Behind the apparent immateriality of artificial intelligence lies a vast industrial infrastructure. Semiconductors, data centers, power grids, cooling systems, copper, water, concrete and global supply chains form the physical foundation of a technology often presented as purely digital. As AI expands, the question of computational power is increasingly becoming a question of the resources required to sustain it.**

Artificial intelligence appears to weigh nothing. A question is typed onto a screen, a few seconds pass, and an answer appears. No engine is visible, no smokestack emits anything, and no material seems to have been consumed. The simplicity of the interface reinforces the image of an almost abstract technology, produced somewhere inside a “cloud” whose very name suggests an absence of physical substance.

The industrial reality is almost exactly the opposite.

Every artificial intelligence model relies on an accumulation of highly sophisticated machines, themselves installed inside industrial buildings, supplied by power grids, cooled by thermal systems and connected through thousands of kilometers of fiber and cables. Their production begins long before they reach the data center, in mines, metal refineries, chemical plants, semiconductor foundries and international logistics networks.

Artificial intelligence therefore belongs entirely to the material world. Its development may even become one of the principal drivers of a new global wave of industrial investment.

I. Behind the Software, an Industrial Machine

At the physical core of contemporary artificial intelligence are computing accelerators, primarily GPUs and other specialized circuits capable of performing enormous numbers of mathematical operations simultaneously.

The sophistication of these components provides an initial indication of the industrial density required by AI. NVIDIA’s Blackwell architecture, for example, incorporates 208 billion transistors across its GPUs and relies on a specialized TSMC manufacturing process. Behind a single chip therefore lies an extraordinarily complex industrial chain: electronic design, design software, lithography, silicon wafers, ultra-pure chemicals, manufacturing equipment, high-bandwidth memory, advanced packaging, substrates and cooling systems.

This chain is also highly concentrated geographically. A significant share of the world’s advanced semiconductors is manufactured in East Asia, while several critical technologies required to produce them are controlled by a limited number of companies located in the United States, Europe, Japan, South Korea and Taiwan.

AI is therefore not dependent solely on algorithms. It depends on a global industrial architecture in which several stages have become genuine strategic chokepoints.

This material dimension becomes even more visible when the chips leave the factories and enter data centers.

II. The Data Center Becomes an Energy Infrastructure

Data centers existed long before artificial intelligence. The internet, e-commerce, banking services, video platforms and cloud computing already depended on immense computing capacity.

AI, however, is changing their trajectory.

The training and operation of contemporary models make intensive use of accelerated servers. According to the International Energy Agency, global electricity consumption by data centers could reach approximately 945 TWh by 2030, more than twice its 2024 level and slightly below 3% of projected global electricity consumption. Between 2024 and 2030, their electricity demand could grow by around 15% annually in the agency’s base case, more than four times faster than electricity consumption from all other uses combined.

AI accounts for a major part of this acceleration. The IEA estimates that electricity consumption from accelerated servers, driven primarily by AI adoption, could increase by approximately 30% per year through 2030. These systems could account for almost half of the net increase in global data-center electricity consumption over the period.

These figures nevertheless require careful interpretation. Not every data center is dedicated to AI, the energy efficiency of chips is improving rapidly, and forecasts depend heavily on the pace at which the technology is adopted. The IEA itself emphasizes that even consumption approaching 945 TWh would still represent less than 3% of total global electricity demand.

The challenge is therefore geographical as much as global.

A data center is not connected to a “global electricity grid.” It connects to a specific local network. The concentration of several facilities in the same region can consequently impose a substantial load on available generation, transmission and distribution capacity.

In the United States, the Department of Energy estimated in late 2024 that electricity demand from data centers could double or triple by 2028, driven in part by AI.

The question then changes fundamentally. It is no longer simply how much electricity AI consumes, but where that electricity can be produced, transported and guaranteed around the clock.

