30 % at Microsoft

AI arms race causes CO2 emissions to rise significantly

More and more IT companies are turning to "green" nuclear power to cover the energy requirements of their data centers.

(Bild: stock.adobe.com/ChristianSchwier.de)

Visits to the internet are leaving ever greater traces. Alongside the digital footprint, which has long been a topic of discussion among data protectionists, the digital CO2 footprint has rarely been in the media spotlight. However, this has changed in the wake of AI's triumphant advance, as its hunger for energy is huge.

According to a recent study by Goldman Sachs, CO2 emissions from data centers could more than double between 2022 and 2030.

ChatGPT and its ilk are already being used millions of times and the current AI energy consumption of a single user request alone is quite something: at 2.9 watt hours, a ChatGPT request costs around ten times as much energy as a Google search - by comparison, an LED light bulb consumes between six and nine watts per hour.

However, according to the study by the investment banking firm, AI - together with electrification and the relocation of industry - will lead to an increase in electricity demand in the USA, for example, that has "not been seen since the early years of this century", according to the analysis published in May 2024. Data centers will consume eight percent of US electricity by 2030, compared to three percent in 2022. Globally, the electricity demand of data centers is expected to increase by around 160 percent.

AI race causes CO2 emissions to rise
The AI race is already having an impact on greenhouse gas emissions, according to a report published by the World Economic Forum last year. Microsoft, an investor in ChatGPT manufacturer OpenAI, reported an increase in its CO2 emissions of almost 30 percent since 2020 due to the expansion of its data centers, while Google's emissions in 2023 were almost 50 percent higher than in 2019 - and here too mainly due to the data centers.

Using the example of ChatGPT, the report illustrates the hunger for energy of AI for its development alone: while the training of a model such as GPT-3 is estimated to consume just under 1,300 megawatt hours (MWh) of electricity - the annual electricity consumption of 130 US households - GPT-4 required 50 times as much energy. Overall, the computing power required for AI growth doubles approximately every 100 days.

Rising gas demand and nuclear power plant comeback
According to analyses by Goldman Sachs Research, servers & co. have so far been able to keep energy consumption at roughly the same level despite an increasing workload through more efficient use, which amounted to around 200 terawatt hours per year in the period from 2015 to 2019 - for comparison: electricity consumption in Austria as a whole was 61.08 terawatt hours in 2023 (source: https://energie.gv.at/). According to estimates, however, the total increase in electricity consumption by data centers due to AI is expected to increase by exactly these previous 200 terawatt hours per year between 2023 and 2030 - and by 2028, AI is expected to account for around 19 percent of data center electricity demand.

According to the study, renewable energies alone will not be able to meet demand; gas demand is expected to rise and nuclear power will also be used. At the end of 2024, for example, it was announced that a reactor at the decommissioned US nuclear power plant Three Mile Island is to be restarted. In this case, the customers are Microsoft's data centers.

On the other hand, there is also hope that AI can help combat climate change. A report by the US management consultancy Boston Consulting Group (BCG), for example, sees the potential for AI applications to help significantly reduce greenhouse gas emissions. According to the study "How AI Can Speed Climate Action", a five to ten percent reduction could be possible by 2030. AI efficiency could also be increased, thereby reducing energy consumption. However, AI currently seems to be faced with the task of delivering on its promises in a sustainable way.