The machine came from somewhere.

Exactly two sources attribute mineral demand to AI and data centers with a defensible denominator, and they agree. The IEA’s 2025 analysis puts data-centre copper demand near 512 kilotonnes by 2030, about 2% of global copper demand. The first peer-reviewed bottom-up model finds copper is 83% of the mineral mass AI data centers will require through 2035. Most of that copper goes into grid transmission and distribution, the wires that carry the power, rather than into anything inside a server. The supply-risk flags both sources raise are gallium, germanium, graphite, and rare-earth processing concentration, with over 90% of gallium refined in one country.

And the claim people reach for first has no credible source behind it. No defensible attribution connects AI demand to cobalt mining. Cobalt demand is driven by batteries, in vehicles and consumer electronics, and GPUs contain a negligible amount. The human-rights documentation from the DRC’s cobalt mines is real, serious, and about batteries; borrowing it for AI weakens both arguments.

Both attributions are models, one year old, with no supply-chain-traced data behind them. That the entire evidence base is two models is itself a finding, and it belongs in any sentence that quotes them.

TSMC, which fabricates nearly every advanced AI accelerator, used 25.55 TWh of electricity in 2024— roughly 9% of Taiwan’s total, with over 80% of the company’s own emissions coming from that electricity, and one projection reaching about 24% of the island’s power by 2030. Water withdrawal ran about 101 million cubic meters in 2023, and the company’s sustainability report states that its water-per-unit target was missed. During Taiwan’s 2021 drought, the worst since 1964, the government paid farmers to fallow fields while chipmakers trucked water in, 20 tons at a time. When fab demand met a hard hydrological limit, agriculture yielded.

Below the company totals, the public evidence comes from a model: imec’s open virtual fab, benchmarked with the major manufacturers, which finds lithography and etch account for about 45% of a fab’s direct and electricity emissions, and that per-wafer emissions riseat advanced nodes. The chemistry has an open end too: the first peer-reviewed measurement of PFAS in fab wastewater found more unidentified fluorinated compounds than identified ones (the researchers call the unknown fraction “dark PFAS”), and no regulatory inventory quantifies fab PFAS industry-wide.

Notice what is missing between those two paragraphs: any per-wafer environmental figure from an actual fab, at any advanced node. Companies disclose totals; models estimate steps; the quantity in between is published nowhere. Every “gallons per chip” number you have seen was built on that absence.

NVIDIA’s own product carbon footprint for the 8-GPU HGX H100 baseboard is widely cited at 1,312 kg CO2e — cradle-to-gate: manufacturing only, no use phase, no transport, no end-of-life, no surrounding server. Memory alone is 42% of the material impact. The generational comparison the company publishes shows embodied intensity falling 24% per exaflop from H100 to B200. That is real engineering progress, and a denominator chosen well: efficiency per exaflop improves while total exaflops shipped explodes, so the per-unit and fleet-wide stories move in opposite directions at once.

The academic anchor runs deeper on one older chip: the first cradle-to-grave, multi-criteria assessment of an A100 finds about 141 kg CO2eq over its life, with manufacturing dominating the impacts carbon accounting doesn’t see: 85% of mineral and metal depletion, 81% of freshwater eutrophication in its training case study. Google’s five-generation TPU assessment, built on first-party data, shows compute-carbon-intensity improving threefold across generations. And at fleet scale, Microsoft’s own report attributes its 30.9% Scope 3 rise since 2020 to datacenter construction and hardware.

The figure in circulation — AI producing 2.5 million tonnes of e-waste a year by 2030 — is real and published, and almost nobody quoting it states its conditions: the most aggressive of the paper’s four scenarios, large language models only, 3-year server lives assumed, extrapolated from demand. In 2026 a peer-reviewed recalibration modeled the same quantity from the supply side, anchored on TSMC’s actual chip-packaging capacity, and landed at 131–225 kilotonnes a year— seven to fourteen times lower, with realistic server lives of 4 to 6 years. Against the world’s 62 million tonnes of annual e-waste, the recalibrated estimate is a fraction of one percent, and even the aggressive scenario is about 3%.

Underneath the projection fight sits a question with four defensible answers: how long does the hardware last? A press claim of 1–3 years traces to an unnamed engineer and a misreading of Meta’s training logs, which record interruptions during one 54-day run — 419 of them, 30.1% GPU-related — not retirements. The same operators’ audited filings depreciate the hardware over 5–6 years, and a prominent investor argues even that flatters the assets. Failure, retirement, depreciation, and obsolescence are four different clocks. The water guide’s lesson holds here without modification: this is not a fact until someone says which clock.