For the past year, projections of AI-driven e-waste have circulated through trade coverage and mainstream media, figures reaching as high as 5 million cumulative tons of AI-related scrap by 2030, sometimes framed as equivalent to discarding billions of smartphones.
A peer-reviewed paper published in Resources, Conservation and Recycling offers a more grounded view. Data scientist Alex de Vries-Gao, who has previously examined AI’s energy and water footprints, built a bottom-up model from actual chip production capacity and server lifespan data rather than extrapolating from computational demand forecasts. His conclusion: AI servers will likely generate between 131 and 224.8 kilotons of e-waste per year by 2030, roughly one-tenth of what earlier projections implied, and comparable to the total annual e-waste output of a country like Denmark or Norway.
Where earlier estimates went wrong
The Wang et al. projection, published in Nature Computational Science in 2024, derived server counts from estimated future computational demand, an approach De Vries-Gao notes is disconnected from what the supply chain can actually produce. By analyzing the capacity limits of TSMC’s chip-on-wafer-on-substrate packaging technology, which has been used for all leading AI chips in recent years, he was able to ground the estimate in verified production output rather than demand projections.
Server lifespan assumptions also drove the gap. Wang et al. used a three-year lifespan; De Vries-Gao found that more recent AI chip designs have a median useful lifespan closer to four years, with a reasonable assumption of four to six years for AI servers overall. Longer lifespans mean slower turnover and less waste per year.
A third factor: earlier estimates treated every server running any AI workload as dedicated AI hardware. In practice, much AI capacity runs on shared infrastructure, where workloads are allocated virtually rather than requiring discrete physical boxes.
The data gap
The paper’s sharpest finding may be its transparency critique. Basic information about AI-optimized server deployments — configurations, installed base, replacement cycles — is not publicly available. That makes independent verification difficult and leaves ITAD and recycling operators making capacity and capital decisions without reliable volume forecasts.
What’s ahead?
The revised numbers don’t eliminate AI hardware as a planning concern. At 131–224.8 kilotons annually by 2030, it is a distinct and growing stream that warrants separate intake and triage processes. In the current IT hardware environment, these additional tasks are not just tied to volumes, because GPU servers and AI-optimized storage have different security profiles, refresh cycles and remarketing paths than conventional enterprise IT.
AI-related hardware also remains a minority contributor to total global e-waste, projected at roughly 62 million tons annually this decade. The practical opportunity is to use AI hardware as a catalyst for modernizing tracking and recovery systems across all streams, rather than treating it as a category that requires an entirely separate response.
