Sustainable Era

The Physical Limits of Artificial Intelligence

Artificial intelligence (AI) has very real physical limits. To understand the landscape now taking shape, we begin in a place seemingly far removed from Silicon Valley: Denmark.

In 2026, the Danish government introduced a bill to change the rules governing access to the electricity grid, moving away from the first-come, first-served principle. Under the proposed framework, essential services, household needs and projects related to electrification and the energy transition could be given priority, while some large electricity consumers - including the largest and least flexible data centers - would be placed lower in the queue.[1] The decision reflects an increasingly tangible problem: in several parts of the country, available grid capacity is becoming scarce, while connection requests are growing faster than the grid can be expanded.

Behind what appears to be a technical decision lies a much deeper question. If electricity-grid capacity is finite, who should come first? A new AI data center, a factory, a hospital or a household? AI may appear immaterial, but behind a software program, a virtual assistant or an answer that appears on a screen within seconds lies a long physical chain of semiconductors, servers, data centers, cooling systems, electricity, transformers, cables, grids and power-generation facilities.

The digital revolution therefore requires a vast industrial infrastructure. According to the International Energy Agency (IEA), infrastructure and equipment investment by the largest technology companies exceeded $400 billion in 2025 and could rise by a further 75% in 2026. For comparison, five Big Tech companies now invest more than the entire global oil and gas industry spends on oil and gas production.[2]

The central question, then, is whether the physical infrastructure - energy, power grids, data centers, semiconductors, cooling systems and raw materials - will be able to keep pace with the speed of the digital revolution.

A Technology Spreading at Extraordinary Speed

Artificial intelligence is no longer a technology confined to research laboratories or specialist companies: its adoption across the real economy is proceeding at extraordinary speed.

According to the Stanford Artificial Intelligence Index Report 2026, in 2025 88% of surveyed organizations reported using AI. 79% were already regularly using generative AI in at least one business function - the form of AI capable of creating new content, including text, images, video or software, from the information and instructions it receives.[3]

Even more striking is the speed of its diffusion. Generative AI reached an adoption rate of approximately 53% among the adult population considered in the study within three years — faster than either the personal computer or the Internet during their respective early stages of adoption.[4]

The spread of the technology has been accompanied by an equally rapid increase in investment. In 2025, global corporate investment in AI reached $581.7 billion, 130% more than the previous year.[5]

Private investment reached $344.7 billion, with a strong concentration in the United States: $285.9 billion, compared with $12.4 billion in China.[6] The gap, however, needs to be interpreted with caution. Stanford notes that private-investment data probably understate China's role, where a significant share of capital directed towards AI is mobilized through government-backed or government-sponsored funds.[7]

Numbers of this magnitude show how rapidly artificial intelligence is penetrating the economy and how much capital it is attracting. But for investment on this scale to translate into lasting growth, it must generate tangible economic results. And this is where the entire bet on artificial intelligence ultimately rests: productivity.

The Promise of Productivity

The great economic opportunity offered by artificial intelligence lies in higher productivity - the ability to produce more with the same amount of labor and resources, or to achieve the same result in less time.

Early studies show significant results across several areas. According to research reviewed in the Stanford Artificial Intelligence Index Report 2026, customer-service agents using AI resolved 14-15% more issues per hour; software developers using AI completed 26% more work; and in one marketing experiment, AI increased the amount of content produced per worker by 50%.[8]

The benefits, however, are not uniform. They tend to be greater in activities involving clearly defined and easily measurable tasks, while becoming more limited - and in some cases even negative - when the work requires greater judgment or more complex forms of reasoning.[9]

The International AI Safety Report 2026 reaches similar conclusions: a recent analysis of productivity studies at the individual-task level found gains generally ranging from 20% to 60% in controlled studies and from 15% to 30% in most experiments conducted in real-world working environments.[10]

These are significant results, but they need to be interpreted with caution. If a programmer can complete a task 25% faster, this does not mean that GDP will increase by the same amount. For productivity improvements in individual tasks to spread across the economy, companies must reorganize processes, train employees, redesign jobs and genuinely integrate the new technology into their operations.

We have seen this before with electricity and information technology: major innovations can take years before efficiency gains at the level of individual activities translate fully into higher productivity across the economy. The Stanford AI Index likewise notes that these effects do not necessarily become visible immediately at the macroeconomic level.[11]

The real uncertainty is how quickly the productivity gains made possible by artificial intelligence will translate into economic growth.

