The five largest technology companies in the world invested over 400 billion dollars in 2025 in their infrastructure, driven almost entirely by demand for AI computing. In 2026, this figure is expected to grow by 75%. To understand what this means, it helps to know that NASA’s annual investment at the peak of the Moon race represented, in constant dollars, approximately 50 billion. What is happening today is of a different nature.

This is not an ordinary technological race. It is the formation of a structural bottleneck: the capacity to produce artificial intelligence at scale depends on a physical resource — electricity — whose concentrated access between a few hands is already drawing the outlines of the next monopoly of the century.

The public debate on AI revolves around jobs. That is the wrong terrain. The real stakes are being played out in the price of the electron and in the question of who controls the infrastructure that transforms that electron into intelligence.

The Essentials

  • The five largest technology companies spent more than 400 billion dollars on infrastructure investments in 2025, with projected growth of 75% in 2026, according to IEA data.
  • This spending finances primarily data centers whose global electricity consumption is projected to nearly double by 2030, rising from 300 to 380 TWh in 2023 to approximately 945 TWh according to IEA projections published in April 2025.
  • Three companies — Microsoft, Google, and Amazon — control more than 65% of global cloud computing capacity, creating a barrier to entry valued at tens of billions per year.
  • Europe and most emerging countries do not have access to sufficiently stable low-carbon electricity markets to compete in this infrastructure, transforming a technological advantage into a geographical and institutional one.
  • Solutions exist: collective purchasing pools, energy efficiency standards, green electricity markets reserved for regional players, and regulation of exclusivity contracts on renewable energy.

Four Hundred Billion Cannot Be Purchased With an Idea

The history of industrial capitalism is the history of a few resources that structured global competition for decades. Coal made England. Oil made the Middle East and American majors. Access to financial capital determined who could innovate in the twentieth century.

Today, the structuring resource is computing. And computing has an immediate physical constraint: electricity.

A large language model like GPT-4 requires tens of thousands of specialized chips running permanently in data centers that consume as much energy as a mid-sized city. Training a next-generation model costs between 100 million and a billion dollars in computing alone, according to estimates from several research teams. Inference — simply answering a request — costs ten to a hundred times more per token than web browsing. These costs are not falling fast enough for new entrants to absorb them.

What creates the barrier is not technological secrecy. Algorithms circulate. Models open up. What creates the barrier is the physical infrastructure that runs them. And this infrastructure is purchased in tens of billions per year, on stable electricity markets, with chip supply agreements that assume contractual relationships with an oligopoly of manufacturers.

Data Center Electricity Consumption Doubles in Five Years

The International Energy Agency estimates that data centers consumed between 300 and 380 TWh of electricity globally in 2023 — less than all of France, whose consumption was approximately 445 TWh that same year. In 2024, this consumption reached 415 TWh. By 2030, it should approach 945 TWh in the IEA’s base scenario, according to its Energy and AI report published in April 2025. This trajectory is now driven by a significant and growing share of AI-related workloads, a proportion that has notably accelerated since 2022.

This near-doubling in a few years is itself conservative. It is based on the assumption that the energy efficiency of chips will continue to improve at the current rate. If demand for AI grows faster than chip improvements — which is the observable trend since 2023 — consumption could reach 1,200 TWh by 2030.

The problem is not electricity in itself. It is reliable, cheap, and low-carbon electricity, in geographically suitable regions. Data centers need cold for cooling, network stability to avoid interrupting calculations, and green electricity to honor the climate commitments that major tech companies have made — under pressure from investors and regulators. These three simultaneous conditions are found in a limited number of geographies: the northeast United States, Scandinavia, a few enclaves in Southeast Asia.

This is where implantation decisions are made. And companies that establish themselves first sign long-term electricity supply contracts — ten, twenty years — that lock in the resource for those that follow.

Who Controls Computing Controls Access to Intelligence

Concentration is not a projection. It is already here.

Microsoft, Google, and Amazon control approximately 65 to 70% of global cloud computing capacity, according to Synergy Research Group data published in 2025. Microsoft, Amazon, Alphabet/Google, Meta, and Oracle complete the picture of an oligopoly of five players that own, operate, or contract virtually all the infrastructure on which global AI runs.

