Artificial intelligence (AI) has become an integral part of our daily lives. But the smart companions of everyday life come at a cost to the environment: An increasing number of AI models are being trained and deployed on energy-intensive servers in large data centers. Training a single neural network produces as much carbon dioxide as operating five cars. The energy consumption of artificial intelligence is also coming under increasing scrutiny. The rapid development of AI technologies requires more and more computing power, which in turn consumes huge amounts of electricity and resources. Unfortunately, the enormous water footprint of AI models – many millions of liters of drinking water extracted or consumed for power generation and server cooling – has gone largely unnoticed.
This article discusses the environmental impact of digital devices. We’re looking into carbon emissions from digital gadgets, energy use from web, and water use from artificial intelligence. The article ends with wishes and recommendations for consumers and for the manufacturers and operators of this technology. Despite the many benefits and potential of AI, its environmental footprint is an important topic of public discussion.
The environmental impact of digital devices
Few of us are aware of the impact that the manufacture and disposal of our digital devices has on the environment. Nor do we know how much energy is consumed in data networks and data centers. This data is rarely published because service providers and manufacturers of digital products are reluctant to share it. In addition, the energy and greenhouse gas emissions required to manufacture the devices are not disclosed on the devices themselves, and the energy requirements of the Internet and data centers are not included in any calculations.
Digital devices have profoundly changed our lifestyles, with many of our activities – education, work, leisure, commerce, consumption, diet, partner choice, and even the digital cemetery – already offloaded to our smart assistants. Yet we are still largely unaware of the environmental footprint they leave behind.
Digital carbon footprint
To fill this knowledge gap, BUND (Bund für Umwelt und Naturschutz Deutschland) commissioned Jens Gröger of the Öko-Institut in Berlin to create a digital carbon footprint for the production and use of digital devices and services.
Jens Gröger (2020). Digitaler CO2-Fußabdruck Datensammlung zur Abschätzung von Herstellungsaufwand, Energieverbrauch und Nutzung digitaler Endgeräte und Dienste. Öko-Institut e.V., Berlin. Digitaler CO2-Fußabdruck | oeko.de.
Jens Gröger aims to create a database that can be used to calculate the ecological footprint caused by using digital devices. He follows the calculation method of the CO2 calculator of the Federal Environment Agency (uba.co2-rechner.de). The environmental impact of digital devices is calculated in terms of CO2-equivalent greenhouse gas emissions (CO2e for short). The unit of measurement for greenhouse gas emissions is the kilogram of carbon dioxide equivalent (kg CO2e).
Digital devices consume more than 700 kg of CO2
Jens Gröger has found that the production and use of digital devices and services in Germany releases about three-quarters of a ton of greenhouse gas equivalent CO2 per capita and year. The production of the devices accounts for the largest share.
The results are shown in the following chart:
Diagram of the CO2 footprint for the intensive use of digital technology, taken from Jens Gröger (2020). Digital CO2 footprint data collection for estimating production costs, energy consumption and use of digital devices and services. Öko-Institut e.V., Berlin. Digitaler CO2-Fußabdruck (oeko.de).
With roughly half a ton of greenhouse gas equivalent annually, the production of digital devices accounts for about half of the carbon footprint of intensive use of digital technologies. The three largest individual contributions to greenhouse gas emissions from manufacturing come from televisions, desktop PCs with monitors, and laptops.
Televisions are also the largest single contributor to the greenhouse gas emissions from the use phase of digital technologies. This is followed by game consoles and routers. The most intensive use of digital technologies takes place on the Internet: Mobile Internet, music streaming, video telephony, video streaming and online storage together generate 138 kilograms of CO2 per year. The largest amount of greenhouse gas equivalents is consumed by the streaming of high-definition movies.
Energy consumption of the Internet
The Internet consumes an enormous amount of electricity, approximately 416 TWh/year. To put that in perspective, that’s more than the entire power consumption of the United Kingdom. From the data centers, to the transmission networks, to the billions of connected devices available to you, they all use electricity and produce carbon emissions equivalent to those produced by global aviation.
Website carbon emissions
When an average website is accessed on the Internet, approximately 0.8 grams of CO2 equivalent are generated per page view. This equates to 102 kg CO2e per year for a website with 10,000 page views per month.
London-based Wholegrain Digital has developed a method for estimating a website’s CO2 emissions. The free tool is powered by renewable energy.
If you would like to calculate the carbon footprint of your website, you can access the CO2 calculator at the following link: Website Carbon Calculator v3 | What’s your site’s carbon footprint?
Total greenhouse gas emissions
In Germany, housing, mobility, food, public infrastructure, and other consumption cause a total of 11.6 tons of greenhouse gas emissions per capita per year. By comparison, the intensive use of digital devices and services produces relatively small amounts of greenhouse gases – up to 1 ton of CO2e per user per year.
However, the hype surrounding digital devices is far from over, and the digital carbon footprint is expected to continue to grow significantly. If we are to become carbon neutral in the coming decades and limit global warming, digital technologies must also reduce greenhouse gas emissions.
However, focusing on the carbon footprint alone may not be enough to enable truly sustainable AI.