How Much Economic Growth Can AI Generate?

The International Monetary Fund study The Global Impact of AI: Mind the Gap seeks to answer this question. Rather than making a forecast, the authors construct a range of scenarios to estimate how productivity gains associated with artificial intelligence could affect the global economy.[12]

In the most conservative scenario, after ten years the level of global GDP would be 1.3% higher than in the baseline scenario. In the most favorable scenario, the increase would reach almost 4%.[13]

This does not mean that global GDP will grow by 4% because of artificial intelligence. It represents the difference relative to the level GDP would otherwise reach in the baseline scenario. The outcome will therefore depend on the ability of different economies to turn technological progress into higher productivity.

And it is precisely here that significant differences between countries emerge. In the most favorable scenario, after ten years GDP would be 5.4% higher in the United States, 4.7% higher in other advanced economies - including Japan, the United Kingdom, Canada, Australia, Norway and Israel -, 4.4% higher in the European Union and Switzerland, 3.5% higher in China and 2.7% higher in low-income countries.[14]

Artificial intelligence could therefore make the global economy richer overall while simultaneously widening the gap between countries. Advanced economies appear better positioned to benefit from the new technology, while low-income economies risk seeing significantly smaller gains.

Three factors in particular account for these differences: exposure to AI, meaning the prevalence of activities and occupations that can benefit from the new technology; the ability to integrate it into the economy, which also depends on digital infrastructure, skills and institutional quality; and access to the necessary technologies, including data and computing capacity.[15]

Simply adopting artificial intelligence is therefore not enough. Turning its potential into economic growth requires infrastructure, skills, capital and access to technology.

A New Geography of Technological Power

The resources required to develop artificial intelligence are highly concentrated, both geographically and among a small number of companies. The United States is home to 5,427 data centers, more than ten times as many as any other country. But it is the production of the most advanced semiconductors that most clearly illustrates how this concentration can make the global AI supply chain vulnerable: according to the Stanford AI Index, Taiwan's TSMC (Taiwan Semiconductor Manufacturing Company) manufactures nearly all of the most advanced chips used for artificial intelligence.[16]

A new geography of technological power is therefore taking shape. The United States accounts for the majority of private AI investment and retains a dominant position in the development of the most advanced models; Taiwan plays a central role in semiconductor manufacturing. China remains the United States' main technological competitor, but the gap has narrowed sharply: as of March 2026, according to Stanford, the best US model outperformed the best Chinese model by just 2.7% on the metric considered in the report.[17]

The first phase of the AI revolution - dominated by the race for computing capacity - has featured companies such as US-based NVIDIA, a leader in processors used for AI; Taiwan's TSMC, central to the production of the most advanced semiconductors; the Netherlands' ASML, which supplies the lithography equipment required to manufacture the most advanced semiconductors; and US-based Broadcom and Cadence Design Systems, active respectively in semiconductors and chip-design technologies.

But computing capacity alone is not enough. Chips have to be installed in data centers, continuously supplied with large amounts of electricity and cooled through dedicated systems. The digital revolution thus encounters its physical limits. The first is energy.

Artificial Intelligence Discovers It Has a Body

In 2025, data centers worldwide consumed approximately 485 terawatt-hours of electricity. To put this into perspective, one terawatt-hour is equal to one billion kilowatt-hours. According to the IEA, by 2030 this consumption could reach approximately 950 terawatt-hours, almost doubling in just five years and accounting for around 3% of global electricity demand. Over the same period, electricity consumption by data centers primarily dedicated to artificial intelligence could triple.[18]

That 3% may appear modest, but the global figure does not tell the whole story. Data centers are large facilities concentrated in particular locations, and their development can require new investment in power generation and local electricity grids.[19] Pressure is also increasing inside the data centers themselves. Servers are grouped into structures known as racks, allowing numerous pieces of computing equipment to be concentrated in a small space. By 2027, according to the IEA, a single rack in an advanced data center could, at peak load, draw as much power as around 65 homes simultaneously. Between 2020 and 2025, moreover, the power density of servers used for AI increased elevenfold and could quadruple again by 2027.[20]

This development is also putting pressure on the supply chains for technologies required by the electricity system. Components far less conspicuous than microchips are becoming strategic. The IEA highlights power electronics and transformers in particular and points to the growing importance of energy-storage systems in ensuring a reliable electricity supply to data centers.[21]

Artificial intelligence is thus revealing its most tangible dimension: behind software and algorithms, a vast physical infrastructure is taking shape. The digital revolution is meeting the old industrial economy.

