It is well-known, that the internet as a whole infrastructure, with its billions of end-user devices and the entire server infrastructure behind that consume a lot of energy. Particularly with the rise of cloud and “off-site” computing, as well as the popularity of streaming services, Content-Delivery-Networks and other developments in networking infrastructure, the number, and energy consumption of data centres world-wide is growing dramatically. [1, 2]
We therefore ask ourselves: How big is the share of the hype surrounding artificial intelligence in this rapid increase and is it worth it to us?
Rising Energy Demand
In 2022, data centres alone consumed about 460 terawatt hours (TWh), with steep upwards tendency. The International Energy Agency (IAE) estimates data centres require between 600 and 1,000 TWh per year by 2026. This is the equivalent to the energy consumption of a larger European country. In comparison, the entire European Union currently produces roughly 4,000 TWh, annually. The IAE concludes that this steep increase is mostly attributed to the growth of Artificial Intelligence (AI) and Crypto Currency (i.e. “blockchain” technology). [1, 2, 3]
Two years ago, around one third of the 8,000 data centres worldwide were located in the US, making up roughly 4% of the country’s energy demand. Meanwhile, estimates indicate that these data centres’ energy consumption will grow by 50% in the coming two years. In Europe, where also quite a substantial amount of data centres are located (around 16%), projections indicate an equally steep rise in energy consumption particularly in financial and IT sectors in Frankfurt, London, Amsterdam, Paris, and Dublin. [2]
For some regions within Europe, the hype around AI has already a noticeable effect. Ireland, for example recently enacted a moratorium against building new data centres, since the existing ones already make up almost 20% (about two thirds of private households’ consumption) of the country’s energy demand. Utility companies in such regions are barely able to keep up with expansion and maintenance of the energy grid. The steep increase is especially obvious here, where energy consumption by data centres increased by 31% between 2021 and 2023 and quadrupled since 2015. [1, 4]
This begs the question what makes AI so power hungry or whether the growth in power consumption comes from the deepened digitalization and increased usage by consumers. While AI surely made its way into many of our day-to-day interactions with computers and the internet, studies suggest that indeed, AI is extremely wasteful even on a per-request basis. This is particularly true for generative AI (that generates text, sounds, images or even videos) and so-called multipurpose models that are generally able to give answers to all sorts of questions. [5, 6, 7]
Depending on a myriad of factors, including the processor structure of the cloud-computer in use as well as the specific content that is being generated, running such model can use the equivalent power of a single smartphone charge (or only about 100th of one charge for simpler tasks and models). Still, the average AI-query costs around ten times more than the “normal”, non-AI search query via any search engine. [5, 6, 7]
The highest energy consumption is however during the training phase of the models. A study on various BLOOM language models found that training and tuning the models costs around 200 Million to 600 Million times more energy than a single query by and end-user. This is particularly grave because models are typically trained, adjusted, tuned and re-trained constantly to stay “ahead of the pack” and gain new skills and features all the time. [5, 7]
This results in vast energy consumption in practice of course. The training of GPT-3, the model that powered ChatGPT when it was publicized in 2022, used up 1,287 MWh of energy – the equivalent of what roughly 400 households in Germany consume per year. [7, 8]
Lack of Efficiency
Though energy consumption is expected to rise dramatically world-wide either way, as we (hopefully not so) slowly move away from burning fossil fuels and start to electrify transportation, heating/cooling and all other kinds of energy demands. While cost of electric energy is currently falling due to expansion of cheap and efficient renewable energy production, it will remain a precious and limited good in the short- and mid-term. [9]
The question remains, whether AI delivers on its promise to increase productivity and boosting economies. Surveys among executives and workers alike suggests otherwise for the time being. A study by the Upwork Research Institute among CEOs and workers in the US, Canada, UK, and Australia summarizes that almost all executives expect AI to boost productivity and 85% of questioned companies demand or encourage its use. However, workers within the same companies appear to think otherwise. Nearly half of them have no idea how AI could even lead to higher productivity of their work and more than 75% think the usage of AI decreased their performance. [10, 11]
Critics support this argument and argue, that the entire premise of generative AI boosting productivity at least in IT jobs is wrong from the start. Writing new code is only a small part of the job – albeit it is typically the one that generates profit. Instead, they say, the sector is risking to loose competence and understanding of project infrastructure, software architecture and maintenance. Likewise, studies suggest that use of generative AI by students may hinder their educational outcomes. [11, 12, 13]
Even global investors like Goldman Sachs criticize the wide-spread use of generative AI as mean to increase productivity. They claim, that it is unlikely the heavy investments in developing, implementing and rolling out AI will ever pay off for companies. They are also sceptical of the proclaimed transformational potential of generative AI and the benefits it may have for the economy as a whole. [14]
Summary
Artificial Intelligence, including generative models, are undoubtedly here to stay. Over the past few years, they have made their way into many aspects of our day-to-day life. Unlike many other technological advances however, we reserve little thought for the resources that go into providing such services, be it the energy for the required cloud infrastructure, the freshwater reserves to keep the hardware cool, harvesting, and usage of rare minerals required for the microchips in them, or the countless hours spend by the workforce in low-paid countries on the other side of the world to feed content to the machines and moderate their outputs. [9, 15, 16]
And of course, not all uses of “AI” are unnecessary per sé, as there are in fact many fields where specialized models could indeed improve the status quo, however the hype and promises that are given by stakeholders in AI are unmatched to what we have ever seen before. Meanwhile, (generative) AI has little to show to actually improving economic outcomes or (more importantly) how we work, live and generally our day-to-day life. Technologists and tech companies claim, that it is only a matter of time until the tipping point is reached and AI (or the current day’s technology) will change the way we live for the better in a blink, which has always been the case too. Social issues can rarely be solved by more tech.
