Generative artificial intelligence (AI) manages your social media feed, studies your bank loan application, drafts legislation and gives you directions.
It will also soon generate millions of tons of waste.
Tech companies are struggling with truckloads of computer parts as they race to build bigger data centers to train better generative AI models — algorithms that use training data to create new content.
A new study published today in Nature Computational Science estimates that generative AI could generate up to 2.3 million tons of electronic waste (also called e-waste) per year by 2030.
To put that into context, that amounts to throwing away 13.3 billion iPhone 15 Pro phones.
“We were surprised by the scale of the expected electronic waste,” said Asaf Tzachor, co-author of the study and researcher at Israel’s Reichman University.
“To our knowledge, this is the first comprehensive study to quantify and project e-waste generated specifically by generative AI technologies.”
Previous studies have focused on AI energy consumption (generating an image uses about the same amount of energy as fully charging your phone) and water consumption (a ChatGPT conversation with 50 questions can use about half a liter).
Generative AI produced approximately 2,400 tons of electronic waste in 2023.
This relatively small figure will skyrocket as massive data centers built to develop products like ChatGPT are stripped of computer components and then replaced with new designs.
Meanwhile, larger and larger data centers are coming online, with $36 billion invested in AI infrastructure worldwide by 2023.
These will add thousands of tons to the e-waste stream.
Microsoft’s rush to build data centers like this one in Wyoming has led to a spike in carbon emissions. (Supplied: Microsoft)
But it’s not all bad news.
The authors found that extending the life of existing computing infrastructure, reusing reusable components, and recycling valuable materials such as copper and gold could reduce electronic waste generation by as much as 86 percent.
Most AI hardware is replaced every three years
While generative AI is widely available on smartphones and desktop computers, the physical infrastructure is largely out of sight, behind the walls of data centers.
The contents of these buildings are well-kept secrets.
To calculate the amount of electronic waste generated by these data centers, the researchers first estimated the amount of computing power that would be needed to train generative AI models under different adoption scenarios.
They then translated this demand into hardware requirements, estimating the number of servers and electronic circuits such as graphics processing units (GPUs).
This may look high-tech, but even the very latest components are outdated within three years. (Supplied: Microsoft)
They then calculated how technological improvements would reduce hardware requirements.
“Given the typical lifespan of hardware – often around three years due to technological advances – we predicted when this hardware would be retired,” said Dr. Tzachor.
“This approach allowed us to estimate the amount of e-waste generated over time.”
In the “aggressive adoption” scenario, where AI would be used by all internet users, 4.5 million tons of electronic waste would be produced between 2020 and 2030.
Most of the e-waste is said to be clustered in Europe, North America and East Asia, where most of the world’s data centers are located.
Relatively simple ‘circular economy strategies’ would have a big impact on e-waste production, the researchers found.
About 2.8 million tons of outdated AI servers could be avoided by extending their lifespan by one year.
Much of the world’s electronic waste is shipped to countries like Kenya, where it ends up in landfills. (Getty Images: James Wakibia/SOPA Images/LightRocket)
Dismantling, refurbishing and reassembling outdated modules such as GPUs so they can be used for other, less intensive forms of computing could reduce electronic waste by 42 percent.
“To our knowledge, the widespread adoption of these practices in industry is still limited,” said Dr. Tzachor.
“Challenges such as data security concerns and the need for high-quality hardware can make reuse and recycling more complex.”
Some measures even make the e-waste problem worse.
US restrictions on sales of high-end GPUs to countries like China are forcing data centers to use outdated server models, resulting in more e-waste.
A one-year delay in getting the latest chips could result in a 14 percent increase in e-waste, the researchers calculated.
The old toasters are piling up
Generative AI e-waste is responsible for a small but increasing share of global e-waste, which is defined as any product with a plug or a battery.
E-waste is now the fastest growing waste stream in the world. According to the UN’s global e-waste monitoring report, a record 62 million tons of electronic waste was generated in 2022.
This figure was 82 percent higher than in 2010 and is on track to rise a further 32 percent to 82 million tonnes by 2030.
The report states:
The 62 million tons of e-waste generated in 2022 would fill 1.55 million 40-ton trucks, roughly enough trucks to form a bumper-to-bumper line around the equator.
Small electronic equipment such as phones, microwaves and vapes make up the largest category of electronic waste by mass.
Less than a quarter of global electronic waste is formally collected and recycled. In Australia, about half of electronic waste is sent to recyclers.
Australians rank fourth in the world in per capita e-waste generation, generating 22 kilograms of e-waste per person per year, more than double the global average.
A recent report shows that most of Australia’s top companies are not including e-waste in their sustainability reports.
‘We are running out of resources’
Damien Giurco, from the University of Technology Sydney’s Institute for Sustainable Futures, said the study’s e-waste projections appeared accurate.
“We need producer responsibility and product stewardship,” said Professor Giurco, who was not involved in the study.
“When you put equipment on the market, you have to be sure [there’s] a process that supports reuse of materials and recycling.”
Nvidia’s recently launched Blackwell platform for AI data centers weighs 1.2 tons. (Supplied: Nvidia)
Companies that import electronic goods or build data centers in Australia are currently not legally required to account for the amount of electronic waste they generate.
That could change if steps are taken to introduce a circular economy framework. A federal government ministerial advisory group released an interim report earlier this year recommending national and sectoral targets for the circular economy and specific circularity requirements for products.
Ministerial advisory group member and CEO of research and advocacy group Circular Australia, Lisa McLean, said tech companies could be held responsible for their e-waste.
“There is more gold and silver in a ton of iPhones than a ton of ore from a gold or silver mine.”
Corporate net-zero commitments also push companies to consider the carbon emissions in their products, or the amount of greenhouse gases associated with a product’s lifespan.
“I think businesses understand that this is an economic response to the crises we face,” Ms McLean said.
“We are running out of resources and we will not reach net zero without a circular economy.”
