New Delhi: The electronic waste produced by artificial intelligence is not limited to data centres and is expected to be 40-60% higher than previous estimates, according to a new study by the global NGO Basel Action Network.
In its report titled The Coming AI Waste Wave, the US-based organisation estimated that the world will generate 30 to 46 million tonnes of e-waste a year from AI alone, largely due to the hyperscaling of hardware and electrical equipment that the AI boom has led to.
“Until now, environmental scrutiny of AI data centres has focused almost entirely on carbon footprint and freshwater consumption,” said the report. “The e-waste impacts, its likely toxicity, and its environmental justice implications have to date largely been ignored.”
Basel Action Network’s findings say that around $7 trillion is expected to be invested globally between 2025 and 2030 just for building AI infrastructure, to add around 219 GW of data centre capacity by 2030. To put it in perspective, one GW of data centre capacity requires 70,000 tonnes of electromechanical infrastructure, which includes servers, power distribution networks, storage, cooling and accelerators.
These are also part of the ‘hidden’ cost of AI, according to the report. Most earlier estimates looked just at data centres, specifically at servers and accelerators, to calculate AI waste. However, these form only 13% of the total electronic infrastructure in a data centre.
When the entire ecosystem is considered, the study said, estimates of e-waste from the AI wave increase by 40-60%.
Why additional AI waste?
AI data centres are driving additional e-waste because, despite rapid expansion, AI has led to the breakdown of Moore’s Law. First coined in 1965 by Intel co-founder Gordon Moore, the law stated that the computing power of a chip would improve roughly every two years, leading to more powerful computers without increasing the average physical hardware required. This was known as vertical scaling.
In the case of AI, however, the technology is moving too fast to sustain vertical scaling. Instead, most companies are expanding computing power by literally adding more chips, leading to increased electromechanical infrastructure, and thus, more waste.
“Rather than each generation of a single chip getting faster (vertical scaling), AI infrastructure scales by deploying incomprehensible quantities of these specialised chips working in parallel (horizontal scaling),” said the report.
The report also explained another phenomenon that is unique to AI cloud computing — the “cattle vs pets” perspective on servers used by AI companies. The phrase was coined by Bill Baker, a cloud engineer at Microsoft. Baker explained how earlier, servers used to be considered as pets by technology companies; and if one malfunctioned, it would be fixed and brought back to its original shape.
However, in the era of AI and cloud computing, servers are considered cattle; if one malfunctions, it is simply replaced by another one, due to the sheer volume of servers that are necessary for today’s computing requirements. This is also the phrase that can help explain the e-waste generation from AI and a departure from a ‘circular economy’ principle.
Currently, there are 12,259 data centres operating across the globe. Of these, 5,388 are in the United States. The report pointed out that the US has an additional 4,871 data centres in the pipeline, even as Europe and Asia are also looking to expand data centre facilities by 2030.
Each of these data centres is expected to generate waste in the form of transformers, power cables, lithium-ion batteries, coolant, and other hardware. Most of this, according to the report, is built to go obsolete or require replacement in 2.5 to 5 years, and can also be categorised as hazardous waste, especially because most of the data centre waste will not be sorted according to categories.
“This paper estimates that by 2050 we will be generating roughly between 196 – 211 million tonnes of e-waste each year,” said the report. “This is roughly a tripling of today’s annual global e-waste generation.”
(Edited by Aakriti Handa)