III. The Battle for the Megawatt

This constraint helps explain a striking transformation within the technology industry: digital companies are becoming increasingly involved in energy.

Access to abundant, stable and predictable electricity is becoming a determining factor in the location of computing infrastructure. Grid-connection delays, transformer availability, transmission capacity and access to new generation can now determine the timetable of a computing project worth several billion dollars.

Microsoft stated in its 2025 environmental sustainability report that it had contracted 34 GW of new renewable energy capacity across 24 countries.

The trend extends far beyond renewable energy alone. Future data-center requirements are contributing to renewed interest in nuclear power, gas-fired generation, electricity storage, new grid infrastructure and, over the longer term, small modular reactors and other energy technologies still under development.

The digital economy is therefore encountering a challenge long familiar to aluminum, steel, chemicals and other electricity-intensive industries: energy availability is becoming a strategic factor of production.

There is, however, an important difference. Data centers can concentrate considerable economic power within a relatively small physical area. Their electricity consumption can consequently increase extremely rapidly within particular regions, while building a power plant, high-voltage transmission line or electrical substation can take several years.

The speed of software is colliding with the slow pace of infrastructure.

IV. An Intelligence That Produces Heat

Almost all electricity consumed by a processor eventually dissipates as heat. As accelerators become more powerful and are grouped into increasingly dense systems, removing that heat becomes a central engineering challenge.

Cooling is therefore inseparable from AI infrastructure.

Traditional systems have relied extensively on cooled-air circulation. But the increasing thermal density of accelerated servers is encouraging the development of liquid cooling, particularly systems that deliver cooling directly to components.

This transition can substantially improve thermal efficiency, but it introduces another resource into the equation: water.

It would nevertheless be misleading to assign a fixed quantity of water to every query sent to an AI system. Requirements vary according to the model, hardware, location, outside temperature, cooling technology and electricity mix. Some of the water associated with computation is consumed directly at the data center, while another portion is consumed indirectly through electricity generation.

Technologies are also evolving rapidly. Microsoft, for example, states that more than 90% of its data-center capacity now uses closed-loop liquid cooling systems that continuously reuse water, and that some direct-to-chip cooling systems can save more than 125 million liters of water per facility each year compared with the architectures they replace.

The water question therefore does not disappear, but neither can it be reduced to a universal formula in which “one query equals X liters.” Its significance depends primarily on location and technology.

A facility operating in a water-rich region does not create the same constraints as one located in an area already exposed to water stress. The material footprint of AI must therefore be analyzed territorially as well as globally.

V. Before Electricity Come the Metals

The environmental discussion surrounding AI frequently concentrates on electricity consumption. Yet electricity represents only the operational component of its footprint.

Buildings must be constructed, servers manufactured, electronic boards assembled, transformers installed, data centers connected to power grids and machines interconnected.

That requires concrete, steel, aluminum, copper, silicon, plastics, optical fiber and numerous specialized materials.

Copper occupies a particularly important position. It is used throughout electrical equipment, cables, transformers, electronic systems and the broader expansion of power grids. AI therefore represents another source of demand for a metal already under pressure from transport electrification, renewable-energy deployment and electricity-grid modernization. The International Energy Agency has repeatedly highlighted the substantial copper and aluminum requirements associated with the expansion of electrical networks.

Semiconductors meanwhile require their own materials and industrial processes. Their production involves extremely pure silicon wafers, multiple metals, industrial gases, specialized chemicals and large quantities of ultrapure water.

The computer running an AI model is therefore merely the visible endpoint of a much longer extractive and industrial chain.

UN Trade and Development has emphasized precisely this dimension of the digital economy: behind dematerialized services lie raw materials, physical equipment and increasing volumes of electronic waste. It also points out that environmental costs associated with extracting and processing resources are often borne by developing economies that capture a much smaller share of the value created further downstream.

This geography produces a fundamental asymmetry: the places where digital services are consumed are not necessarily those where their material footprint is greatest.

