What we'll cover
Every discussion of AI's environmental impact tends to focus on electricity and carbon emissions. But before a single query is ever run, a long and complex supply chain has already extracted materials from the earth, refined them using energy-intensive processes, fabricated them into chips in some of the most technologically demanding manufacturing facilities humans have ever built, and shipped them across the world.
This session examines the upstream environmental and social costs of AI: the critical minerals that make AI hardware possible, where they come from, what extracting them costs, and the geopolitical tensions that surround their supply. These are questions that receive far less attention than energy consumption, yet they connect AI directly to some of the world's most pressing issues around environmental justice, human rights, and global supply chain governance.
Optional readings
IEA (2021): "The Role of Critical Minerals in Clean Energy Transitions": the fullest institutional analysis of critical minerals demand, supply geography, and policy options, and the best single source for supply concentration data. Technical but authoritative.
Amnesty International (2016): "This Is What We Die For: Human Rights Abuses in the DRC Cobalt Supply Chain": foundational investigative report on cobalt supply chain conditions. Conditions have partly improved since publication but the report remains an important reference.
Siddharth Kara, "Cobalt Red: How the Blood of the Congo Powers Our Lives" (2023): a journalist and academic's account of artisanal cobalt mining in the DRC. Compelling and accessible, though note it is an advocacy-oriented narrative rather than an academic study.
Global E-waste Monitor: published by the UN Institute for Training and Research; the authoritative source for global e-waste statistics.
Responsible Minerals Initiative: industry-led organisation running the Responsible Minerals Assurance Process (RMAP) for smelter auditing. Useful for understanding what "responsible sourcing" claims mean in practice.
UK Critical Minerals Intelligence Centre (British Geological Survey): publishes regular assessments of criticality and supply risk for individual minerals.
Defining critical minerals
The term "critical minerals" has a specific meaning in policy contexts: it does not simply mean important or useful. A mineral is typically classified as critical when it meets two criteria simultaneously.
The two criteria
- Economically essential: The mineral is required for technologies or industries that are central to the modern economy: in particular, digital technology, defence, and the clean energy transition
- Supply-concentrated: Production or processing is dominated by a small number of countries or companies, creating significant risk of disruption
A mineral can be physically abundant in the earth's crust and still be "critical" if its processing is controlled by a single country. Gallium is a good example: not particularly rare, but roughly 99% of primary (low-purity) production comes from China (USGS Mineral Commodity Summaries, 2025).
Two overlapping categories for AI
AI infrastructure requires minerals from two overlapping categories:
- Electronics-critical minerals: Used directly in chip fabrication and electronic components: silicon, gallium, germanium, hafnium, tantalum, copper, gold, indium
- Energy-transition minerals: Used in batteries (for uninterruptible power supplies, data centre backup, cooling systems) and in the clean energy infrastructure powering data centres: cobalt, lithium, nickel, rare earth elements
The AI hardware boom is driving demand in both categories simultaneously, compounding pressure that already exists from the clean energy transition.
From mine to data centre: the supply chain
Let's trace some of the key materials that go into modern AI infrastructure, from the GPU chips to the servers to the data centre cooling and power systems.
Key minerals in AI infrastructure
| Mineral | Where used in AI infrastructure | Top producers | Key risk |
|---|---|---|---|
| Silicon | The base of all semiconductor chips; wafers on which transistors are built | China (~75% of polysilicon), Japan (wafers) | Refining energy-intensive; Chinese dominance in polysilicon |
| Gallium | Gallium nitride (GaN) used in power electronics and some compound semiconductors | China (~99% of primary production, USGS 2025) | Chinese export controls announced July 2023 (effective August 2023) |
| Germanium | Optical fibre, some semiconductor processes, infrared optics in data centres | China (~60%), Canada | Chinese export controls announced July 2023 (effective August 2023) |
| Hafnium | High-k dielectric layers in modern transistors (sub-5nm chips, including NVIDIA H100/H200) | France, USA, Russia (as by-product of zirconium) | Niche supply; disruption from geopolitical tension |
| Copper | Chip interconnects, PCB traces, power delivery, cooling pipes, data centre wiring | Chile (~27%), Peru, DRC, China | Demand surge from AI + electrification; long mine development timelines |
| Cobalt | Lithium-cobalt-oxide batteries (UPS systems, laptop batteries); some chip manufacturing chemistries | DRC (~70%) | Artisanal mining, child labour, environmental damage |
| Rare earth elements | Permanent magnets in server fans, cooling pumps, hard drives; also in EV motors and wind turbines | China (~60% mining, ~85% processing) | Near-monopoly on processing; radioactive waste from mining |
| Tantalum | Capacitors in virtually all electronic devices including servers and networking equipment | DRC (~40%), Rwanda, Australia | Conflict mineral history; artisanal mining conditions |
| Indium | Indium tin oxide in displays, some thin-film processes | China (~55%), South Korea, Japan | No primary mining; recovered only as by-product of zinc refining |
The complexity of chip fabrication
Modern AI chips like the NVIDIA H100 are among the most complex objects humans manufacture. A single chip contains billions of transistors at feature sizes of 4–5 nanometres (roughly 10–15 atoms wide), built through hundreds of sequential fabrication steps, each requiring ultrapure materials and tightly controlled chemical processes. The slightest contamination at any step can ruin the chip.