The Second Phase: From Semiconductors to the Power Grid

The first phase of the artificial intelligence revolution could be summarized in a relatively short chain:

AI → semiconductors → servers → IT services.

A much longer chain is now emerging alongside it:

AI → data centers → cooling → electricity → transformers → cables → grids → power generation.

It is along this new chain that activities and infrastructure are coming into play that, until only a few years ago, we would hardly have associated with artificial intelligence. The growth of data centers is increasing demand for power and cooling systems, transformers, cables and electricity-distribution equipment.[22]

But the transformation extends far beyond the boundaries of the data center. Rising electricity demand requires investment in transmission and distribution grids, new interconnections and, in some cases, new generation capacity. The AI industrial chain is therefore progressively extending from semiconductors to the entire electricity system.[23]

This growing demand for electricity is also encouraging the search for new ways to generate, store and manage energy. Alongside renewable energy and storage systems, interest is growing in microgrids and on-site generation and, over different time horizons, in technologies such as advanced geothermal energy and small modular nuclear reactors (SMRs).[24]

This second phase is less visible than the race for semiconductors, but it could prove equally important. The development of AI will depend not only on the availability of increasingly powerful processors, but also on the ability to power and cool them and to connect them to grids capable of supporting their demand.

Computing capacity is therefore encountering a very tangible constraint: producing ever more powerful semiconductors is not enough if the infrastructure required to power and cool them cannot expand at the same pace.

A Truly Sustainable Revolution?

Faced with a technological transformation of this magnitude, sustainability cannot be assessed solely in terms of energy consumption. The question is whether artificial intelligence can continue to develop over time without creating new environmental, economic and social imbalances.

On the one hand, advances in hardware and software are making AI increasingly efficient: according to the IEA, in recent years the energy required to perform a single AI task has fallen by at least a factor of ten each year. On the other hand, the technology is spreading rapidly and applications are being developed that require far more energy, such as video generation and some advanced reasoning systems. Greater efficiency per task therefore does not necessarily translate into lower overall consumption: AI can reduce the energy required to perform specific tasks, but if its use expands rapidly, the resulting increase in overall demand can offset those savings.[25]

The IEA also stresses that rising demand from data centers raises issues that go beyond the amount of electricity required. These include security of supply, grid capacity, energy costs and the availability of the equipment needed to support this expansion.[26]

But the environmental impact of AI will be determined not only by how much energy it requires, but also by how that energy is produced. Data-center growth can accelerate investment in low-emission energy sources and new energy technologies, but the need to secure continuous power generation quickly is also encouraging new fossil-fuel projects, particularly natural gas, especially in the United States.[27]

Paradoxically, artificial intelligence can also help improve grid efficiency: it is already being used to monitor transformers and other equipment, identify potential failures in advance, reduce the risk of outages and make better use of existing infrastructure. The IEA estimates that the use of AI in the energy sector could generate savings of more than 13 exajoules by 2035, equivalent to around 3% of global final energy consumption, provided the barriers currently limiting its adoption can be overcome.[28]

The relationship between artificial intelligence and sustainability is therefore ambivalent: AI can increase pressure on the energy system while at the same time providing tools to make it more efficient. Other dimensions must also be considered. Data centers must continuously dissipate the heat generated by their equipment, and some cooling systems also require water. Building digital and electricity infrastructure also requires components and raw materials, some of which come from highly concentrated supply chains.[29]

Energy, resources and infrastructure therefore show how the growth of artificial intelligence raises questions that extend far beyond technology. Its sustainability must consequently be assessed from a broader perspective encompassing environmental, economic and social dimensions.

It is environmental because the development of AI requires energy, infrastructure and raw materials and, in some cases, directly consumes water. It is economic because investment on this scale must generate sufficient productivity gains and returns to justify its cost. And it is social because artificial intelligence can profoundly transform the labor market, changing occupations and the skills they require while replacing some activities currently performed by people. There is also a distributional question: who will bear the costs of the transformation, and who will reap its benefits?

