Posted: 27 Sep. 2022

Artificial intelligence: Worth the environmental cost?

Learning the hard truth about AI’s carbon footprint

AI seems like a no-brainer in many business contexts. But the carbon emissions produced to run it may be a very strong reason to question its use. Before jumping on the AI bandwagon, business leaders should scrutinise its application carefully. 

Artificial intelligence has been welcomed inside the walls of companies in nearly every sector around the world. No one is denying its utility; it’s performing life-saving medical monitoring and diagnosis, chatting with customers to help respond to their needs, and revolutionising the productivity and safety of goods shipping. It also has major benefits in the area of sustainability, detecting illegal GHG emissions and deforestation while optimising the watering and fertilisation of crops. 

But there is a price to be paid. This remarkable innovation is a massive consumer of IT resources, meaning a massive electricity over-consumer. It’s also a significant user of physical resources, including gold and the rare earth metal neodymium. 

For all the efficiency AI adds to operations, training an AI system is inefficient. Even to recognise a cat, it needs more than 10,000 pictures of cats. According to one study, to train a transformer – a type of deep-learning algorithm – the necessary electricity can emit more than 626,000 pounds of CO2 (284 tonnes); that’s nearly five times the emissions of the average American car. 1

And training a learning algorithm is just the beginning. The process of deploying it in real-time applications, known as interference, consumes even more energy than the training does.2  Executing the learned algorithm requires as much as 40 times the energy a comparable C++ code program would. Consider that Python, a commonly used interpreting language, consumes 75.88 units of energy; C++ consumes just 1.34 units.3  Then there’s the collection of all the data needed to improve AI algorithms, which restarts the whole energy-consuming, data-gathering cycle again.

Finally, we need to discuss the impact of AI on the usage of physical resources. AI depends on storing and processing huge amounts of data, which must be stored and maintained using devices. If we continue our unlimited storage of data, to feed our AI, the use of neodymium in 2025 will exceed the current supply in Europe 120 times.

AI is here to stay, and produces astonishing results. But there are certain applications for which business leaders should opt out of AI; just because we can do something, it doesn’t mean we ought to. The AI research seems to be chasing progress by building larger and larger models, such as GPT3 and ResNeXt. But these models have shown diminishing performance in terms of accuracy.4 Take the mega-model GPT3 as an example: It has lower returns than the smaller GPT2 but requires significantly more resources. 2, 5

So it’s safe to conclude that bigger is not always better. But the opposite ideology dominates the current AI research and application agenda. It threatens to cause future environmental damage with even greater emissions.4  Even those unconcerned about AI replacing human workers, or its vast carbon footprint, might be swayed by the sheer company cost of all that environmental waste.

The bottom line underneath AI use is this: You have a choice, and the consequences must be weighed against the value of application. Artificial intelligence should not be a no-brainer. 

1 Ben Dickson, “AI could save the world, if it doesn’t ruin the environment first”, PC Mag, 17 April 2020,, last accessed 27 April 2022. 

2 Rob Toews, “Deep Learning’s Carbon Emissions Problem”, Forbes, 17 June 2020,, last accessed 25 May 2022. 

3 Kesk -*-, “What Are the Greenest Programming Languages?”, Medium, 2 December 2021,, last accessed 11 July 2022. 

4 Roy Schwartz, Jesse Dodge, Noah A. Smith, Oren Etzioni, “Green AI”, Allen Institute for AI, July 2019,, last accessed 25 May 2022. 


Eric Onderdelinden

Eric Onderdelinden


I'm a very experienced enterprise architect and lead the Enterprise Architecture offering within the Netherlands. I'm specialized in Enterprise Architecture with a focus on the private sector. Besides in depth knowledge of Industrial processes and products I bring a wealth of experience in data and technology , including cloud services and desired agility, to the table. Recent assignments include pre-merger assessment and post merger integration. I work on project dealing with application portfolio rationalization, business case development and TCO. I publish on a regular basis in IT and business magazines. I support companies worldwide with the establishment and maturing of their EA practice. I'm a teacher in the master class enterprise architecture organized by the NAF. Besides that I'm a memebr of the board for Platfrom Digitale Wendbaarheid. Currently I'm working on a PhD concerning the value contribution of IS in M&A.