50,000 tons of copper for an AI data center: The dark truth about the AI ​​boom

16-year wait: This unnoticed raw material shortage could burst the AI ​​bubble

Mountains of scrap metal and billions of liters of water: What the new AI infrastructure is really costing us

When tech giants wax lyrical about artificial intelligence, abstract terms like algorithms, parameters, and the cloud dominate. But the reality of AI is frighteningly physical. The industry is devouring unimaginable amounts of resources to build gigantic hyperscale data centers: tens of thousands of tons of copper and steel, billions of liters of drinking water, and rare technology metals that are pushing global supply chains to the brink of collapse. While the public debate mostly focuses on electricity consumption, a look behind the scenes reveals a much larger, strategically concealed material debt. From exploding commodity prices and intractable mining bottlenecks to a looming wave of e-waste, the AI ​​boom is proving to be one of the most aggressive and geopolitically explosive resource consumers in industrial history.

The AI ​​industry as a secret resource plunderer – What's really behind the billions in investments

When tech companies unveil their latest AI models, they talk about billions of parameters, training data, and the future of human civilization. The word copper is rarely mentioned. And even less frequently do we hear about the tens of thousands of tons of steel, the millions of cubic meters of concrete, the critical rare earth elements, or the accelerating e-waste problem that arises behind every new language model. The public debate is fixated on two narratives: energy consumption in kilowatt-hours and water consumption in liters. Both narratives are accurate, but incomplete. Because the physical material debt generated by the AI ​​boom is far more extensive, structurally entrenched, and geopolitically explosive than the usual sustainability reports from tech companies would suggest.

Copper as the new oil: Why 50,000 tons are just the beginning

The Copper Development Association has circulated a figure that still hasn't received the attention it deserves: A single hyperscale AI data center can consume up to 50,000 tons of copper. For comparison, a conventional data center uses between 5,000 and 15,000 tons. The jump isn't linear—it's a quantum leap. A single AI data center thus consumes more copper than three conventional facilities combined.

This number becomes real when you understand what copper is used for in a modern AI data center. The metal isn't a single component, but a ubiquitous material that permeates virtually every function of the facility. Power distribution, high-performance cables, transformers, busbars, connectors, cooling systems – all of it relies on copper. Nvidia's latest GB200 NVL72 unit alone contains over 5,000 copper cables with a total length of more than 3.2 kilometers. And the thermal design power of a single NVIDIA H100 chip is already 700 watts, which places extreme demands on heat dissipation – and thus on copper-based cooling systems.

For comparison, Microsoft's $500 million data center in Chicago alone required 2,177 tons of copper. This shows that even medium-sized projects already consume thousands of tons, while the largest AI facilities can actually reach the aforementioned 50,000 tons.

Copper is simply irreplaceable in its function. Only this metal can efficiently conduct heat to the outside of devices, and only copper offers the electrical conductivity required for power distribution in a high-performance data center. The investment bank Goldman Sachs aptly described copper as the oil of the AI ​​age – a formulation that is more economically precise than it initially sounds.

The consequences for the global copper market are significant. According to an analysis by BloombergNEF, copper demand from AI-powered data centers will average around 400,000 tons per year over the next decade, peaking at 572,000 tons in 2028. By 2035, the cumulative copper tied up in data centers could exceed 4.3 million tons. That's roughly the amount Chile—the world's largest copper producer—mines in six months. JP Morgan forecasts a global copper deficit of around 4 million tons by 2030, while S&P Global expects copper demand to rise by around 50 percent to 42 million tons by 2040.

Metal's price is soaring: How the AI ​​boom is reshaping the markets

The price of copper tells a story that most AI narratives overlook. In 2025, the price of copper on the London Metal Exchange surged by more than 43 percent—its best annual performance since 2009. At the beginning of 2026, the price broke through the $13,020 per ton mark for the first time before retreating to around $12,500. Goldman Sachs expects prices to remain permanently above $12,000 until the end of the decade.

The price drivers are multifaceted and mutually reinforcing. On the demand side, three major sectors are now competing for the same metal: the energy transition with electric vehicles and wind turbines, the expansion of power grids, and AI data centers. On the supply side, structural deficits are evident that cannot be remedied by any short-term investment. Mine disruptions in key producing countries such as Chile, Indonesia, and the Democratic Republic of Congo, a strike at the Mantoverde mine, and years of underinvestment have depleted the system's buffers.

The crucial structural bottleneck, however, lies not in the geology, but in time. From the discovery of a copper deposit to commercial production, an average of 16.2 years elapse. For a new copper mine, almost 12.4 years must first be spent on exploration and feasibility studies before any construction investment is even made. The consequence is brutally simple: The mines intended to meet the copper demand of 2030 should have been discovered as early as 2014 and financed by 2015. This did not happen.