Capital, Too, Can Become a Constraint

The development of artificial intelligence requires investment on an ever-increasing scale. The ability to finance it may therefore become another constraint on its expansion.

According to the IEA, total investment in data centers between 2026 and 2030 could reach $3.9 trillion. Investment on this scale is unlikely to be funded solely from the balance sheets of large AI companies and will require increasing reliance on debt and equity financing.[30]

The pace of investment will therefore also depend on market expectations regarding the economic returns from AI and on financing conditions. If returns prove lower than expected, or if financial conditions become less favorable, growth in data-center investment could slow, with consequences throughout the chain of infrastructure and technologies supporting AI's development.[31]

The Future of Work Remains an Open Question

The impact of artificial intelligence is not limited to capital, energy and infrastructure. It inevitably extends to work.

According to the International AI Safety Report 2026, one study estimates that around 60% of occupations in advanced economies and 40% in emerging economies are highly exposed to artificial intelligence. This does not mean that these occupations are destined to disappear, but rather that some of the tasks they involve could be automated or performed with the support of artificial intelligence.[32]

The consequences for employment remain difficult to assess, and the evidence available so far is mixed. Recent studies conducted in the United States and Denmark have found no effect on overall employment, while other research points to declining employment or fewer opportunities for early-career workers in occupations most exposed to AI.[33]

A fundamental question therefore remains open: to what extent will AI replace activities currently performed by people, and to what extent will it instead raise labor productivity while reshaping occupations and the skills they require?

Investing in the New AI Economy

For investors, the transformation described so far suggests looking beyond the companies that directly develop artificial intelligence. The growth of AI now involves a much broader industrial chain, extending from computing capacity to the physical and energy infrastructure required to support it. McKinsey highlights how the expansion of data centers is generating growing demand for electrical, thermal and mechanical infrastructure.[34]

The first part of the value chain remains centered on computing capacity, which we already encountered in the first phase of the AI revolution. From an investor’s perspective, it encompasses several complementary segments: from NVIDIA and Broadcom’s processors and accelerators, to TSMC’s manufacturing of the most advanced semiconductors, and finally to ASML’s equipment and Cadence Design Systems’ technologies for chip design and verification.

But as data centers expand, the transformation progressively extends to the entire physical infrastructure required for their operation: semiconductors, data centers, power and cooling, transformers and cables, electricity grids, and energy generation and storage.

In this second part of the chain, Schneider Electric and Vertiv provide technologies for powering and cooling data centers, while ABB operates in electrical infrastructure and distribution and Comfort Systems USA installs mechanical and electrical systems. Prysmian occupies a distinctive position because it operates in both electrical infrastructure and the digital connections required by data centers. The increasing power density of AI systems is driving demand for transformers, distribution systems, cooling and power electronics and could ultimately change the electrical architecture of data centers themselves.[35]

The evolution of artificial intelligence is also opening up a new energy frontier. The need to secure large amounts of electricity within relatively short time frames is increasing interest in new generation capacity, storage systems, microgrids and generation directly connected to data centers. Alongside mature renewable-energy technologies, growing attention is also being directed towards technologies capable of supplying continuous power, such as advanced geothermal energy and next-generation nuclear power, although their levels of maturity and development timelines differ significantly.[36]

Several listed companies already illustrate this convergence between digital infrastructure and energy. NextEra Energy is involved in developing new generation capacity intended, among other uses, to power large data-center projects, while RWE combines power generation and supply with the development of sites equipped with the necessary grid connections. Bloom Energy provides distributed-generation systems capable of directly powering data centers, allowing them to secure new electricity capacity without relying exclusively on the pace of grid expansion. Advanced geothermal energy represents a more recent frontier: Fervo Energy, listed since 2026, is one of the most significant examples of this technology moving towards industrial scale.[37]

Alongside listed groups, a new generation of companies is emerging to address the bottlenecks created by this transformation. Emerald AI develops systems that allow data centers to adapt part of their electricity demand to grid conditions; Amperesand is moving solid-state transformers designed in part for AI data centers towards commercialization; VEIR is testing superconducting systems capable of transmitting large amounts of electricity through limited spaces; and Phaidra uses artificial intelligence to manage and optimize data-center cooling systems.[38]

These are private companies and technologies at different stages of maturity. But their development points to a broader phenomenon: the physical limits of artificial intelligence are themselves becoming a field of innovation. Transformers, grids, cooling systems, storage and intelligent electricity management - technologies that for years remained far from the center of attention surrounding AI - are taking on new strategic importance.

