MetallKIDD was a research project at the Wuppertal Institute, funded by the Federal Environment Ministry from 2021 to 2023, with Fraunhofer IOSB, the Folkwang University of the Arts and the VDI Technology Center. It looked at how AI increases the demand for technology metals, which are rarely recycled.

Together with Vera Austrup and Manuel Bickel I built a tool that estimates this demand. It breaks an AI system down to the teraflops one use requires and compares them with the teraflops of the hardware running the inference. It distinguishes between inference on the device and at hyperscalers, and it includes efficiency gains and the replacement rates of hyperscaler hardware. The example in the file is BingGPT on a smartphone.

According to the tool, the phones of AI users could take over about 10 percent of inference in 2023 and 36 percent by 2050, if each one gives a tenth of its compute. A decentralised setup like this could only be run by companies with control over the hardware, such as Apple or Microsoft. Background inference would also reduce battery life. Data protection would be possible, since only individual layers are passed between devices, but inference would be considerably slower.

The tool also calculates the energy demand. The number of AI users is expected to grow, while hardware and software become more efficient. In the tool, the efficiency rate stands for the hardware that is actually saved each year. The scenarios show how much the result depends on it.

The tool only considers chip performance in teraflops and petaflops. Memory is not included, although the production capacity for RAM is currently a limiting factor.

It is published as version 0. The values are our best estimates and can be replaced with your own. Instead of one fixed scenario, it shows how different settings of the variables lead to different results.

Update 2026

Since 2023 the measurements have moved further than the assumptions. Google puts the median text prompt to Gemini at 0.24 Wh [1], a peer-reviewed study by Microsoft researchers puts large models at 0.31 Wh [2]. The 3.71 Wh in the tool are about twelve times that. They match long reasoning queries instead, for which the same study finds 3.91 Wh [2]. Models with reasoning use on average 30 times the energy of models without [3]. Today the tool would have to count tokens rather than queries.

The memory gap has become the main constraint. An eight-GPU server in the tool carries 640 GB of high-bandwidth memory. Current NVIDIA systems carry 1.44 TB and 2.1 TB [4]. In NVIDIA’s own footprint of a B200 board, memory accounts for 49 percent of the emissions from manufacturing [5]. The IEA expects the shortage of this memory to last until at least the end of 2027 [6], and contract prices for ordinary DRAM nearly doubled in the first quarter of 2026 [7].

Recalculated with the tool’s energy method and only the values above, the result for Germany in 2026 depends almost entirely on the share of reasoning queries. With plain queries the model needs about 8,900 GPUs and 59 GWh a year, less than in 2023 with almost three times the users. With reasoning queries only, it needs 112,000 GPUs and 740 GWh.

Data centres used 415 TWh of electricity in 2024 [8] and 485 TWh in 2025. The IEA expects 950 TWh by 2030, around 3 percent of global electricity demand. AI-focused data centres grew by 50 percent in 2025 alone [6]. In Germany data centres used 21.3 TWh in 2025 [9].

China produces 99 percent of the world’s primary gallium [10]. By 2030 data centres could need up to a tenth of today’s supply [11]. How much e-waste AI hardware will leave behind is disputed: one study expects 1.2 to 5 million tonnes in total by 2030 [12], a later recalculation 131 to 225 kilotonnes a year in 2030 [13].

All fifteen metals in the tool are on the EU’s 2023 list of critical raw materials except indium [17]. For nine of them, one country supplies more than half of world production, China for gallium, magnesium, neodymium, silicon metal and tungsten. Recycling covers little: the EU puts the recycled share of input at zero for gallium, beryllium, strontium and silicon metal, and only tungsten reaches 42 percent [17]. Some of the tool’s recycling rates are higher than the EU’s, for titanium 91 against 1 percent. Measured against EU demand, the scenario stays at or below about one percent for every metal, for one service in one country.

Phones now run language models of about three billion parameters themselves [14]. The split between device and cloud has replaced the idea of phones lending compute to a shared network. The renewal rate still holds: hyperscalers write their servers off over five to six years, which is 17 to 20 percent a year. Amazon shortened part of that again in 2025, citing the pace of AI [15]. About a third of people in the EU aged 16 to 74 used generative AI in 2025 [16]; the tool started from 20 percent.

Rebuilt as a script, version 0 reproduces exactly, but errors in it move the main result. One of the four methods divides by the GPUs a model instance needs instead of counting them, and phone compute starts growing two years early. Corrected, server metal rises by 15 percent, to 27.2 tonnes by 2050; the figures above are the corrected ones. The circuit board, which carries three quarters of the server metal, is entered with 4,000 g and has no documented source; with the 265 g from the documentation, the result would halve.

Which of the four methods is used changes the result by a factor of 275. Across 10,000 runs over the documented ranges of all inputs, the median is 3.1 tonnes, and 90 percent of the runs lie between 0.07 and 23.6 tonnes. The corrected default sits near the top, because it combines a low efficiency gain, a high energy per query and the heavier board. The working paper and the code are linked below.

A version 1 would count tokens instead of queries and treat the share of reasoning as a variable of its own. It would take memory into account next to compute, use current hardware with its weight and cooling, and calculate energy as a time series like the metals. Phones would be modelled as devices running their own small models rather than as part of a shared network. It would also fix errors in version 0: palladium is recycled at the rate for neodymium, one criticality column multiplies by a power figure in gigawatts, and several recycling rates lie well above the EU’s.

Sources

1. Amin Vahdat, Jeff Dean: Measuring the environmental impact of AI inference. Google Cloud, August 2025.
2. Oviedo et al.: Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling. Joule, 2026.
3. Sasha Luccioni, Boris Gamazaychikov: AI Energy Score v2. Hugging Face, December 2025.
4. NVIDIA: DGX B200 and DGX B300, product pages.
5. NVIDIA: PCF Summary for NVIDIA HGX B200. July 2025.
6. IEA: Key Questions on Energy and AI. April 2026.
7. TrendForce: DRAM revenue, first quarter 2026. June 2026.
8. IEA: Energy and AI. April 2025.
9. Bitkom, Borderstep: Rechenzentren in Deutschland. November 2025.
10. USGS: Mineral Commodity Summaries 2026, Gallium.
11. IEA: Energy and AI, AI and energy security. 2025.
12. Peng Wang et al.: E-waste challenges of generative artificial intelligence. Nature Computational Science, 2024.
13. Alex de Vries-Gao: Recalibrating global artificial intelligence e-waste estimates. Resources, Conservation and Recycling, 2026.
14. Apple: Updates to Apple’s On-Device and Server Foundation Language Models. July 2025.
15. Annual reports (Form 10-K) of Microsoft, Alphabet, Meta and Amazon, 2022 to 2026, SEC EDGAR.
16. Eurostat: Use of generative AI tools in 2025. December 2025.
17. Milan Grohol, Constanze Veeh: Study on the Critical Raw Materials for the EU 2023. European Commission, 2023.