At the same time, the trade policy dimension under the US tariff system distorts global copper flows. UBS analysts estimate that the US at one point held around half of the world's available copper stockpiles, even though the country accounts for less than ten percent of global copper demand. This market distortion drives up international premiums and exacerbates supply risks for Europe and Asia.

Steel, concrete and aluminum: The hidden building fabric of AI infrastructure

Copper is the most prominent, but by no means the only material that is fading into the shadow of AI narratives. Building a hyperscale data center is a massive industrial project requiring vast quantities of conventional building materials that don't appear in any tech pitch.

Steel is the backbone of every data center. It is needed for load-bearing structures, roof constructions, wall systems, equipment supports, and security infrastructure. Smaller data centers under 10,000 square meters already consume around 1,500 to 2,000 tons of steel and 10,000 cubic meters of concrete. For hyperscale facilities, which today reach capacities of 150 megawatts to well over one gigawatt, these figures multiply accordingly. In addition, the increased floor loads from heavy server racks—from the traditional 2.5 to 5 kilonewtons per square meter to the now required 12 to 15 kN/m²—necessitate thicker concrete slabs and reinforced steel structures.

A study commissioned by Greenpeace and conducted by the Öko-Institut (Institute for Applied Ecology) has determined that the expansion of AI-specific data centers alone will require approximately 920 kilotons of steel and around 100 kilotons of critical raw materials by 2030. Aluminum, also an essential material, is used in data centers for exterior cladding, HVAC systems, cable trays, and server enclosures, primarily due to its low density and corrosion resistance. Silver is used in server circuit boards and integrated circuits; tantalum, on which the US is 100% dependent on imports, is found in critical capacitors; platinum and palladium are used in semiconductors.

Concrete is known for its disproportionately high carbon footprint: According to the UN, the construction industry is responsible for 38 percent of global CO₂ emissions, and concrete alone accounts for eight percent of global greenhouse gases. The construction phase of a data center generates significant amounts of so-called embodied carbon, meaning CO₂ that is produced not during operation, but during material extraction, transport, and construction. These emissions are often not reported, or only partially reported, in operators' sustainability reports because regulatory reporting has historically focused on operations.

The water paradox: Three billion liters per plant per year

While the water consumption of AI data centers has entered the public debate, it is still grossly underestimated. A single 100-megawatt data center can require around 2.5 billion liters of water per year – depending on the cooling technology and location. Large data centers can consume up to 19 million liters of water per day, according to estimates by Allianz Commercial, which is equivalent to the daily consumption of a city with up to 50,000 inhabitants.

The cooling mechanism is crucial for understanding the water problem. With the widespread use of evaporative cooling towers, between 70 and 85 percent of the water used simply evaporates into the atmosphere. This water is irretrievably lost to the local water cycle. When Google and Microsoft were preparing their large language models in 2021 and 2022, both companies recorded increases in their water consumption of 34 and 20 percent annually, respectively. Google's data centers consumed around 20 billion liters of water in 2022 – roughly equivalent to the annual consumption of 2.5 million Europeans.

According to a study by the University of California and the University of Texas, training OpenAI's GPT-3 model required approximately 5.4 million liters of water. Of this, 700,000 liters were used for cooling the data centers alone, while the remainder was consumed in the supply chain for server manufacturing and power generation. A British government analysis estimates the additional, AI-driven global water demand by 2027 at between 4.2 and 6.6 billion cubic meters. The Öko-Institut (Institute for Applied Ecology) predicts that the water demand of data centers will almost quadruple to 664 billion liters by 2030.

Microsoft has unveiled a new data center design that uses no water for cooling and, according to the company, saves more than 125 million liters of water per year per facility. This innovation is commendable, but still far from setting the global standard. The vast majority of AI infrastructure being built worldwide relies on conventional evaporative cooling – particularly in regions where water is still readily available but already under ecological strain.

Rare earths and technology metals: The invisible Achilles' heel

Besides commodity raw materials like copper, steel, and aluminum, there is a second, strategically even more critical layer of materials: rare earths and technology metals. Without gallium, there are no high-performance LEDs or high-frequency chips. Without indium, there are no touchscreens or 5G antennas. Without germanium, there are no modern semiconductors. Without tantalum, there are no miniaturized capacitors. Without neodymium and dysprosium, there are no high-performance permanent magnets for cooling fans and pumps.

All these metals have one thing in common: China controls their global supply to an extent unmatched by any other raw material supply chain. When China placed gallium and germanium exports under control in August 2023, prices skyrocketed within weeks. Since the beginning of 2025, there has even been a complete export ban on heavy rare earth elements. For the Western AI industry, this represents a structural dependency that cannot be resolved in the short term through any diversification strategy.