Generally speaking, it is likely a good idea to question the promises given around any tech (including “AI”) critically and using it where it makes sense (or is fun!), all the while keeping in mind what this usage entails.
Sources
- Baraniuk, C. (2024): Electricity grids creak as AI demands soar. URL: https://www.bbc.com/news/articles/cj5ll89dy2mo
- International Energy Agency (2024): Electricity 2024 – Analysis and forecast to 2026. URL: https://www.iea.org/reports/electricity-2024 [p. 31ff]
- International Energy Agency (2024): Electricity – Europe. URL: https://www.iea.org/regions/europe/electricity (viewed: 2024-10-05)
- Shortt, R: (2023): Households cut electricity use, consumption by data centres up. URL: https://www.rte.ie/news/business/2023/0612/1388694-power-consumption-by-data-centres-jumped-31-last-year
- Luccioni, A.,Jernite, Y., Strubell, E. (2023): Power Hungry Processing: Watts Driving the Cost of AI Deployment? URL: https://doi.org/10.48550/arXiv.2311.16863
- Vincent, J. (2024): How much electricity does AI consume? URL: https://www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption
- EnviaM Group: Stromverbrauch Künstliche Intelligenz. URL: https://www.enviam-gruppe.de/energiezukunft-ostdeutschland/verbrauch-und-effizienz/stromverbrauch-ki
- Statistisches Bundesamt (2023): Stromverbrauch der privaten Haushalte nach Haushaltsgrößenklassen. URL: https://www.destatis.de/DE/Themen/Gesellschaft-Umwelt/Umwelt/UGR/private-haushalte/Tabellen/stromverbrauch-haushalte.html (viewed: 2024-10-05)
- Calvert, B. (2024): AI already uses as much energy as a small country. It’s only the beginning. URL: https://www.vox.com/climate/2024/3/28/24111721/climate-ai-tech-energy-demand-rising
- Monahan, K., Burlacu, G. (2024): From Burnout to Balance: AI-Enhanced Work Models. URL: https://www.upwork.com/research/ai-enhanced-work-models
- Doctorow, C. (2024): AI’s productivity theater. URL: https://pluralistic.net/2024/07/25/accountability-sinks/
- Geuter, J. (2024): Productivity gains in Software Development through AI. URL: https://tante.cc/2024/08/27/5312/
- Marx, P. (2024): Will generative AI have many long-term benefits?. URL: https://disconnect.blog/roundup-will-generative-ai-have-many-long-term-benefits
- Koebler, J. (2024): Goldman Sachs: AI Is Overhyped, Wildly Expensive, and Unreliable. URL: https://www.404media.co/goldman-sachs-ai-is-overhyped-wildly-expensive-and-unreliable
- Marx, P. (2023): A.I.’s dirty secret. URL: https://www.businessinsider.com/chatgpt-ai-will-not-take-jobs-create-future-work-opportunities-2023-2
- Marx, P. (2024): The data workers whose labor powers AI. URL: https://disconnect.blog/roundup-the-data-workers-whose-labor-powers-ai
Jan is co-founder of ViOffice. He is responsible for the technical implementation and maintenance of the software. His interests lie in particular in the areas of security, data protection and encryption.
In addition to his studies in economics, later in applied statistics and his subsequent doctorate, he has years of experience in software development, open source and server administration.