As AI expands rapidly, data centres around the world are scaling up at pace. That growth also creates electronic waste (e-waste). In new research, data scientist Alex de Vries-Gao estimates that AI servers could generate 131–224.8 kilotonnes of e-waste per year by 2030, a lower figure than previously assumed.
Global AI use is rising fast. De Vries-Gao has previously examined what this means for energy demand and for the technology’s global CO₂ and water footprint. In his latest study, published in Resources, Conservation and Recycling, he focuses on another consequence of AI’s growth: the volume of electronic waste associated with the hardware. One of his key findings is that AI servers could produce around 131–224.8 kilotonnes of waste annually in 2030. “That is comparable to the total annual e-waste generated by countries such as Denmark, Norway or Austria. So we are talking about very substantial waste streams,” the researcher says.
Lower than earlier projections
De Vries-Gao’s analysis responds to an earlier study that suggested total AI-related e-waste could rise to five million tonnes by 2030. The VU researcher arrives at a lower estimate by first assessing how many AI servers can realistically be manufactured, a constraint earlier projections largely did not account for. Based on currently available data, he also argues that the commonly assumed three-year lifespan of this equipment may be too pessimistic, and that servers are likely to remain in use for longer in practice. As a result, projected waste volumes are considerably lower than in earlier estimates.
A caveat
A lower estimate is not a reason for complacency, De Vries-Gao stresses. “AI still generates an enormous amount of waste. That’s why it remains crucial to extend the use of these materials, invest in high-quality recycling, and improve data collection so we can map the problem more accurately,” he says. The major companies behind data centres, he adds, have a responsibility here: “Without greater transparency, there is a risk that the scale and real impact will be underestimated. There is still a lot to improve.”
