Generative artificial intelligence (GenAI) is here to stay. Today, millions of people use tools like ChatGPT, Midjourney or Suno AI to create various types of content. As users, we open an application, type a prompt and voilà! What we want to see appears as if by magic. In real terms, there is little magic in this. What we see is the product of a complex process which includes collecting a huge amount of data, training language models and inference processes.
These tasks are not usually performed on our devices. The companies behind GenAI applications have data centers with huge computing power that do most of the work. It is no secret that this type of infrastructure requires the latest processing units, for example NVDIA’s H100, but it is a bit difficult to size up the amount of resources involved because it is not so easy to find information about it. Specifically, a lot of energy and water.
A ChatGPT query, a bottle of water
Data centers consume water, but how much water? As we say, finding an answer is not an easy task, but we have some data that can help us understand the issue. The current rise of GenAI has been accompanied by a significant global increase in water consumption by Microsoft. According to an environmental report from the Redmond company, this metric grew by 34% from 2021 to 2022. Interestingly, ChatGPT was launched in November 2021, and its model had been training for some time before.
If you’re wondering why we’re linking ChatGPT, a product from OpenAI, to Microsoft, it’s because the company behind the most famous AI chatbot of the moment has a strategic relationship with the creators of Windows. Basically, OpenAI uses Azure AI’s cloud infrastructure to train and run its AI products. But it’s not just Microsoft. Google’s worldwide water consumption soared by 20% in the same period, a time when AI was (and still is) booming.
A study published by The Washington Post sheds some more light on this scenario. Researchers from the University of California concluded that a request to write a 100-word email on ChatGPT results in a consumption of 519 milliliters of waterThis is an average calculation, but it helps us get a more tangible idea of what happens with these data centers. We are talking about little more than a half-liter bottle of water, like the one we can Buy in any market.
Water is a key resource for cooling systems in the largest data centers, although its consumption depends on the location of the centers, the time of year and the weather conditions. Environmental concerns related to the operation of such thirsty equipment are not new. Technology companies, often driven by regulations and activists, have long been working on new solutions to limit resource consumption.
We’ve seen Microsoft submerge an experimental data center in the ocean to address the temperature issue, and we’ve also seen companies like Meta and Google wanting to put servers in geographic locations where water seems to be in short supply. As we noted above, GenAI doesn’t happen by magic, and it looks like data centers will continue to multiply in different parts of the world behind multi-million dollar investments in a race to dominate the sector.
Images | Freepik | Microsoft | WorldOfSoftware with Bing Image Creator
At WorldOfSoftware | Meta and Spotify warn the EU: regulation will cause it to miss the AI train