AI and Carbon Footprint

AI, especially the big models, is hungry for data and operating power. These systems work on millions of samples to learn. Training a model can take weeks, even months in some cases. Thousands of processors operate simultaneously, and considerable amounts of electricity are consumed during this process.

For example, the required energy to train a language model such as GPT-3 produces approximately 550 tons of carbon dioxide emissions. This is equal to the emissions resulting from a plane flying 700 times from America to Europe. And this is only the training process. The model continues to consume energy each time it is used.

Furthermore, AI systems are hosted at giant data centers since they operate over clouds. The servers at these facilities operate continuously 24/7. Cooling systems are utilized to lower temperatures, and millions of operations are performed every second.

Consequently, every question we pose to "ChatGPT" generates carbon emissions, albeit in minimal amounts. These minimal impacts create a significant environmental load when combined with billions of users.

The real power behind AI is the giant data centers. These centers are giant facilities in which thousands of servers operate simultaneously, supported by cooling systems. They are the physical heart of the digital world, although they are not visible.

Please check out our previous blog post here on Green Data Centers.

A data center is full of computers operating constantly. These machines generate high heat. Heating systems require a significant cooling infrastructure in order to continue to operate. And cooling systems work with electricity. The result? Double energy consumption: both for servers and for the systems that keep them cool.

According to the International Energy Agency, data centers consume 1.5% of the electricity in the world. This ratio is even more than the total energy need of some countries. Furthermore, this consumption is expected to increase. As AI-based services become more widespread, the pressure on data centers is also growing.

Environmental impact of data centers is not only limited to energy. Some facilities use large amounts of clean water for cooling purposes. Moreover, these centers are generally connected to energy grids operating with fossil fuels. In other words, they do not only consume electricity but also emit carbon.

This question needs to be asked at this point: Do we become more insensitive to the planet as we make AI smarter?

Concrete facts are of great importance when discussing carbon footprints. So, how much emissions does AI generate?

According to a study conducted by the University of Massachusetts Amherst, the training of only one big language model can generate 550 tons of carbon dioxide emissions. Another example is related to Google’s AI systems. According to the company’s own data, only AI operations consumed millions of kilowatts of energy.

And the server infrastructure used for AI models emerging as a result of the cooperation between Microsoft and OpenAI requires high energy as well. Both the training and daily usage of these systems consume much more energy than annual electricity consumption of an average household.

As the data clearly shows, AI produces significant amounts of carbon emissions both during training and during usage. This situation make the notion of “invisible digital pollution” even more visible.

Technology companies are taking some steps to mitigate environmental impacts of AI. Giants such as Google, Microsoft, and Amazon are pledging to power their data centers with renewable energy. This transition, however, progresses slowly while not being equally realized in every region.

It is not possible to reduce carbon footprint of AI by only changing the source of the energy. The real issue here is in the size and efficiency of the models. Companies are increasingly developing bigger and stronger models every year. These models are being trained with more data, thus requiring more energy. In other words, each new model emits more carbon than the previous one.

Furthermore, the system uses energy not only during training, but also during usage. This consumption increases incrementally when an AI model is used by millions of people every day. This means that training them once does not mean it is finished.

Reducing carbon footprint of AI is not under the responsibility of only technology companies. This is an issue that should be jointly dealt with by individuals, institutions and public policies.

What can we do as individuals? The first step is to be aware. Knowing that each “free” digital service has an environmental cost can cause us to question our consumption habits. Even not using the systems we do not really need, not operating the big models just out of curiosity can create a difference.

Institutions, especially companies developing digital products or processing data, can prioritize eco-friendly options in their infrastructure preferences. Algorithms consuming less energy, simpler designs, and data centers with high energy efficiency play a critical role in this sense.

And policy makers should take steps towards transparency and regulation. Reporting carbon footprints can be made mandatory for the companies. Moreover, offering incentives for green technology investments provides not only economic but also environmental gains.

This multi-player problem should also have a multi-dimensional solution. It is not only the production of technology that will be decisive for a sustainable future, but also how we use it.

Click here to read our blog post titled “Climate Crisis and The Role of Individuals”.

As AI revolution steams full ahead, it is impossible to ignore the environmental impacts of this technology. Reducing our carbon footprint is not only the responsibility of technology companies, but also of individuals and policy makers. A sustainable digital future can only be possible through conscious steps we will collectively take.