GenAI: Among the many aspects derived from the large-scale deployment of generative artificial intelligence, there is an important but underestimated one about which little is said. I am referring to the electrical energy that determines the use of the many software tools in this field, which makes it difficult to quantify precisely the energy consumption associated with the technology. Firstly because, this depends on several factors, starting with the machine learning model and the training process.

The other reason depends on the large companies involved in the development of AI models, such as OpenAI, Microsoft, Google and Meta, who chose silence on the amount of energy used in the development of their respective solutions when they realised that GenAI had become profitable and thus a primary element on which to base future business aims. In the past, OpenAI used to provide details on the hardware and timeframes it devoted to the training of the first models it developed. Still, as reported by several researchers, secrecy fell on ChatGpt and Gpt-4. 

It may be curious that tools designed to answer any question cannot give information on how much they consume, yet the chatbots of the companies mentioned above fail in this case. However, given how much sustainability is a primary factor in the public debate, despite the greenwashing behind which so many companies hide, it is impossible to continue ignoring that the amount of energy AI tools burn is a cause for concern, not least because the short- and medium-term future will be characterised by a steady increase in such consumption.

Generating images costs more than texts

In the absence of certainty, we have to make do with estimates for now, bearing in mind that if the training of an AI model is the phase that requires the most energy of all, the next phase must also be analysed: the use of models by people to generate output, which enables the various software that creates text, images and videos to achieve what is required.

The person who tried to understand how much an AI model can consume was Alexandra Sasha Luccioni, a researcher at Hugging Face. This French-American company develops AI-based apps. One of her studies with a team of researchers from Carnegie Mellon University offers an interesting insight into the relationship between AI and energy because they performed tests on 88 different AI models by repeating the prompts a thousand times to estimate the energy cost.

Thus, they found that text requires much less consumption than images: in the former case, it takes 0.002 kWh to process a text and 0.046 kWh to create one, while image-generating models use an average of 2.907 kWh per thousand attempts. For further comparison, the average smartphone consumes 0.012 kWh to complete a recharge, an amount of energy similar to what it takes to obtain an image using GenAI software. This data, albeit relative, also says that AI models need more energy to generate output than to classify input.

Classifying AI models for conscious choices

The focal point, however, is another, as Luccioni points out: ‘The generative AI revolution has a planetary cost that is completely unknown. How to solve this gap is the question that is being attempted to be answered. Alex de Vries, a PhD student at the VU in Amsterdam who has previously calculated the energy consumption of Bitcoin, used Nvidia’s GPU sales forecasts and energy specifications as the basis for the development of GenAI models, estimating that by 2027, the AI industry could consume between 85 and 134 terawatt-hours per year, with the latter figure representing the annual energy needs of the Netherlands.

Beyond the numbers, a concept on which de Vries’ study gives pause for thought is the relationship between improvements in the efficiency of AI models and the increase in demand and use of the same software by people. Since predicting whether the former will compensate for the latter is impossible, a potential remedy is to find a system to classify AI models. For Luccioni, it would be enough to replicate what has been done with household appliances.

A solution would allow people to be informed and thus make an informed choice of whether, when and for what purposes to use generative artificial intelligence. With the younger generations being much more sensitive to climate change and the need for more sustainable lifestyles, perhaps we could find more balance, consume less energy and reduce the damage we cause to the environment.