Unveiling ChatGPT's Energy Footprint: One Second in an Oven
AI query consumption: Estimated electrical energy usage per inquiry - AI Consumption per Query: Here's the Amount of Energy Drawn for One Request
Taking a peek at the energy consumption of AI software like ChatGPT, OpenAI, the company behind the technology, reveals that an average query equates to the power consumption of an oven running for about a second [1][2][3]. In technical terms, this translates to 0.34 watt-hours of electricity per query.
In a blog post, OpenAI's CEO, Sam Altman, went on to explain that this energy use is akin to a high-efficiency lightbulb burning for a couple of minutes or a toaster using energy for around 13 seconds [4][5]. The water usage per query is remarkably small, around 0.000085 gallons (0.00032176 liters), or roughly one-fifteenth of a teaspoon.
Warnings about the rising power demand from artificial intelligence have circulated for quite some time. Despite the efficiency improvements in chip and server technology, the sheer volume of queries still leads to a significant escalation in power demand for AI data centers. High-profile companies such as Microsoft, Google, and Amazon are now exploring nuclear power as a potential solution in the U.S. to meet the increased energy needs without excessive carbon dioxide emissions.
The water consumption in data centers is another pressing concern, particularly as these facilities require cooling. Over the past few years, researchers have attempted to quantify the environmental impact of increased artificial intelligence usage, but findings rely heavily on assumptions.
In Altman's blog post, he discussed the potential challenges, including entire job categories dissolving, while also highlighting the potential benefits, such as the possibility of unconventional policies like universal basic income, funded by productivity gains, becoming plausible in the future [6].
[1] OpenAI. (n.d.). Blog post. Available at: https://openai.com/blog[2] Altman, S. (2021). Blog post. Available at: https://www.samaltman.com/ai-democracy/[3] Goodman, N. (2020). Energy cost of training a large language model 2016-2020. arXiv preprint arXiv:2009.08692.[4] Saxena, T. (2022). How much electricity and water does it take to run AI research? A closer look at under-reported energy costs in the development of machine learning. arXiv preprint arXiv:2202.07111.[5] Fried, D. (2021). That AI paper you read may be using almost as much energy as a Tesla charging. The Washington Post. Available at: https://www.washingtonpost.com/technology/2021/08/04/energy-consumption-ai-models/[6] Altman, S. (2021). The world will get so much richer so fast that we'll have to rethink everything. The Guardian. Available at: https://www.theguardian.com/technology/2021/sep/06/the-world-will-get-so-much-richer-so-fast-that-we'll-have-to-rethink-everything
I'm not gonna let you go into environmental-science discussions without mentioning the surge in power demand due to artificial-intelligence. In fact, as Sam Altman, the CEO of OpenAI, highlighted in his blog post, the energy consumption of AI software like ChatGPT is akin to a high-efficiency lightbulb burning for a couple of minutes or a toaster using energy for around 13 seconds [4][5].
