Photo courtesy of Michael Shainblum/Stills
When new machines and systems like ChatGPT emerge—with the promise of making our lives easier and our work more efficient—it can be easy to accept a sort of technological determinism. But culture shapes technology, not the other way around: The questions we ask now, in nascent stages of development, have the power to shape its future.
One of our primary curiosities around AI is how it affects the planet. The environmental impact of AI is not well-documented—the industry is unregulated, which means the companies that are creating AI models are not required to track or disclose this information. But scientists like Shaolei Ren, PhD, a computer engineering professor who researches AI and sustainability, are beginning to uncover its impact. When I chatted with Ren, he shared his latest estimates.
A Q&A with Shaolei Ren, PhD
WHAT IS AI?
AI, or artificial intelligence, is a very broad term that describes a machine learning model—a model that’s taught, by being given lots of data, how to generate responses to requests (like a human’s question) that it has never seen before.
An AI model is created primarily using two phases: training and inference. Training, as its name suggests, is the process in which a lot of existing data is used to teach the server what type of information you want it to generate and how—you’re basically trying to optimize the model. This usually takes about two weeks.
The inference phase is when the AI model is actively being used for data requests. This phase lasts for the duration of time the AI model is used.
WHAT IS THE IMPACT OF AI MODELS ON THE ENVIRONMENT?
When we talk about the environmental impact of AI models, the training phase is typically the most environmentally draining (even though it lasts for only a couple of weeks) because thousands and thousands of servers and GPUs (graphics processing units) are needed to train the model. This is the case regardless of whether the AI model will have millions of users or just a few hundred, because similar amounts of resources are needed to train all of them.
While the model is being used—in the inference phase—its environmental impact will depend on the number of users: The more users, the greater the effect.
Here are what researchers estimate for the energy, carbon, water, and land consumption of AI models—mostly from the training phase, but some assessments were made in the inference phase, when the data was available.
Energy
Training a large AI model—GPT-3 or GPT-4, used in ChatGPT, for example—requires more than 1,000 megawatt hours of energy. For context, each US household uses about one megawatt hour of energy per month. That means that training a large AI model like ChatGPT, which takes around two weeks, uses the same amount of energy that 100 households in the US use in a year.
We estimate that when using a large AI model, asking one question is the equivalent of turning on an LED bulb for about half an hour (or 0.003 to 0.004 kilowatt hours of energy). When multiplied by millions of users, that’s a lot of energy.
Carbon
The carbon footprint for training a large AI model (which can be more complicated to track) involves two types of carbon: the operational carbon, which is associated with the electricity usage, and the embodied carbon—the carbon footprint used to make the AI chips and servers. The exact numbers significantly vary depending on many factors, including the algorithms, datasets, and hardware that are used. But the combined carbon footprint for training one large AI model is often equivalent to a few hundred households’ carbon emissions each year.
Water
For direct water consumption—the water that’s evaporating into the atmosphere (compared to water withdrawal, which is water that is used and returned to the water system to be recycled—our estimates show that training GPT-3 uses 700,000 liters during two weeks of training. For indirect water consumption—water that’s evaporated for electrically generating the model—the amounts are likely two to three times more. And it’s estimated that 10 to 50 queries from ChatGPT consume about 500 milliliters (or 16 ounces) of water.
Land
Land footprint describes how much land you need to support an AI model. One large data center takes up as much land as half a football stadium. In just Northern Virginia, where lots of AI models are being engineered, there are more than 100 data centers. We’re still gathering data on the land footprint of AI throughout the US.
ARE THERE WAYS TO REDUCE THE ENVIRONMENTAL IMPACT OF AI?
Generally, if you have a smaller model, then you should expect to use fewer resources. That means that companies can try to develop smaller ones, but the current trend is still showing that the models are getting larger and larger. For example, the Llama 3.1 language model recently released is five times larger than its previous version. Once companies have the capability to develop powerful models that can answer complex questions—that’s what they’re aiming for—I think the next stage will be to make AI models more efficient.
It makes sense that we need to use resources to evolve our technologies, but we need to be careful about the overall impact. I’m hoping that the positive benefit will outweigh the negative environmental impacts sooner than later.
