It’s strenuous to ignore the excitement about the potential benefits of generative AI, from improving employee productivity to advancing scientific research. While the rapid development of this fresh technology has enabled the rapid deployment of powerful models across many industries, the environmental consequences of this generative AI gold rush remain challenging to quantify, let alone mitigate.
The computing power required to train generative AI models, which often have billions of parameters, such as OpenAI’s GPT-4, can require a staggering amount of electricity, leading to increased carbon emissions and pressure on the electrical grid.
Moreover, deploying these models in real-world applications, enabling millions of people to apply generative AI in their everyday lives, and then tuning the models to improve their performance consumes immense amounts of energy long after the model is developed.
In addition to the demand for electricity, a lot of water is needed to cold the equipment used to train, deploy and tune generative AI models, which can strain municipal water supplies and disrupt local ecosystems. The growing number of generative applications of artificial intelligence has also stimulated demand for high-performance computing equipment, increasing the indirect environmental impact of its production and transportation.
“When we think about the environmental impact of generative AI, we don’t just think about the electricity consumed when a computer is connected. There are much broader consequences that reach down to the system level and persist depending on the actions we take,” says Elsa A. Olivetti, professor in the Department of Materials Science and Engineering and leader of MIT’s new lab decarbonization mission Climate project.
Olivetti is senior author of a paper published in 2024.The implications of generative artificial intelligence for climate and sustainability,” co-authored by MIT colleagues in response to an institute-wide call for papers exploring the transformative potential of generative AI, in both positive and negative directions for society.
Demanding data centers
The electricity demand of data centers is one of the main factors influencing the environmental impact of generative AI, as data centers are used to train and run deep learning models behind popular tools such as ChatGPT and DALL-E.
A data center is a temperature-controlled building that houses computing infrastructure such as servers, data storage drives, and networking equipment. For example, Amazon has more of them 100 data centers around the worldeach of which has approximately 50,000 servers that the company uses to run cloud services.
Although data centers have existed since the 1940s (the first were built at the University of Pennsylvania in 1945 to support the first general-purpose digital computerENIAC), the development of generative artificial intelligence has dramatically increased the pace of data center construction.
“What sets generative AI apart is the power density it requires. It’s essentially just computation, but a generative AI training cluster can use seven to eight times more energy than a typical computational workload,” says Noman Bashir, lead author of the Impact paper, who is a computational and climate impact specialist at the MIT Climate and Sustainability Consortium (MCSC) and postdoc at the Computer Science and Artificial Intelligence Laboratory (CSAIL).
Researchers estimated that data center energy demand in North America increased from 2,688 megawatts at the end of 2022 to 5,341 megawatts at the end of 2023, driven in part by demand for generative artificial intelligence. In 2022, global electricity consumption by data centers increased to 460 terawatts. This would make data centers the 11th largest consumer of electricity in the world, between Saudi Arabia (371 terawatts) and France (463 terawatts), according to a study by the Organization for Economic Co-operation and Development.
Electricity consumption in data centers is expected to reach 1,050 terawatts by 2026 (taking data centers to fifth place on the global list, between Japan and Russia).
While not all data center computing uses generative AI, the technology is a major driver of energy demand growth.
“The demand for new data centers cannot be met sustainably. The pace at which companies are building new data centers means that most of the electricity to power them must come from fossil fuel power plants,” says Bashir.
It’s difficult to quantify the power needed to train and deploy a model like OpenAI’s GPT-3. In a 2021 research paper, researchers from Google and the University of California, Berkeley estimated that the training process alone consumed 1,287 megawatt hours of electricity (enough to power about 120 average U.S. homes for a year), generating about 552 tons of carbon dioxide.
While all machine learning models require training, a hallmark of generative AI is the rapid fluctuations in energy consumption that occur at different stages of the learning process, Bashir explains.
Power grid operators must have a way to absorb these fluctuations to protect the grid, and they tend to do so diesel-based generators for this task.
The growing influence of inference
Once a generative AI model is trained, the energy demand does not disappear.
Each time the model is used, for example by someone asking ChatGPT to summarize an email message, the computer hardware performing these operations consumes power. Researchers estimated that a ChatGPT query uses about five times more energy than a regular Internet search.
“But the average user doesn’t think about it much,” says Bashir. “The ease of use of generative AI interfaces and the lack of information about the environmental impact of my actions means that as a user I have little incentive to limit my use of generative AI.”
In established AI, energy consumption is distributed fairly evenly between data processing, model training, and inference, the process of using a trained model to predict fresh data. However, Bashir expects that the electricity demand for AI generative inference will eventually dominate as these models become ubiquitous in so many applications, and the electricity needed for inference will escalate as future versions of the models become larger and more elaborate.
Additionally, generative AI models have a particularly low shelf life due to the growing demand for fresh AI applications. Companies release fresh models every few weeks, so the energy spent training previous versions is wasted, adds Bashir. Novel models often apply more energy for training because they usually have more parameters than their predecessors.
Although most of the research literature has focused on the electricity demand of data centers, the amount of water used by these facilities also has environmental impacts.
Chilled water is used to cold a data center by absorbing heat from computer equipment. It’s estimated that for every kilowatt-hour of energy used by a data center, two liters of water are needed for cooling, Bashir says.
“Just because it’s called ‘cloud computing’ doesn’t mean the hardware lives in the cloud. “Data centers are present in our physical world and, due to their water consumption, they have a direct and indirect impact on biodiversity,” he says.
Computer equipment in data centers has its own, less direct impact on the environment.
While it’s challenging to estimate how much energy would be needed to produce a GPU, the type of powerful processor that can handle the intensive workloads of generative AI, it would be more than enough to produce a simpler processor because the manufacturing process is more elaborate. On top of the GPU’s carbon footprint are emissions from the transportation of materials and products.
The sourcing of raw materials used to produce GPUs also has environmental implications, which can involve grubby mining procedures and the apply of toxic chemicals in processing.
Research firm TechInsights estimates that the three major manufacturers (NVIDIA, AMD, and Intel) will ship 3.85 million GPUs to data centers in 2023, up from about 2.67 million in 2022. This is expected in 2024 .this number will escalate even more. .
The industry is on an unsustainable path, but there are ways to encourage responsible development of generative AI that supports environmental goals, Bashir says.
He, Olivetti and their colleagues at MIT argue that this will require a comprehensive consideration of all the environmental and social costs of generative AI, as well as a detailed assessment of the value and perceived benefits.
“We need a more contextual way to systematically and comprehensively understand the implications of new developments in this space. Because of the speed at which improvements were being made, we didn’t have a chance to catch up on how to measure and understand the trade-offs,” says Olivetti.