The Water Footprint of AI
In an effort to address global challenges such as climate change, artificial intelligence models are increasingly being trained and deployed on energy-intensive, high-performance computers in large data centers. Unfortunately, the enormous water footprint of AI models – the tens of millions of liters of water extracted from groundwater or consumed to generate electricity and cool servers – has been largely ignored. Without action, this growing water consumption could lead to a climate catastrophe.
Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren (2023). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. University of Colorado Riverside and University of Texas Arlington.29.10.2023. https://doi.org/10.48550/arXiv.2304.03271.
The Earth’s natural ecosystem is a water cycle. However, available and usable freshwater resources are extremely limited and unevenly distributed across the planet. Four billion people, or about two-thirds of the world’s population, already suffer from acute water shortages for at least one month a year.
Big thirsty data centers
The researchers, led by Pengfei Li, explained that large data centers, where many AI models are trained and deployed, are known to be very energy-intensive. This is especially true for large models such as GPT-3 and GPT-4 for voice services. Together, they consume 1 to 2 percent of the world’s electricity. Less well known is that data centers are also extremely thirsty. Data centers use vast amounts of water to cool servers on site, to generate electricity at a off-site facility, and to manufacture servers in the supply chain.
Google data centers use large amounts of potable water
The researchers estimate that in 2022, Google’s data centers alone account for 25 billion liters of water withdrawn directly from the hydrologic cycle and nearly 20 billion liters of water used for on-site cooling. This represents a 20% increase in total data center water withdrawals and consumption compared to 2021, while Microsoft’s total water consumption increased by 34% over the same period. One reason for these significant increases is likely to be the growing demand for artificial intelligence applications.
In addition, the researchers estimate that Google, Microsoft, and Meta collectively used 2.2 billion cubic meters of water on-site and off-site in 2022. By comparison, Denmark withdraws only half as much water from the water cycle in a year, including municipal, industrial, and agricultural use.
Training GPT3 model for voice service
Scientists at the Universities of Colorado Riverside and Texas Arlington have calculated that training the GPT-3 model for voice services at Microsoft’s state-of-the-art data centers in the United States could consume a total of up to 5.4 million liters of water – 700,000 liters of which are used on site. Or on a scale you can imagine: GPT-3 uses a 500 ml bottle of water for about 10 to 50 reactions, depending on when and where it is used. These numbers could increase for the much larger GPT-4 model.
Influencing the water footprint
In addition, the researchers point out that the choice of when and where to train a large AI model can have a significant impact on the water footprint. To reduce the water footprint, it is advisable to avoid high temperature hours during the day. To reduce the carbon footprint, however, it is better to take advantage of sunny hours when solar energy is abundant. Unfortunately, the researchers do not offer a way out of this dilemma.
What the scientists would like and what they recommend
Scientists unanimously want more transparency from AI model developers and data center operators. They also recommend that industry and academia work together to research less compute-intensive algorithms and more energy-efficient hardware.
Not only manufacturers, service providers, data center operators, software developers, and legislators have a special responsibility to reduce greenhouse gas emissions, but also consumers themselves. We can reduce our personal carbon footprint by changing our individual digital consumption patterns.
Buy only high-quality, durable products
Digital equipment must therefore become more energy efficient not only in use, but also in production. To this end, digital devices should be of high quality and repairable, for example through modular design, replaceable batteries and spare parts supplied by the manufacturer. We should also choose devices that guarantee a long-term supply of software updates and spare parts.
As consumers, we should also make sure that we buy high-quality, durable products and use them for as long as possible.
Surfing the Internet
A laptop or tablet is a better alternative than a desktop PC with a monitor for surfing the web. When it comes to Internet transmission options, Internet access via wired LAN or wireless WLAN is significantly more energy-efficient than access via cellular networks.
Internet routers act as an access point to the Internet. As such, they are typically in use around the clock to maintain Wi-Fi or control connected phones, smart home devices, or other network components. However, they also have power-saving features that typically require the user to enable. For example, Wi-Fi networks can be switched off at certain times (e.g. at night) to save energy, as the constant standby mode of routers consumes a disproportionate amount of energy.
Additional tips
When transmitting video to small screens, lower screen resolutions can be used without a noticeable difference in quality. Or, for videoconferences or online seminars, inactive participants can turn off microphones and video cameras to reduce the amount of data. In addition, simple phone calls can save significantly more CO2e than video calls. Cloud storage and online mailboxes should also be emptied on a regular basis.
Can AI still help?
Thorsten Staake, business information scientist and expert in energy systems, puts the results into perspective: The core statement is correct. But if programming a navigation app consumes as much CO2 as five cars, but emissions can be saved through intelligent traffic control, AI can also make a contribution to climate protection.
Adrian Lobe (2019). ENERGIEVERBRAUCH:KI ist alles andere als grün. Spektrum.de. 26.07.2019. Künstliche Intelligenz verbraucht für den Lernprozess unvorstellbar viel Energie – Spektrum der Wissenschaft
That’s why it’s important that we consumers develop an awareness of digital consumption. In other areas of consumption, education for a climate-friendly lifestyle is already more advanced.
Please find the German translation of this article here: Künstliche Intelligenz: Zu hungrig und zu durstig