Evidence is growing that the carbon footprint of artificial intelligence (AI) is on the rise. However, experts claim that the technology is already adapting and will soon even be helping humanity battle climate change.
Big data, machine learning, AI – these are bywords for everyone who has even the slightest idea of what goes in tech these days. OpenAI’s ChatGPT, the viral AI-generated chatbot, is used by more than 100 million people worldwide.
Yes, some worry AI will take away their jobs, others even raise concerns over the future of humanity – in an open letter in late March, scores of tech leaders and AI researchers warned there was risk of losing control of our civilization. Overall, a “bigger is better” attitude has been adopted, as thousands of new machines are built every day.
But the most urgent problem seems to be the environmental cost of computation. After all, each new chatbot or image generator is built with ever more electricity, and this obviously implies that the practice is to blame for a significant quantity of carbon emissions that contribute to global warming.
Cybernews chatted to industry experts to find out what ideas to mitigate the issue are being floated around. For most, though, the problem is also part of the solution – tech companies allegedly grasp the need to use energy more efficiently and are already making huge strides towards achieving that.
Energy costs rising
“As an artificial intelligence language model, I don’t have a direct carbon footprint because I don’t consume energy or produce emissions in the same way that humans or machines do.” This is how ChatGPT answers the question about its carbon footprint.
But it continues: “The servers and data centers that host my computing infrastructure consume a significant amount of energy and produce carbon emissions.” And that is precisely the key to the enigma.
Already in 2019, the Massachusetts Institute of Technology’s bimonthly magazine Technology Review reported that the cloud where computing takes place had a larger carbon footprint than the entire airline industry and generated around 2% of global emissions.
To illustrate, a single data center might consume an amount of electricity equivalent to 50,000 homes. And AI training actually consumes more energy than traditional types of computing – just one model’s training can burn more electricity in a year than 100 US homes.
The Technology Review also said that training just one AI model equates to emitting more than 626 pounds of carbon dioxide – nearly five times the lifetime emissions of an average American car.
Google's researchers discovered that AI accounted for 10-15% of the tech company's overall electricity consumption, Bloomberg reported recently.
Because the latter was measured at 18.3 terawatt hours in 2021, that would imply that Google's AI uses around 2.3 terawatt hours of energy yearly, assuming a mean average of 12.5% taken from the above range.
That is comparable to the electricity used by all the residences in, say, Singapore, where almost six million people live.
The models – which are becoming bigger, by the way – need retraining, too. If you leave your model unmodified for a year or two, it will probably have no idea about, say, COVID-19 or Russia’s war in Ukraine. This, again, costs a lot of energy.
Just like mining Bitcoin: the New York Times just published a story about how much electricity these trillions of calculations per second use, and how the public usually pays the price.
Pollution accelerating
The process is quite straightforward: Microsoft, Google, and, for that matter, OpenAI use cloud computing, which depends on myriads of servers inside huge data centers, to train AI algorithms.
These data centers, though, need to be cooled, usually by energy-intensive air conditioning units. Some simply blow air around the entire building, other firms experiment with floor planning or “hot” and “cold” aisles of server racks. Regardless of which method is adopted, a lot of energy is needed.
Most experts Cybernews has reached out to – now and earlier – recognize the problem and stress it is vital to fundamentally understand that profitability and environmental impact are not mutually exclusive.
“The increasing use of digital solutions and AI is providing ever more insights and benefits for our world and businesses, helping us be more productive, efficient, and sustainable,” Justin Bean, an author and speaker focusing on sustainability in tech, told Cybernews.
“But data itself has a carbon footprint – data processing, analysis, and storage are part of the emissions problem. Decarbonization of data centers is a key climate priority as we try to convert dirty data into clean data.”
Chris Noble, CEO and co-founder of Cirrus Nexus, a cloud-management platform helping firms measure and reduce carbon emissions from cloud operations, agrees: “As energy demand for IT doubles every four years, bringing online clean, reliable, quality energy will become even more challenging for energy providers.”
“The solutions will require a comprehensive strategy encompassing government regulations, incentive programs, public education, industry investment, commercial adoption, and changes in how organizations operate IT environments,” Noble told Cybernews.
Sasha Luccioni, a researcher at Hugging Face, a platform where users share pre-trained AI models, datasets and demos, went one step further. Last year, she tested how much energy the training of the firm’s own large-language model, BLOOM, required.
The answer was 25 metric tons of carbon dioxide emissions. However, Luccioni also estimated emissions by OpenAI’s GPT-3 – it has allegedly released more than 500 tons of carbon dioxide.
For a single model which Chat-GPT is based on, it’s a lot. 500 tons of carbon dioxide emissions is the equivalent of around 600 flights between London and New York – and ChatGPT-4 is already available, too.
“We certainly have to be careful of the value created by the output of AI as it relates to the impact on the planet. If the AI is doing something basic but throwing a lot of computing power and energy resources at the problem, does that make sense for the planet?” Vince Lynch, an AI expert, told Cybernews.
Available solutions
Luckily, there’s room for improvement – and some companies are already implementing measures that allow them to save energy. For example, Microsoft, Google, and Amazon, the three largest cloud providers in the US, have all made commitments to be carbon neutral or negative.
It’s telling that Google, which aims to run its offices and data centers completely on carbon-free energy by 2030, has deployed none other than AI to control cooling in the buildings. Big Tech firms are also purchasing renewable energy.
Some of the biggest machine learning jobs might be relocated to more carbon-friendly world regions. A good example is the city of Montreal in Canada, as a number of data centers there run on hydro-electricity. Solar or wind power is also perfectly fine.
Companies should also consider simply using cloud-based services that allow AI workloads to be distributed across multiple data centers. This would reduce the energy consumption of any single facility.
“AI can only fix the climate if it fixes itself first. Since AI is likely to see its use grow exponentially, it’s all the more necessary to scale responsibly – that is, to use data centers that run on low-carbon electricity.”
Alexis Normand.
According to Alexis Normand, CEO and co-founder of Greenly, a carbon accounting platform, a proven way for AI developers to reduce emissions is to be more flexible and smarter in timing the training and maintenance sessions. For instance, training at off-peak hours when energy is less expensive and more abundant reduces the demand during peak times.
“Less energy consumption is registered within data centers by running models on more efficient processors at off-peak hours to minimize workload,” Normand told Cybernews, before also agreeing a greener grid would be helpful.
“AI can only fix the climate if it fixes itself first. Since AI is likely to see its use grow exponentially, it’s all the more necessary to scale responsibly – that is, to use data centers that run on low-carbon electricity.”
Some labs have also successfully cut down the size of their AI models. In February, for example, Meta has released LLaMA, a large-language model several times smaller than the ones OpenAI is building.
Finally, less is more. Phil Tee, CEO and co-founder of Moogsoft, an AI provider, thinks that businesses should only store data the machine actually needs continuously.
“Far too many tech vendors choose to take a ‘more the merrier’ approach to data storage. A culture based on excess has left technicians clinging to every piece of data, even when said data no longer holds significance. It is high time to eliminate what companies no longer need by focusing not on data at rest, but data in motion,” Tee told Cybernews.
Your email address will not be published. Required fields are markedmarked