Why does my AI have a carbon footprint?

Just because it's in the metaverse doesn't mean it's not contributing to climate change 🫠

This week, Biden signed an executive order on “advancing U.S. leadership in artificial intelligence infrastructure.” The order is all about providing federal support for AI’s enormous land and energy needs, including prioritizing and expediting permitting for data centers being built on federal lands, and potentially speeding up environmental reviews.

Biden’s order comes at a time when citizens all over the world, from Chile to Spain, Arizona to Ireland, have had to contend with whether critical resources like electricity and water will be used to support their communities, or fuel the production and operation of data centers. In many cases, these are communities already vulnerable to scarce resources, unprecedented droughts, and a strained electric grid.

Data centers are no doubt a necessary part of our lives. The boundary between our online activities and our in-person experiences is blurry, and data centers house the digital world that we’re all dependent on. But we live on a planet with limited resources, under the constant threat of climate disaster. We need to thoughtfully and democratically choose how we allocate those resources and what technologies and infrastructure we prioritize.

But before we get in too deep discussing these tradeoffs, it’s worth unpacking why AI needs so much land, energy, and water in the first place.

It’s all virtual, so why is there an environmental footprint?

We’re all more or less aware that using plastic and flying on airplanes require natural resources and have an impact on the environment. What’s less obvious is AI’s reliance on natural resources, which is easily obfuscated by all the talk of super-human intelligence.

But every step of the AI lifecycle has a very real environmental footprint, from the extraction, manufacturing, and shipping of physical hardware we use to train, host, and store AI models, to the energy and water required to power and cool these servers (I love this topic so much I designed an entire free course on it for AI developers).

HARDWARE

AI models run on servers and chips in data centers. Creating all this infrastructure requires clearing land for data centers, extracting raw materials for CPUs/GPUs/etc, manufacturing, and shipping. There are environmental costs associated with each of these steps.

If we’re talking specifically about the carbon emissions from AI, this is the embodied carbon, which refers to all the emissions produced in the supply chain for a given item or product. Your shoes, your phone, your coffee cup, and your AI all have embodied carbon because resources had to be extracted, manufactured, and transported, all before you could use these things.

TRAINING

Once all of the physical infrastructure is built, software engineers write code that runs on the hardware in these data centers.

The process of training the generative AI models we’re all using these days involves massive amounts of data that needs to be crunched and processed by a computer. All of this processing power requires energy in the form of electricity to power and cool the machines that perform this computation.

To be clear, it’s not just generative AI models that require energy. Anything we do in the digital world requires computational power, which requires energy. But training generative AI models is particularly energy intensive because these models are so large and require so much data.

Training GPT-3, for example, used ~1,300 megawatt-hours of electricity. That’s about equivalent to the energy used by 123 american homes each year. Unfortunately the exact energy consumption of more recent generations of generative AI models (which are significantly larger) is unknown because most players in the space do not disclose this information.

So where does this energy come from? Well it’s produced by power plants that are sourcing energy from carbon emitting fossil fuels like coal or gas, or non-carbon emitting sources like wind or solar. This energy then flows through a network of transmission lines and substations, which is known as the electric grid, to power our homes, offices, restaurants, hospitals, and even data centers.

The electricity used to power data centers is no different from the electricity used to cool or heat your home, and this is why AI has a carbon footprint. It requires energy in the form of electricity, which likely came from carbon emitting sources like coal or natural gas operating on the electric grid.

INFERENCE

AI inference, which is just a fancy word for when a user interacts with and uses an AI model to get a prediction, is also power hungry. It used to be that model training was the most energy intensive part of the AI lifecycle, but that’s changed with generative AI for two reasons: 1) Model size/complexity 2) Popularity and use.

Generative AI models are large (we don’t call them large language models for nothing). When you send a message to ChatGPT/Claude/Gemini, the model has to perform a lot of mathematical computation to process the input text and generate output text. Researchers estimate that for ChatGPT, even assuming a single query per user, the energy costs of operating the model for inference surpass its training energy costs after just a few weeks or months.

In addition to these generative AI models just being more compute intensive, we’re also using them all the time all over the place! So while any individual message to ChatGPT might use a negligible amount of energy, those numbers add up fast when we consider the scale of this technology and how generative AI is being shoehorned into every corner of our lives. Compared to older non-generative AI models, this new generation of generative AI models can use up to 30x more energy answering the same set of questions.

Just think about that next time you use ChatGPT to calculate the tip at dinner instead of using an actual calculator…

In addition all of the electricity required to power servers during both training and inference, water is required to cool the servers in data centers. Doing all that math can make these machines heat up (you know, like when you have 1000 chrome tabs open and your laptop fan starts hyperventilating). Researchers estimate that an average AI datacenter “uses around 550,000 gallons of water daily.”

AI can’t fix climate change

It’s true that everything we create has an environmental footprint. In comparison to all of the other industries that contribute to environmental degradation, the ~2%-4% that cloud computing contributes to overall emissions might seem small. But that number is growing rapidly, as evidenced by the number of cloud providers who will miss their climate targets because of AI. And cloud computing already emits more global greenhouse gas emissions than the entire commercial airline industry.

Last year, the US National Security Advisor stated that we will need to add “tens or even hundreds of gigawatts” of energy to the grid to power data centers for AI. Putting that number in context, the entire US electric grid is about 1200 gigawatts of capacity. So 100s of gigawatts could amount to 10% or even 30%+ of our overall electric capacity being used to power AI.

Among the justifications for prioritizing this massive allocation of energy to data centers, Biden’s statement on the executive order begins with “cutting-edge AI will have profound implications…keeping communities safe by mitigating the effects of climate change.” This sentiment echos the delusional attitudes of Sam Altman, Eric Schmidt, and other tech pundits who claim something to the tune of yes we know AI is the reason we’re all failing to hit our climate targets, but it’s okay because one day AI will generate an idea so good it will solve climate change.

The unfortunate truth is that AI is not going to solve climate change. I do strongly believe that AI will contribute in meaningful ways to the discovery of new materials that have low environmental impact, to making renewable energy more reliable, and to helping us better predict and understand our changing climate. However, AI is incapable of solving the structural, cultural, and emotional problems that comprise our climate disaster. As MIT energy editor James Temple eloquently put it, “technological advances are just the start—necessary but far from sufficient to eliminate the world’s climate emissions.”

Our world runs on fossil fuels. Our economy is built on the linear path from resource extraction to landfill. We are obsessed with overconsumption, and burdened by sunk costs. All of these problems will only get worse the more we expand data centers, extract rare earth minerals for servers and cables, and redirect energy and water to power and cool AI. Technology can help us chip away at the problems we face, but prioritizing AI at all costs is only a distraction from the true obstacles threatening humanity.

Our cities are on fire, our land is drying up, our forests are vanishing. Reversing the impacts of climate change is already hard enough. Let’s not make this problem even harder on ourselves.