Technology metals like gallium and indium are often only produced as byproducts in the extraction of other raw materials. This means that even if the price rises and demand increases, production cannot simply be ramped up. It is tied to the primary production of the respective main metal. This inelasticity on the supply side is a structural characteristic of the technology metals market that significantly exacerbates the risks of an AI-driven demand spike.

The geopolitical dimension is further exacerbated by the fact that supply routes for critical raw materials are increasingly exposed to geopolitical disruptions. According to the UN, eleven percent of all global trade flows through the Strait of Hormuz – a route that transports strategic raw materials for chip manufacturing and which has recently come under considerable pressure due to the Iran conflict. Disruptions to these corridors not only increase transport costs but also force insurers to drastically raise war risk premiums.

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

E-waste: The trillion-ton time bomb in the AI ​​lifecycle

One issue that never appears in the glossy brochures of AI companies is the dramatically short lifespan of the hardware they use. Analysts predict that most AI processors will be technically obsolete after three to five years because the development cycles of chips and AI accelerators involve a significant performance leap every 12 to 18 months. This not only means that billions of dollars in investments lose value in just a few years, but also that the raw materials used in their construction end up in an extremely short recycling cycle—a cycle for which the global recycling infrastructure is not designed.

A study by the Chinese Academy of Sciences, published in the journal Nature Computational Science, estimates that the cumulative e-waste from LLM hardware alone will reach up to 9 million tons worldwide by 2030 in conservative scenarios. In a scenario with rapidly increasing user adoption, this figure could be around 2.5 million tons per year by 2030. For comparison, total global e-waste amounted to approximately 62 million tons in 2022. AI data centers add a new, previously almost nonexistent component to this stream.

The Öko-Institut warns that the expansion of data centers and AI capacities will generate up to five million tons of additional electronic waste by 2030. This scrap contains valuable materials such as copper, gold, silver, cobalt, and rare earth elements, which could theoretically be recovered. In practice, however, both the technical capacity and the economic incentives for comprehensive recycling are lacking. Many of these devices end up in informal recycling facilities in the Global South, where the extraction of valuable metals takes place under hazardous conditions.

The hidden cost structure: What an AI data center really costs

When the industry discusses the costs of AI data centers, it typically cites figures like five to twenty billion dollars per large facility. What's regularly missing is an honest full-cost accounting that includes all direct and indirect resource costs.

Copper is estimated to account for up to six percent of a data center's capital costs. For a $10 billion project, that would equate to $600 million for copper alone. With copper prices now exceeding $12,000 per ton and a requirement of 50,000 tons, this results in a copper cost of approximately $600 million per facility – and rising, because copper prices are under structural upward pressure. Every percentage point increase in the price of copper drives up the construction costs of a hyperscale data center by millions.

Added to this are the costs of grid expansion. The energy demands of data centers have already prompted several governments to take drastic measures. In the US, President Trump mandated in March 2026 that tech companies like Google, Microsoft, Amazon, Meta, and OpenAI sign a Ratepayer Protection Pledge, requiring them to bear the full costs of new power plants and grid expansion themselves. While this model offers short-term protection to residential electricity customers, it shifts infrastructure costs into the companies' operating expenses and thus into the prices of their services. At the end of 2025, Ireland enacted stringent regulations requiring new data centers to operate their own battery storage or power plants and cover at least 80 percent of their electricity needs with newly installed renewable energy sources.

Allianz Commercial's projections are sobering: estimates predict that spending on AI infrastructure will reach approximately seven trillion US dollars by 2030. To justify these investments, consumers and businesses would need to invest around 800 billion US dollars in AI products, according to calculations by the Wall Street Journal – and this over the entire lifespan of the data centers currently under construction. At the same time, the industrial insurer Allianz Commercial anticipates that tight schedules, a shortage of skilled workers, and skyrocketing raw material prices are increasingly jeopardizing these construction projects.

The ecological debt of mining: Who pays the price in the Global South?

The discussion about AI's resource consumption usually ends where the supply chain becomes opaque: at the mine. However, copper mining in the major producing countries of Chile and Peru is anything but a neutral process.

In Chile, the world's largest copper producer, mining leads to massive water consumption in the Atacama Desert, one of the driest regions on Earth. The open-pit mining process and subsequent smelting cause significant soil and air pollution, as well as profound disruption to local ecosystems. In Peru, research by the organization Facing Finance has shown that German copper imports are demonstrably linked to human rights violations: instead of promised improvements in living conditions, social and environmental conflicts are plaguing the mining regions. These external costs do not appear in any tech companies' balance sheets. They are borne by the affected populations.

The mining industry itself faces a fundamental capacity problem. Mining experts speak of a supply gap of up to ten million tons of copper by 2040 – roughly equivalent to Chile's current annual production. Declining ore grades in new deposits, rising development costs, longer permitting processes, and increasing resistance from affected communities are further lengthening the already extremely long lead times. A new copper mine discovered today could not begin production until 2042 at the earliest. This is not a technical weakness – it is the physical reality of an industry designed for decades to come, which is now encountering a demand curve that is exponential, not linear.

