The majority of people think of Artificial Intelligence (AI) as part of the applications on their phones or chatbots, but the real story of environmental implications occurs behind the scenes, in massive facilities called “data centers.” These are buildings filled with dozens, hundreds, or even thousands of computers that process each 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 new generation of data centers, specifically built to handle AI services, are fundamentally different from their predecessors. They require much more powerful processors called GPUs (Graphics Processing Units) that consume a lot more electricity and require significantly more 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 (Generative AI) refers to tools and systems capable of generating text, images, and/or videos. While its use perplexes many people, others react unimpressed by its real-world applications or fear its potential to displace jobs.
A recent report from the Pew Research Center reflects mixed feelings from the public and experts, showing that only 17% of people in the United States think AI will have a positive impact in the next 20 years, compared to 56% of AI experts. Additionally, 51% say they are more concerned than excited compared to 17% of experts. An area where both groups agree is the need for greater control and regulation of AI.
An increasing concern about the use of Generative AI is the environmental implications of the technology at every step of its life cycle. Recently, Sam Altman, CEO of OpenAI, suggested that the benefits of AI outweigh its costs. These statements do not consider who bears these costs to health and the environment. Nor do they consider who can decide which compromises are acceptable. Communities breathing polluted air near data centers or experiencing increases in energy costs are on the front lines of the impacts of the technology and far from its benefits.
Sam Altman ends by saying, “The intelligence too cheap to measure is within our reach.” Referring to the fact that AI will become incredibly cheap to produce, so cheap that the cost becomes insignificant. ‘Too cheap to measure’ is a phrase from 1954 during the early stages of nuclear energy development, when similar promises were made about electricity costs that ultimately were not fulfilled, but many of its impacts did.
To understand the multiple environmental implications, let’s travel back through the processes from when a user enters a message into a GenAI tool until they receive a response in the form of text, image, or video. Through this post, I describe what I have learned about this topic, from the most well-known implications of artificial intelligence in terms of energy to other lesser-known impacts regarding water usage and air pollution.
### The Energy Cost of Asking ChatGPT for Something
A user is on their computer and types a message into ChatGPT and clicks send. That request goes to OpenAI’s servers (the company that created ChatGPT). The servers are located in data centers; there are several companies that provide data center services, and some of the largest tech companies like Meta, Google, Amazon, have their own data centers. Data centers are buildings with racks of processing units or computers. These units only process the data input from your device (like your laptop or phone). For practical purposes, imagine computers stacked on top of each other that do not require a monitor, keyboard, or mouse.
The processing of the request is done on Graphics Processing Units (GPUs) and/or Central Processing Units (CPUs). The difference is that GPUs have more processing power and consume more energy. Most GenAI requests are made on GPUs, regardless of whether the message requests a text, image, or video response. They were named because they were originally developed to process graphics for video games and other 3D graphics applications. They are now used for parallel processing, allowing multiple calculations to be done simultaneously. The more “work” a GPU does, the more energy it requires. In general, more energy will be needed to produce an image, or many images (videos), than text.
To put this into 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 around 114 joules, approximately equivalent to running a microwave for a tenth of a second. However, larger and more powerful models can use 6706 joules per response, enough energy to run that same microwave for eight seconds or take a person 120 meters on an electric bike, according to the report.
The same report estimates that generating a standard quality image requires around 2282 joules, while creating a five-second high-quality video can consume over 3.4 million joules, over 700 times the energy of generating a high-quality image, equivalent to cycling 61 kilometers on an electric bike or running a microwave for over an hour.
Considering individual messages may not seem like much without the context of how many requests the servers receive in a day. It is estimated that ChatGPT receives over a billion requests a day to generate text and tens of millions to generate images. According to the same article published in MIT Technology Review, the electricity to process those requests is equivalent to the energy used by over 3000 households in a year.
These calculations do not include the energy to generate video and do not include the requests that other large companies receive through their own AI models, such as 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 reduce their accuracy, one undeniable fact is the increase that has already been observed in the electricity data centers use. In 2018, data centers used 1.9% (76 TWh) of the total electricity consumed in the United States. In 2023, it increased to 4.4% (176 TWh) of the total electricity consumption in the U.S., and projections until 2028 range from 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 amount of energy is converted into heat. This is where part of the water consumption comes in, to prevent the servers from overheating.
### Water Usage in Data Centers
Similar to your phone or computer, servers processing AI requests heat up and release heat into the room. Warm environments can damage electronic components or reduce their efficiency. There are several ways to cool servers and data center facilities, such as air conditioning systems that require large amounts of electricity but use little water, or water-based cooling that is often preferred because it is cheaper. Cooling is done through machines known as Computer Room Air Handlers (CRAH). In short, these machines take the warm air rising inside the room, cool it, and return it to the bottom of the room.
### How Does This Work?
Inside the machines are coils with cold water. The warm air transfers heat to the water, cooling the air in the process. The result is warmer water that needs to be cooled again. The warm water goes to cooling towers where some of the water evaporates, that is where the high water usage occurs. When the water evaporates, it absorbs energy in the form of heat, lowering the temperature of the remaining water, which cools. The evaporated water needs to be replaced. According to a recent report from the Lawrence Berkeley National Laboratory, U.S. data centers consumed 66 million cubic meters of water directly in their facilities (also known as direct water) in 2023.
If 66 million cubic meters is a significant amount depends on what we compare it to. If we compare it to the amount of water used for agriculture, for example, then it might be said to not be too much water. 66 million cubic meters convert to 53 thousand acre-feet, a measure of water usage in U.S. agriculture, and one acre-foot can be thought of as the area of a football field filled with a foot (30cm) of water. I live in the Central Valley of California, where I can compare this to the water usage 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 U.S. data centers, 11 thousand acres (4492 ha) of almonds could be irrigated. In California, there are around 1.56 million acres of almonds. If we compare it to residential water usage, then 66 million cubic meters equate to the water usage of over half a million people in a year in the United States.
As an expert in water with a Ph.D. in hydrology and having worked on water and climate change for over 10 years, I have other concerns. One of them is that it is showing an exponential growth trend. According to the same report from the Lawrence Berkeley National Laboratory, direct water used by data centers in the U.S. in 2014 was 21 million cubic meters. This means that over the course of 10 years, water usage tripled and data centers are growing in number and size.
The other major concern is that those numbers only reflect the direct water use by data centers. It does not include other water uses in the life cycle of artificial intelligence, such as water use for resource extraction and manufacturing for the building itself and the hardware and microchips that form the processing units. It also does not take into account the indirect water use by power plants that produce electricity for data centers, estimated at 800 million cubic meters in 2023, over 10 times the direct water usage.
Furthermore, these comparisons do not tell the whole story because half of the data centers are located in regions where water is already scarce. Data centers are strategically positioned near population centers and in areas with lower electricity costs, but this also means that they often compete with local communities and agriculture for limited water resources, particularly in drought-prone regions like California, Arizona, and parts of Texas.
### Data Centers Also Contribute to Air Pollution and Health Degradation
The energy consumed by data centers not only contributes to carbon emissions but also generates air pollution that directly harms human health. Throughout the life cycle of AI, from chip manufacturing to data center operation, significant amounts of criteria air pollutants, such as fine particles (PM2.5), nitrogen dioxide (NO2), and sulfur dioxide (SO2), are released into the atmosphere.
A recent article by Han et al. at UC Riverside and Caltech titled “The Unpaid Toll: Quantifying the Public Health Impact of AI” quantifies the public health impacts throughout the AI life cycle. They estimate that the public health burden of U.S. data centers in 2030 will be valued at over $20 billion per year, comparable to the emissions from road vehicles in California. The cost comes from increased cases of asthma and other cardiopulmonary diseases caused by poor air quality.
These pollutants come from three main sources. Firstly, data centers rely on backup generators that use diesel and emit substantial amounts of air pollutants during operation, testing, and maintenance. Secondly, the electricity powering data centers often comes from fossil fuel power plants that release air pollutants by burning coal and natural gas. Thirdly, the manufacturing of AI hardware (which requires highly refined materials) and materials to build data center buildings (iron for steel and resources for cement) generate contaminant emissions.
As in other cases of environmental justice, these environmental impacts are not distributed equitably. Air pollutants can travel hundreds of kilometers from their sources, but the most affected communities are often low-income areas that receive few economic benefits from data centers.
### How to Learn More?
Understanding the environmental implications of AI is crucial to be more aware of its use and to ask representatives to design and implement appropriate policies and regulations. We need policies that require community involvement from the planning stages, that demand environmental protection, that encourage transparent and accessible information from technology developers, and that ensure the benefits of technological advancement are shared equitably.
I had the pleasure of collaborating with a group of exceptionally intelligent, attentive, and talented students from Computer Science and Mathematics majors at Harvey Mudd College and Scripps College. Our collaboration culminated in an educational website about Artificial Intelligence and its implications for energy and water. You can explore the interactive elements and share them with others who may be interested in learning more about the environmental costs of artificial intelligence.