This is why semiconductor manufacturing is so geographically concentrated: TSMC in Taiwan fabricates the chips for NVIDIA, Apple, AMD, and many others. There is essentially no substitute facility for the most advanced nodes. A single disruption (whether from geopolitical tension, natural disaster, or energy shortfall) would affect global AI capacity.
Geographic concentration and geopolitical risk
The physical supply chain for AI hardware is more geographically concentrated, and therefore more fragile, than most people realise.
China's dominant position
China occupies a uniquely powerful position in the critical minerals supply chain for AI, through its dominance of processing and refining. Even where minerals are mined elsewhere, they are often shipped to China for processing before being used in manufacturing.
Key figures:
- ~85% of global rare earth processing
- ~99% of primary gallium production (USGS 2025)
- ~60% of germanium production
- ~75% of polysilicon refining (the raw material for silicon wafers)
- Majority of lithium chemical processing, even though most lithium is mined in Australia and South America
In 2023, China imposed export controls on gallium and germanium, both classified as critical for advanced semiconductor manufacturing. This was widely understood as a response to US restrictions on advanced chip exports to China, and a signal that critical minerals are a lever in geopolitical competition.
Case study: the export-control arc, 2023–2026
The 2023 controls turned out to be the opening move in a rapid escalation that shows how a supply chokepoint behaves once it is actually used:
- December 2024: one day after a new round of US chip-export rules, China banned exports of gallium, germanium, antimony and superhard materials to the United States outright, moving from licensing friction to outright prohibition for the first time
- April 2025: in response to sweeping new US tariffs, China placed seven medium and heavy rare earth elements, and the permanent magnets made from them, under export licensing, applied worldwide rather than to the United States alone
- October 2025: the controls were expanded to five further rare earths, to refining and magnet-making technology, and, for the first time, to foreign-made products containing even small amounts of Chinese-origin rare-earth content
- November 2025: within weeks, after a US–China leaders' meeting, China suspended the December 2024 ban and the October expansion for one year; the April 2025 licensing regime stayed in force
The suspension is as informative as the controls. A restriction that can be paused for a year and reinstated on notice works as a dial: the materials flow, but each flow depends on licensing decisions in Beijing, and the leverage survives the pause. For AI hardware, whose power electronics, magnets and chips sit directly on these supply chains, mineral supply has become a live geopolitical variable.
Case study: the TSMC concentration problem
Taiwan Semiconductor Manufacturing Company (TSMC) fabricates the overwhelming majority of the world's most advanced chips, including essentially all of the NVIDIA GPUs that power AI infrastructure. In 2024, TSMC held approximately:
- ~90% market share at the most advanced nodes (≤5nm)
- ~64% of global semiconductor foundry revenue overall (2024)
Taiwan sits 160 km from mainland China, which claims it as its own territory. The concentration of advanced chip manufacturing in a geopolitically contested location is considered by many security analysts to be one of the most significant systemic risks to the global economy. The US CHIPS Act (2022) and EU Chips Act (2023) are both partly responses to this risk, investing in domestic semiconductor manufacturing capacity to reduce dependence on Taiwan.
TSMC itself is investing in new fabs in Arizona, Japan, and Germany, though these facilities remain years from full production and are unlikely to match the leading edge of TSMC's Taiwan operations for some time.
Other supply concentrations
| Country / region | Dominant position | Relevance to AI |
|---|---|---|
| Democratic Republic of Congo | ~70% of global cobalt mining | Cobalt in batteries for UPS and mobile devices; tantalum in capacitors |
| Chile, Argentina, Bolivia ("Lithium Triangle") | ~55% of global lithium reserves | Lithium-ion batteries for data centre backup power and cooling |
| Taiwan | ~90% of leading-edge chip fabrication | All advanced AI chips (NVIDIA, AMD, Apple) manufactured here |
| Netherlands (ASML) | 100% of extreme ultraviolet (EUV) lithography machines | EUV machines are required to manufacture chips at ≤7nm; ASML is the only manufacturer |
| Japan | Dominant in semiconductor chemicals and some specialty materials | Photoresists, chemical mechanical planarisation slurries, and other process chemicals used in chip fabs |
Environmental and social costs of mining
The concentration of mining in particular countries and regions means that the environmental and social costs of AI hardware are borne unevenly, and often by communities in the Global South that have little connection to the AI systems being built.