VI. A Deeply Geopolitical Value Chain

This material dependence transforms artificial intelligence into a geopolitical issue.

Computing power requires simultaneous control over several industrial layers: chip design, advanced semiconductor manufacturing, memory, packaging, lithography equipment, materials, energy, electricity grids, data centers and telecommunications infrastructure.

Few countries control the entire chain.

The United States occupies a dominant position in several processor-design segments and AI platforms. Taiwan plays a major role in advanced semiconductor manufacturing. South Korea is central to memory production. Japan remains important in several industrial materials and manufacturing technologies. Critical lithography and chip-production technologies depend on European suppliers. China, meanwhile, possesses immense industrial capacity, a vast electricity infrastructure and significant positions in the processing of numerous raw materials.

Access to computing power is consequently becoming an attribute of power comparable, in certain respects, to access to energy or industrial capacity.

The OECD now explicitly discusses national compute capacity, emphasizing that national artificial intelligence strategies must be aligned with available computing resources.

Algorithmic sovereignty without material capacity therefore remains limited. Possessing researchers and developing models may not be sufficient if processors must be imported, data centers cannot be adequately powered or network infrastructure is insufficient.

AI is thus forcefully reintroducing geography into a digital world that had long claimed to transcend it.

VII. The Efficiency Paradox

It would nevertheless be incorrect to assume that the consumption of an artificial intelligence system necessarily increases in proportion to its performance.

Processors are becoming more efficient. Model architectures are evolving. The numerical precision required for certain operations is decreasing. Software is optimizing hardware utilization. Data centers are improving their cooling systems, and the same computational tasks can progressively require less energy.

In theory, these advances should reduce the footprint required for each unit of computation.

At the system level, however, they can produce the opposite result.

This is the classic mechanism of the rebound effect: when a resource becomes more efficient and less expensive to use, total consumption can increase because its applications multiply.

Cheaper image generation allows more images to be generated. More efficient inference enables AI to be incorporated into more software. Falling model costs make it possible to process more documents, videos, scientific data and everyday queries.

Improving the energy efficiency of an individual query therefore provides no guarantee that overall consumption will decline.

The essential question becomes the relationship between two rates of change: the rate at which computing efficiency improves and the rate at which demand for computation grows.

For now, IEA projections suggest that the latter will remain sufficiently high to produce a substantial increase in the overall electricity consumption of data centers.

VIII. The Material Cost Must Be Compared With What AI Replaces

The material footprint of artificial intelligence cannot, however, be examined solely as an isolated cost.

A technology can consume significant resources while enabling even greater savings elsewhere.

AI can help optimize electricity grids, improve industrial processes, predict equipment failures, reduce certain forms of travel, optimize logistics, accelerate research into new materials and improve building energy management. The IEA itself considers both dimensions simultaneously: artificial intelligence increases energy demand, but it may also improve the efficiency and operation of energy systems.

Assessing its actual environmental impact therefore requires comparisons between complete systems.

If a digital simulation eliminates the need for thousands of physical experiments, its environmental balance cannot be reduced to the electricity consumed by the data center. If a model improves the energy efficiency of a factory, the resulting savings must be included. Conversely, if AI primarily produces an explosion of new digital activities that did not previously exist, its footprint largely adds to existing consumption.

The OECD similarly recommends distinguishing between the direct environmental impacts of computation and equipment and the indirect effects, both positive and negative, generated by applications. It also highlights persistent gaps in measurement and transparency.

A serious debate therefore consists neither of presenting AI as inherently sustainable nor of estimating its footprint from a handful of spectacular examples. It requires measuring what AI consumes, what it replaces and what it genuinely enables society to save.

IX. AI as a New Heavy Industry

The expression may appear paradoxical. Yet as computing infrastructure expands, artificial intelligence is acquiring several characteristics traditionally associated with capital-intensive industries.

