Artificial Intelligence

The Energy Implications of Artificial Intelligence: A Very Brief Overview

"We just interact with the technology and we’re not actually aware of how much resources – electricity, water, space – it takes.” – Sandra Wachter, University of Oxford as quoted in Sparks, 2023.

Much of the discussion about Artificial Intelligence (AI) in education centers on the implications of its use by students, especially in relation to academic integrity. However, ethical and environmental concerns are worth discussing, especially as faculty embrace AI. This blog post will briefly look at the energy requirements of AI models and possible future scenarios that could unfold due to the increasing usage of them.

A recent study by Alex de Vries, published in the journal Joule, demonstrates some of AI models' current and potential energy usage requirements. Currently, a typical Google request uses approximately .3 Wh of energy and a ChatGPT request uses just under 3 Wh. If Google was to move to an AI-powered search system, we could see requests that use close to 9 Wh of energy per interaction (de Vries, 2023). Put into more dire terms, in what de Vries calls a “worst case scenario,” AI-powered Google requests could consume as much electricity per year as the entire country of Ireland – 29.3 TWh per year.

A less dire prediction of future AI energy consumption looks at the company NVIDIA and its share of the market for AI servers. NVIDIA currently holds approximately 95% of the market share for AI servers and is slated to deliver 100 000 servers in 2023. If operating at full capacity, combined, these servers would use 5.7-8.9 TWh per year, which is substantially less than the Google AI worse case usage scenario noted above. Having said that, if we look ahead to server production, in 2027 NVIDIA could potentially be shipping 1.5 million servers which points to an astounding yearly consumption of 85.4 – 134.0 TWh of electricity (de Vries, 2023).

Coupled with the large amount of energy AI servers use is their massive consumption of water. As Ren (2023) points out, AI servers consume water in two ways: 1. AI server energy consumption generates a lot of heat. Data centers use cooling towers and/or air from outside to dissipate that heat outdoors and prevent servers from overheating. These cooling towers use large amounts of fresh water. 2. Electricity generation consumes water through cooling at thermal power plants and nuclear plants as well as through evaporation at hydropower plants. By 2027 we could potentially see total water withdrawal through AI operations of 4.2 – 6.6 billion cubic meters per year which would be 4-6 times what Denmark withdraws (Ren, 2023).

As we move forward with AI models, it’s imperative we look beyond simply whether these tools will enhance teaching and learning and look at the overall implications of their usage. Current predictions show a massive increase in energy consumption and water withdrawals in the very near future with AI model usage. Although Open AI states they “give considerable thought about the best use of our computing power” that simply may not be enough and Legislation to force companies to be more transparent about their energy use ultimately may be the only way to push companies to act more responsibly (Sparkes, 2023).

References

de Vries, A. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191–2194. https://doi.org/10.1016/j.joule.2023.09.004

Ren, S. (2023). How much water does AI consume? The public deserves to know - OECD.AI. Retrieved December 12, 2023, from https://oecd.ai/en/wonk/how-much-water-does-ai-consume

Sparkes, M. (2023). Should we be worried about AI’s growing energy use? New Scientist. Retrieved December 12, 2023, from https://www.newscientist.com/article/2396064-should-we-be-worried-about-ais-growing-energy-use/