Cobalt: the DRC case
The Democratic Republic of Congo produces roughly 70% of the world's cobalt, much of it through artisanal and small-scale mining (ASM): informal operations using hand tools rather than industrial equipment.
- Child labour: Investigations by Amnesty International and others have documented children working in cobalt mines in Katanga province, in conditions that expose them to toxic dust and cause serious health impacts
- Occupational health: Artisanal miners lack protective equipment; cobalt dust causes lung disease ("hard metal lung disease")
- Environmental damage: Mining runoff contaminates local water supplies and agricultural land
- Traceability gap: Artisanal cobalt often enters the supply chain through intermediaries that make tracing its origin difficult for downstream technology companies
Lithium: the water problem
The Lithium Triangle (Chile, Argentina, Bolivia) holds more than half the world's lithium reserves, concentrated in salt flats (salares) in the Atacama Desert, one of the driest places on earth.
- Brine extraction: Lithium is extracted by pumping lithium-rich brine to the surface and evaporating it in large ponds; this process consumes large volumes of water in an extremely arid region
- Indigenous communities: Many salares are in or near territories of indigenous communities (Atacameño, Aymara, Quechua) who rely on local water sources for subsistence farming and livestock; water diversion threatens these livelihoods
- Ecosystem impacts: The salares are fragile ecosystems supporting flamingo populations and endemic species; the brine ponds disrupt these habitats
- Governance: Bolivia's large reserves remain largely undeveloped partly due to disputes over sovereignty and benefit-sharing
Rare earth mining: the radioactivity problem
Rare earth elements (neodymium, dysprosium, and others used in the magnets in server fans, hard drives, and cooling pumps) are typically found in geological associations with radioactive thorium and uranium.
- Radioactive waste: Mining and processing rare earths generates radioactive tailings, waste material that must be carefully managed to prevent contamination
- China's environmental legacy: Decades of poorly regulated rare earth mining in Jiangxi province have left significant environmental damage: stripped hillsides, contaminated rivers, and legacy radioactive waste
- Lynas (Australia/Malaysia): Australia's Lynas Rare Earths processes ore in Malaysia; the facility has faced protests over concerns about radioactive waste disposal
- Why processing matters: Even "clean" mining in countries with strict environmental standards often sends ore to China for processing, where environmental regulations may be weaker
On discussing environmental justice
The environmental and social costs of critical mineral extraction are real and documented, but this area is also one where media coverage can be selective or sensationalised. When engaging with specific claims about conditions in mining regions, try to trace them to primary sources: peer-reviewed research, investigative journalism from established outlets (The Guardian, Bloomberg, Reuters), or reports from organisations like Amnesty International or the Business & Human Rights Resource Centre rather than secondary summaries.
The end of the line: e-waste
The critical minerals story does not end when hardware is manufactured. What happens when AI servers, GPUs, and data centre equipment reach the end of their lives is a significant and growing problem.
The scale of the problem
- The Global E-waste Monitor estimates that ~62 million tonnes of e-waste was generated globally in 2022, the equivalent of roughly 107,000 jumbo jets
- Only about 22% of global e-waste is formally collected and recycled; the rest is either landfilled, incinerated, or handled informally
- AI's rapid hardware upgrade cycles (driven by the competitive pressure to deploy the latest GPUs) accelerate this waste stream
- A GPU generation that was state-of-the-art in 2022 (A100) may be decommissioned for newer H100s or H200s within 2–3 years, even when physically functional
Where e-waste goes
- Much of the world's e-waste (including from data centres in wealthy countries) is exported to informal recycling sites in West Africa (Ghana's Agbogbloshie), South Asia (Pakistan, India), and Southeast Asia
- Informal e-waste recycling involves burning cables to recover copper, using acid baths to extract gold, and other processes that release toxic substances (lead, mercury, dioxins) and expose workers to serious health risks
- This trade is technically illegal under the Basel Convention, but enforcement is inconsistent
- The valuable materials (gold, copper, palladium) are recovered; the hazardous residues are often left in the local environment
The recycling gap
Despite the high value of materials in AI hardware, formal recycling rates for critical minerals from e-waste are very low:
- Rare earth elements: <1% end-of-life recycling rate globally
- Cobalt: Higher recycling rates from lithium-ion batteries (>50% in some jurisdictions), but recovery from non-battery e-waste is low
- Indium: <1% end-of-life recycling rate
- Gold: Higher rates (~15–20% from electronics) but significant losses in informal processing
The recycling gap is partly a technology problem (recovering materials from complex multi-layer chips is difficult) and partly an economics problem (virgin mineral extraction has historically been cheaper than urban mining from e-waste).