It requires enormous upfront investment. It depends on energy infrastructure. It mobilizes global raw-material supply chains. It requires complex industrial facilities. It depends on specialized suppliers that are difficult to replace. It faces constraints involving land, electricity, water and regulation.

The data center is gradually becoming less comparable to a conventional IT building and more comparable to an industrial installation connected to a wider energy ecosystem.

This transformation could have significant economic consequences.

For several decades, part of the digital economy benefited from an exceptional characteristic: the extremely low marginal cost of reproducing information. Copying software, a photograph or a digital file required virtually no additional material resources.

Generative artificial intelligence partially alters this logic.

Producing another response requires computation. Producing a billion responses requires vastly more computation. Information remains digital, but its generation carries a continuous physical cost.

The AI economy therefore remains subject to the fundamental constraints of the industrial world: capital, energy, resources, production capacity and infrastructure.

X. The Next Frontier May Not Be Algorithmic

The global competition surrounding artificial intelligence is generally described as a race between models: which system will be the most intelligent, the fastest or the most capable?

That interpretation may gradually become insufficient.

As algorithmic performance advances, material constraints could become increasingly important. Possessing the best model is of limited value without the accelerators required to deploy it at scale. Possessing the accelerators is insufficient if electricity capacity cannot power them. Building the data centers is insufficient if the grid cannot absorb their load.

The technological frontier therefore becomes an infrastructure frontier.

Countries capable of combining abundant energy, robust electricity networks, financial capacity, regulatory stability, international connectivity and access to semiconductors may acquire a structural advantage in the AI economy.

Conversely, countries that remain merely consumers of artificial intelligence services may find themselves importing not only their technology but, indirectly, their computing capacity as well.

A new digital geography is consequently emerging around several very old resources: land, energy, water, metals and capital.

Conclusion — Behind the Cloud, the Physical World

Artificial intelligence has a peculiar characteristic: the simpler its interface becomes, the more complex the infrastructure required to produce it.

A few words typed on a smartphone can trigger operations performed by a machine containing tens or hundreds of billions of transistors, installed alongside thousands of other processors, inside a building connected to an electricity grid and global telecommunications infrastructure.

Behind that machine are semiconductor factories. Behind those factories are industrial equipment and chemical supply chains. Behind electricity networks are copper, aluminum, transformers and power plants. Behind the entire system are investments now measured in hundreds of billions of dollars.

None of this means that artificial intelligence is materially unsustainable. Efficiency gains can be substantial, cooling systems are evolving, processors are improving, and AI itself can contribute to more efficient use of resources.

But it does mean that its growth is not independent of the physical world.

The defining question of the next decade may therefore not simply be who builds the most advanced artificial intelligence, but who possesses the industrial ecosystem capable of operating it at scale.

Because behind the models, parameters and algorithms remains a much older reality.

Artificial intelligence needs chips. Chips need factories. Factories and data centers need water and electricity. Networks need metals. And all of these infrastructures require land, capital and time.

The cloud was never a cloud.

It is an industry.

Main Sources

  • International Energy Agency, Energy and AI, 2025; projections concerning data-center electricity consumption and accelerated servers.
  • U.S. Department of Energy / Lawrence Berkeley National Laboratory, research on the evolution of data-center electricity consumption in the United States.
  • OECD, research on national compute capacity and Measuring the Environmental Impacts of Artificial Intelligence Compute and Applications.
  • UN Trade and Development, Digital Economy Report 2024, particularly regarding material resources, electronic waste and the geographical distribution of the digital economy’s environmental footprint.
  • Microsoft, Environmental Sustainability Report 2025, data concerning data centers, cooling, water use and contracted renewable-energy capacity.
  • NVIDIA, technical documentation for the Blackwell architecture; accelerator characteristics and manufacturing processes.
  • International Energy Agency, research on critical minerals and copper and aluminum requirements associated with electricity infrastructure.