Land use: The invisible footprint of AI infrastructure

Another rarely discussed aspect of AI's hunger for resources is land consumption. Hyperscale data centers today no longer require just a few hectares, but often hundreds of hectares of land – for the server buildings themselves, but also for power supply, cooling infrastructure, backup systems, and the associated power distribution and substations. The demand for suitable sites near stable power grids and sufficient water supplies is already driving up real estate prices in traditional data center regions like Virginia, Amsterdam, and Frankfurt.

According to McKinsey, 200-megawatt systems are no longer uncommon, and projects exceeding one gigawatt are actively being planned. Power density per server rack has increased from an average of eight kilowatts in 2022 to 17 kilowatts for AI-enabled racks in 2024 – and this trend is continuing. The implications of this for space requirements and infrastructure planning are not yet sufficiently addressed by regulations in most regions.

In Virginia alone, the largest data center location in the US, demand for network capacity is expected to rise to 12.1 gigawatts by 2025 – an increase of almost 30 percent compared to the previous year. In the state, one in four kilowatt-hours already goes toward cooling and operating the digital infrastructure. In Germany and Europe, the planning and approval processes for large-scale infrastructure projects represent a separate bottleneck: It often takes seven to twelve years for new substations and high-voltage power lines to be approved, built, and commissioned.

Carbon footprint of construction: What nobody wants to measure

The sustainability reports of major tech companies focus with remarkable consistency on one key metric: the PUE (Power Usage Effectiveness) value, i.e., the ratio of total electricity consumption to IT electricity consumption. A low PUE is considered an indicator of technological efficiency. What this metric fails to capture is the so-called embodied carbon – the embedded CO₂ footprint generated during the extraction of raw materials, their processing, transportation, and the construction of the facility.

As power grids become increasingly decarbonized and the operational carbon footprint of a data center decreases accordingly, the relative share of embedded carbon in the overall balance is growing. For the next generation of data centers, which are intended to be powered by renewable electricity, embodied carbon could already account for half or more of the total lifecycle emissions. This consequence has so far barely registered in the public debate.

The Öko-Institut (Institute for Applied Ecology) has calculated that CO₂ emissions from data centers will rise from 212 million tons in 2023 to 355 million tons in 2030 – despite the assumed massive expansion of renewable energies. In the US, 55 percent of the electricity used for data centers is still generated from fossil fuels such as coal and natural gas. As long as this remains the case, every new AI data center that goes into operation means not only an increased demand for copper, steel, and water, but also a direct rise in CO₂ emissions – with all the associated consequential costs for society, health, and the climate system, which also do not appear in the balance sheets of tech companies.

Structural conclusions: The costs of invisibility

What conclusions can be drawn from this analysis? First, a sobering observation: The narrative of AI as a primarily digital, intangible technology is a myth. AI is one of the most material-intensive technology investments in human history. It consumes copper, steel, concrete, aluminum, rare earth elements, and water in quantities that dwarf any other past technology boom.

The key economic question is: Who bears these costs? Currently, allocation follows the principle of maximum externalization. Mining companies and the communities they affect bear the environmental and social costs of raw material extraction. Municipalities and grid operators bear the costs of the overburdened infrastructure. Future generations bear the costs of climate change and electronic waste. And taxpayers in democratic societies subsidize grid expansion, which would not be necessary on this scale without the AI ​​boom.

The market failure is structural. Copper prices, construction costs, and energy prices internalize a growing share of real costs, but environmental damage in Chile, human rights violations in Peru, and long-term climate costs remain unpriced. Without a full-cost accounting system that incorporates these externalities, the AI ​​industry operates with effectively subsidized access to raw materials—at the expense of those without bargaining power.

The second conclusion concerns the strategic implications for Europe and Germany. Copper, gallium, germanium, indium, and rare earth elements are raw materials for which Europe is almost entirely dependent on imports. The AI ​​boom exacerbates this dependence and increases geopolitical vulnerability. China has demonstrated its willingness and ability to use export controls as a tool of foreign policy pressure. Europe lacks an adequate response to this.

The third conclusion is perhaps the most important: the pace of AI infrastructure expansion and the pace of raw material development are fundamentally incompatible. AI data centers are built in two to five years. New copper mines take 16 years. New rare earth projects take even longer. The market will close this gap through the price mechanism—by rising raw material prices, rising construction costs, and ultimately, rising prices for AI services. Who will ultimately bear these costs is still undecided. What is clear, however, is that the bill will be substantial.

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here simply call me at +49 7348 4088 965. My email address is [email protected]:or

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs