The mass the accelerator does not carry
The Basel Action Network builds a reference 100 MW AI facility at about 7000 metric tons across five equipment classes: networking, power distribution, backup power, servers plus accelerators, and cooling. In the same text, servers and GPUs are 13 percent by weight. Copper in cabling and busbars at a 100 MW site is about 2700 tonnes on its own.[1]
On that denominator a GPU-focused waste estimate sees about one seventh of the facility mass. BAN writes that adding power, networking and cooling can put the total 40 to 60 times above the most cited academic projections. The range it produces for 2025 to 2050 is 395 million to 617 million tonnes.[1]
The teardown assumption carries the watt
The figure is tied to an assumption that power infrastructure is ripped out when racks move from a 5 kW to 15 kW band to a 50 kW to 140 kW AI band. Accelerator turnover is taken as two to three years against five to seven years for general-purpose servers; the model assumes the whole server is replaced when the Nvidia GPU generation is upgraded. That is a ladder from announced capacity expansion to waste mass, not a delivered decommissioning certificate.[1]
Another reading is that copper busbars and cooling towers outlast one GPU generation and that teardown stays partial. BAN's text takes planned buildout and fast turnover as inputs; it does not show in how many published 100 MW sites that cycle has already run.[1]
The mass line to watch
Announced megawatts still sit on the capacity rung of this ladder. If an operator in the 100 MW class publishes separate tonnages for cabling, busbars, cooling and accelerators by the end of 2026, the 7000-tonne reference drops to a site measurement; if it does not, the figure stays a waste counterpart of planned watts.[1]