Before delving into the topic of frugal AI, it’s important to set the stage. The unprecedented climate crisis we’re facing began with the Industrial Revolution in themid-19thcentury, which sowed the seeds of our current consumer society. Climate change isn’t the only environmental threat—there’s also water stress, resource depletion, and biodiversity loss—but it is undoubtedly the most visible and the best documented, and therefore the one that can help us better understand the others.
A sector that continues to grow at an ever-faster pace
The digital sector is difficult to pin down because its impact is widespread. According to ADEME, it accounted for 2.5% of France’s carbon emissions in 2022. In recent years, the sector has experienced strong growth, and forward-looking studies primarily project scenarios in which this growth will continue, at least in the medium term.
A quick calculation based on public data from the IPCC’s SSP1-19 scenario—one of the most optimistic—highlights the absurdity of this growth. If the sector grows according to the lowest growth forecast, the digital sector would emit six times more than the target set by the scenario for reducing global CO₂ emissions by 2050! Even if the sector’s growth were to stagnate at today’s level, it would account for three-quarters of total emissions… In such a world, what would be left for the rest of us?
If we focus on AI, we see a clear turning point starting in 2012. The sector’s growth then accelerated dramatically, with computing power requirements doubling every 5–6 months instead of every 24 months—a figure that had previously remained stable under Moore’s Law. This date coincides with the development of AI models based on deep learning, made possible by the use of graphics processing units (GPUs) to perform the calculations underlying deep learning and by the growth of open data on the Internet. It’s worth noting that AI is not limited to learning via deep neural networks, but these are undeniably the most resource-intensive.
A new milestone was reached in 2023 with the explosion of generative models such as the ChatGPT chatbot. Although it is difficult to provide precise figures—given that “tech giants” like OpenAI, Meta, and Microsoft, which are behind the largest models, no longer disclose this data—this widespread adoption is deeply concerning.
The Impact of Generative AI on the Climate
ChatGPT is based on the GPT-3 model, which has now been replaced by an improved version, GPT-4. It is not the only one, but it is the most popular and one for which data is available. The model on which it is based has 176 billion parameters and required 552 metric tons ofCO2 equivalent to train it in California. In terms of electricity consumption (a more objective metric in that it does not depend on the energy mix), the model ran for days on nearly 4,000 high-performance Nvidia GPUs, whose energy consumption was estimated at 1,283 MWh (megawatt-hours, or 1,000 kWh).
The usage phase consumes even more energy! Every day, the roughly 10 million users consume 564 MWh of electricity. Recent announcements by the CEOs of OpenAI and Microsoft regarding orders for hundreds of thousands of GPUs to power future versions are staggering in terms of energy consumption and environmental impact. With its current production capacity, Nvidia is far from being able to produce that many.
ChatGPT is just the tip of the iceberg in this vast landscape. Today, AI is a driving force behind the exponential growth of the digital sector, with an explosion in the number of applications and services that use generative AI. Of course, AI development at this pace is not sustainable as it stands.
How can we design more resource-efficient AI?
We will only be able to sustain this growth if AI enables significant emissions reductions across all other sectors. This is the prevailing view that AI will help us emerge from the crisis. Despite far too many useless or questionable applications, there are benefits for society—particularly in simulating and analyzing complex physical phenomena, such as studying scenarios to combat the climate crisis. But we must ensure that these solutions do not ultimately make things worse! For example, AI will enable companies that rely on fossil fuels to optimize their operations and thusemit even more CO₂.
Everywhere, we hear talk of “frugal AI” without the term being clearly defined. In everyday language, moderation is often understood as the appropriate response to excessive alcohol consumption. In the context of AI, however, it refers more to simplicity (which is clearly insufficient here), moderation, or even abstinence. Frugality and moderation are often considered synonyms; it is also possible to view frugality as pertaining to the functioning of technical systems, while moderation refers to their use within the context of social practices.
These two dimensions complement each other in that any technical system is designed for specific uses, which are thereby facilitated and encouraged. Thus, the more user-friendly the system appears, the greater its impact becomes: this is known as the rebound effect. However, the most relevant definition is the negative one: according to Le Robert, the opposite of frugality is gluttony. It is therefore possible to view frugality and moderation as a virtue that is measured in negative terms, based on the amount of resources that are not consumed.
However, defining “frugal AI” is difficult for several reasons. On the one hand, existing analyses often focus on model training and/or the usage phase, but overlook the full lifecycle of the service or product. This includes data production, use, and storage, as well as the hardware infrastructure deployed, from manufacturing through to the end of life of all equipment involved. On the other hand, for a service recognized as beneficial to society, it would be appropriate to estimate the volumes of data involved in the process and the indirect positive effects resulting from its deployment. For example, an energy-optimization system for an apartment can increase comfort or enable the deployment of new services thanks to the savings achieved.
Putting AI on a diet: a necessary but insufficient step
Today, the terms “frugality” and “simplicity” are often used interchangeably with “energy efficiency”: we conceive and develop a solution without considering its environmental cost, and then improve it from that perspective at a later stage. Instead, we should assess the effects early on, before rolling out the service—even if that means abandoning it.
Frugal AI is therefore characterized by an intrinsic contradiction, given the insatiable appetite for energy and data required today to train large models and put them to use, with little regard for the considerable risks to the environment. When it comes to AI, frugality must go far beyond mere efficiency: it must first and foremost be compatible with planetary boundaries. It must also question usage patterns early on—even to the point of eliminating certain services and practices—based on comprehensive and rigorous life-cycle analyses.
The goals underlying these technological developments should at the very least be debated collectively. Behind the argument of increased efficiency lies competition among national sovereignties or rivalry among companies seeking colossal profits. There is nothing about these goals that should not be evaluated through an ethical lens.
An evaluation of algorithmic systems using Contemporary Environmental Ethics even allows us to redefine the concept of moderation on a different basis. Indeed, despite their variety, these ethical frameworks do not view Nature (water, air, materials, and living beings) as resources available solely to the human species, which is engaged in technological competition and industrial hedonism. In conclusion, one could argue that a prospect is now opening up for responsible AI research that is as formidable as it is difficult to achieve: proposing models and systems that are as compatible as possible with such a “strong” definition of moderation.