Responses: policy, industry, and research
The risks around critical minerals supply chains have prompted responses at multiple levels. These are worth knowing, both as context and as a basis for evaluating claims made by technology companies and governments.
Policy responses
- US CHIPS and Science Act (2022): $52 billion for domestic semiconductor manufacturing; restrictions on companies receiving funds from expanding in "countries of concern" (primarily China)
- EU Critical Raw Materials Act (2023): Sets targets for domestic mining, processing, and recycling of critical minerals by 2030; reduces dependence on single-country sourcing to max 65% for any mineral
- Minerals Security Partnership: US-led coalition of 14 countries (including Australia, Canada, UK, EU) coordinating on critical mineral supply chains
- Export controls: US has restricted exports of advanced AI chips to China; China has responded with export controls on gallium, germanium, and graphite, escalating through the 2024–26 arc described above
Industry responses
- Responsible sourcing programs: Most major technology companies have cobalt and conflict minerals policies; the Responsible Minerals Initiative (RMI) runs a widely-used smelter audit program
- Recycling investment: Apple has invested in its Daisy robot that disassembles iPhones for material recovery; broader e-waste recycling programs are expanding
- Supply chain diversification: Companies are attempting to source from multiple countries and producers rather than relying on single sources
- Battery chemistry shifts: Lithium iron phosphate (LFP) batteries, which do not use cobalt, are gaining market share, reducing cobalt demand in some applications
Research and alternative materials
- Cobalt-free battery chemistries: LFP, sodium-ion, and other alternatives are being developed and deployed to reduce dependence on cobalt
- Rare-earth-free magnets: Research into iron-nitride and other materials that could replace neodymium-iron-boron magnets in motors and actuators
- Gallium nitride (GaN) alternatives: Silicon carbide (SiC) can substitute in some power electronics applications with a different sourcing profile
- Urban mining technology: Improved hydrometallurgy and biotechnology processes for recovering rare earths and other metals from e-waste; still largely pre-commercial at scale
The "green" energy transition and critical minerals
One of the more complex ironies in this space: the clean energy transition that AI companies point to when claiming to reduce their carbon footprint also requires large quantities of the same critical minerals as AI hardware. Electric vehicles need cobalt and lithium; wind turbines need rare earth magnets; solar panels need indium and other minerals.
This means that AI infrastructure growth and clean energy growth are competing for the same mineral supply, and that the environmental and social costs of critical mineral extraction are simultaneously relevant to both. Policies that aim to decarbonise electricity also increase pressure on critical mineral supply chains in ways that must be managed carefully.
Summary and key takeaways
The physical supply chain for AI infrastructure extends far beyond data centres and electricity bills:
- AI hardware requires a wide range of critical minerals: from silicon and copper (abundant but processing-intensive) to gallium, germanium, hafnium, cobalt, and rare earth elements (scarce or concentrated)
- Geographic concentration creates systemic risk: China dominates processing for most of these materials; Taiwan is the sole manufacturer of leading-edge chips; the DRC produces the majority of cobalt
- Export controls are already being used as geopolitical leverage: China's 2023 restrictions on gallium and germanium escalated to an outright ban on US-bound exports in December 2024 and worldwide rare-earth licensing in 2025, then a one-year suspension negotiated in November 2025; supply access is now a bargaining chip that can be dialled up and down
- Mining carries environmental and human costs that are borne unequally: Cobalt mining in the DRC, lithium extraction in the Atacama, and rare earth processing in China all involve significant environmental and social impacts, largely in communities far removed from AI's benefits
- E-waste is a growing problem: Rapid GPU upgrade cycles accelerate hardware disposal; current formal recycling rates for most critical minerals are very low
- Policy and industry responses exist but are insufficient to the scale of the challenge: The CHIPS Act, EU Critical Raw Materials Act, and responsible sourcing programs are real steps, but critical mineral demand from both AI and the clean energy transition is growing faster than supply chains are diversifying
Next (4.4): We bring the session together with practical guidance: tools for measuring your own AI footprint, the choices that make the most difference, and AI's potential as a tool for addressing environmental challenges rather than just creating them.