From Growth to Progress

The evidence examined so far illustrates the scale of the transformation under way. 88% of surveyed organizations are already using artificial intelligence, global corporate investment in AI reached $581.7 billion in 2025, and data centers could consume around 950 terawatt-hours of electricity by 2030. According to the International Monetary Fund study discussed earlier, under the different scenarios considered, AI could raise the level of global GDP over a ten-year period by between 1.3% and almost 4% above the baseline scenario.

Energy, capital and infrastructure will determine the speed and scale of this transformation. But a deeper, existential question remains: whether, and to what extent, we will be able to turn the opportunities offered by artificial intelligence into human progress.

From the perspective of economist Amartya Sen, development cannot be assessed solely through growth in income or wealth; it must also be considered in terms of expanding each person's capabilities, so that individuals have genuine opportunities to fulfil themselves — in other words, to live a life they have reason to value.[39] This perspective raises a fundamental question in relation to artificial intelligence as well: will the AI revolution genuinely improve people's lives through better services, new knowledge, greater innovation and new opportunities, or will its benefits remain concentrated in the hands of a few?

The issue, therefore, concerns not only the economic outcomes of this transformation, but also the real opportunities it will create and who will have access to them. Artificial intelligence can help expand those opportunities, but it cannot decide which ends should be pursued, nor what direction the progress it helps accelerate should take. These are choices that humanity as a whole will be called upon to make, and they will become increasingly important as AI's capacity to shape the evolution of our societies grows. Turning this new economic and technological power into genuinely shared progress will be one of humanity's great responsibilities.

V.C.M

Bibliography

Bengio, Y. et al., International AI Safety Report 2026, Department for Science, Innovation and Technology, DSIT 2026/001, 2026.

Cerutti, E. M., Garcia Pascual, A. I., Kido, Y., Li, L., Melina, G., Mendes Tavares, M., Wingender, P., The Global Impact of AI: Mind the Gap, IMF Working Paper No. 2025/076, International Monetary Fund, April 2025.

Cleantech Group, 2026 Cleantech Outlook: Sovereignty, Decentralization, and the “Pressure Cooker” Effect, 2026.

Cleantech Group, 2026 Global Cleantech 100 Trend Watch, 2026.

DeOrsey, A., Fervo's IPO Establishes Enhanced Geothermal Baselines, Presents New Competition to Fission, and Adds a New LDES Class, Cleantech Group, May 19, 2026.

Folketinget, L 23 – Forslag til lov om ændring af lov om elforsyning og lov om fremme af vedvarende energi [Proposal for an Act Amending the Electricity Supply Act and the Act on the Promotion of Renewable Energy], introduced August 20, 2026.

Gilani, Z., Bending the Limits for Compute: Optimizing Flexibility for Data Centers, Cleantech Group, September 2, 2026.

Goodpaster, M., Patel, M., Sachdeva, P., Huang, S.-Y., The $7 Trillion Data Center Build-Out: How Industrials Can Capture Their Share, McKinsey & Company, March 27, 2026.

Government of Denmark, Bred politisk opbakning til akutplan for elnettet [Broad Political Support for Emergency Plan for the Electricity Grid], Ministry of Climate, Energy and Utilities, July 1, 2026.

Hernandez Diaz, D., Tai, H., Noffsinger, J., Data Centers: Catalyzing Innovation in Energy, McKinsey & Company, June 26, 2026.

International Energy Agency (IEA), Key Questions on Energy and AI, Paris, 2026.

McKinsey & Company, Powering AI: How Real Is the Risk of Overbuilding?, July 31, 2026.

Pflugmann, F., Bauer, H., Gigliotti, L. et al., Charging Ahead: Electrification Equipment Trends to Watch in 2026, McKinsey & Company, March 5, 2026.

Sen, A., Development as Freedom, Alfred A. Knopf, New York, 1999.

Shehabi, A. et al., 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, 2024.

Stanford Institute for Human-Centered Artificial Intelligence (HAI), The 2026 AI Index Report, Stanford University, 2026.

Zegeye, A., Lee, D., Goodpaster, M., Sachdeva, P., Huang, S.-Y. et al., The Shift to 800-Volt DC at Data Centers: Implications for Providers, McKinsey & Company, July 30, 2026.