for more control and regulation of AI.

One growing conversation in the use or non-use of GenAI is the environmental implications of the technology in every step of its life cycle. Recently, Sam Altman, CEO of OpenAI, suggested that AI’s benefits outweigh its costs. These statements overlook who bears these health and environmental costs, and who gets to decide what trade-offs are acceptable. The communities breathing polluted air near data centers or those experiencing increased energy costs are the ones at the frontline of the technology’s impacts and often last its benefits.

Sam Altman ends by saying, “Intelligence too cheap to meter is well within grasp.” Implying that AI will become incredibly cheap to produce, so cheap that the cost becomes negligible. ‘Too cheap to meter’ is a quote from 1954 during the early development of nuclear energy, when similar promises were made about electricity costs that ultimately didn’t materialize, but many of its impacts did. Also, didn’t ChatGPT start free, then move to $20 per month for advanced features, with enterprise plans costing hundreds or thousands of dollars? So when Altman talks about intelligence becoming ‘too cheap to meter,’ we have to ask: cheap for whom?

To understand the multiple environmental implications, let’s journey backward through the processes from when a user inputs a prompt into a GenAI tool to when they receive a response in the form of text, an image, or a video. A user is on their computer and writes a prompt into ChatGPT and clicks send. That request goes to OpenAI servers (the company that created ChatGPT). Servers are located in Data Centers; there are multiple companies that provide data center services, and some of the bigger tech companies like Meta, Google, Amazon, have their own data centers. Data centers are physical buildings with racks of processing units or computers. These units only process the data you input from your device (e.g., personal computer or phone). So, for practical purposes, imagine computers stacked on each other that don’t require a monitor, keyboard, or mouse.

The processing of the prompt happens in Graphics Processing Units (GPUs) and/or Central Processing Units (CPUs). The difference is that GPUs have higher processing power and are more energy-intensive. Most of GenAI requests happen on GPUs regardless of whether your prompt is asking for a text, image, or video response. They acquired their name because they were originally developed to process graphics for video games and other 3D graphics applications. Today they are used for parallel processing, which allows performing multiple calculations simultaneously. The more “work” a GPU is doing, the more energy it requires. In general, it will take more energy to produce an image, or many images (videos), than text.

To put this in perspective, calculations by O’Donnell and Crownhart in a MIT Technology Review report show that a single query to a small AI text model uses about 114 joules, roughly equivalent to running a microwave for one-tenth of a second. However, larger, more powerful models can use 6,706 joules per response, enough energy to run that same microwave for eight seconds or carry a person 400 feet on an e-bike, according to the report.

The same report estimates that generating a standard-quality image requires about 2,282 joules, while creating a high-quality five-second video can consume over 3.4 million joules—more than 700 times the energy of generating a high-quality image, equivalent to riding 38 miles on an e-bike or running a microwave for over an hour.

Of course, looking at an individual prompt may not seem like much without the context of how many prompts servers get in a day. Estimates are that ChatGPT receives over a billion requests per day for generating text and tens of millions for generating images. According to the same article published in MIT Technology Review, the electricity to process those prompts in one day is equivalent to the power used by over 3,000 homes for a year.

These calculations do not include the energy to generate video and do not include the prompts that other large companies receive through their own AI models, like Microsoft Copilot, Google Gemini, X’s Grok, and other companies developing other AI tools and models.

While many of these calculations have limitations and assumptions that bring uncertainty, what is certain is the observed increase in the electricity that data centers are already using. In 2018, data centers were using 1.9% (76 TWh) of total United States electricity consumed. In 2023, it increased to 4.4% (176 TWh) of the total U.S. electricity consumption and projections to 2028 fall within 6.7% to 12% (300+ TWh to 500+ TWh). All these estimates come from the 2024 United States Data Center Energy Usage Report.

As GPUs are not perfectly efficient, a substantial portion of the energy is turned into heat. That’s where part of the water consumption comes in, to keep those servers cool.

Similarly to your phone or computer, the servers processing AI tool requests get hot and release heat into the room. Hot environments can damage electronic components or reduce their efficiency. To cool the servers and data center facilities, there are multiple methods. Air conditioning systems require massive amounts of electricity but use little water, while water-based cooling methods are often preferred because they are cheaper. The cooling happens through machines known as Computer Room Air Handlers (CRAH). In short, these machines take the hot air that rises inside the room, cool it, and return it to the bottom of the room.

Inside the machines there are coils with chilled water. The hot air transfers heat to the water, cooling the air in the process. The outcome is hotter water that needs to be chilled again. The hot water goes to cooling towers where part of the water is evaporated–that’s where the high water use happens. When water evaporates, it absorbs energy in the form of heat, decreasing the temperature of the remaining water, that becomes cooler. Freshwater needs to replace the water that was evaporated. According to a recent report by Lawrence Berkeley National Laboratory, U.S. data centers consumed 66 billion liters of water directly at their facilities (also known as direct water) in 2023.

While 66 billion liters sounds significant—and it is—whether this represents a large amount depends on what we compare it to. If you compare it with water use for agriculture, for example, then it would arguably not be too much water. 66 billion liters convert to 53,507 acre-feet, a measure of water use in agriculture and you can think of each acre-foot as the area of one football field filled with one foot of water. I live in California’s Central Valley, where I can compare this to the water use of almond trees (5 acre-feet per acre of almonds per year in the southern part of the Valley). That would mean that with the water used by all data centers in the United States you could irrigate 11,101 acres of almonds. In California there are about 1.56 million acres of almonds. If we compare it to residential water use, then 66 billion liters is equivalent to the water use of over half a million people in one year.