This concentration has a direct consequence for the AI economy: companies that are not part of this club do not have access to computing power on the same terms. They buy computing on demand, at retail prices, on spot markets. The major tech companies have negotiated electricity supply contracts at prices significantly lower than spot rates, according to Wood Mackenzie analyses. They amortize their chips over hundreds of millions of hours of computing. Their marginal cost of intelligence is structurally lower than that of any competitor.

Daron Acemoglu and Simon Johnson, in Power and Progress, pose the fundamental question: technology does not benefit everyone, except when forced to do so. AI computing illustrates this mechanism with almost didactic clarity. Technology progresses. Capabilities accumulate. But access to these capabilities follows the distribution of invested capital, not the distribution of needs or ideas.

AI Advances Quickly But Its Infrastructure Digs a Lasting Gap

The pace of model improvement is spectacular. What AI did not know how to do in January, it masters by September — this acceleration in capabilities is real and documented. But it masks a growing asymmetry: gains in capacity benefit first those actors who can afford the infrastructure to exploit them.

The analogy with the nineteenth century is not perfect, but it is instructive. When railroads structured the American economy, companies that controlled the rails dictated transport prices for three decades. It was not railroad engineering that posed a problem — engineers moved freely. It was the physical infrastructure and its concentration in a few hands that ultimately necessitated the dismantling of railroad trusts by Theodore Roosevelt starting in 1902.

The question for the years ahead is not whether AI will be capable of doing more. It will be capable of doing more. The question is whether those not part of the computing oligopoly will have access to the same tools on terms compatible with normal economic competition.

For emerging countries, this issue is particularly structuring. Most have neither the electricity markets nor the transmission networks needed to host competitive data centers. They will be clients of infrastructure, never producers. This is an unprecedented form of technological dependence, which adds to dependence on raw materials and finance. We have seen how thirteen African countries attempted to take control of their mining destiny by banning raw mineral exports — for computing power, no equivalent lever exists yet.

Concrete Solutions Exist, They Require Political Will

It would be analytical error to describe a bottleneck without identifying the mechanisms that can regulate it.

Several avenues are being explored, at different stages of maturity.

The first is regulation of renewable electricity supply contracts. Major tech companies sign Power Purchase Agreements (PPAs) with renewable energy producers that lock in capacity for twenty years. In Europe, where the green energy market is still being built, proposals are circulating to reserve part of this capacity for regional players. The European Commission has integrated this question into its Data Act and its work on the AI Act, without yet producing binding measures on the infrastructure itself.

The second is the pooling of computing access. Several national initiatives — in Germany, in France with the digital sovereignty program, in Japan — are attempting to create shared computing pools accessible at administered rates for research laboratories, SMEs, and administrations. These programs remain modest compared to private volumes. France has announced a significant number of A100 chips available for research via the Clara program, a volume that nonetheless represents only a few weeks of the marginal investment of a single hyperscaler.

The third is the regulation of energy efficiency. By imposing performance standards per unit of computing, regulators can accelerate replacement of inefficient infrastructure and reduce the advantage of players installed on amortized old equipment. Europe is working on metrics of this type as part of its energy efficiency directive, but implementation remains slow.

The fourth avenue is transparency on costs. Today, cloud computing prices are opaque. Large companies sign confidential contracts with massive discounts. Startups pay list price rates. Making weighted average rates public would allow antitrust regulation to assess whether these gaps constitute an abuse of dominant position. European antitrust authority has opened preliminary investigations into the cloud market, but no structuring decision is expected before 2027 at the earliest.

What 2030 Will Say About 2025

The generational issue is real. Infrastructure decisions made between 2023 and 2028 have a lifespan of fifteen to twenty years. Data centers built today will be in service in 2040. Electricity contracts signed today run until 2045. The cost advantages accumulated by incumbent players during this window will not be erased by technology progression alone.

That is the difference between ordinary technological concentration and what is being observed. In most industries, competitive advantage erodes over time because new entrants can replicate infrastructure at lower cost. In AI computing infrastructure, advantage strengthens: the earlier an actor has invested, the lower its marginal cost, the more it can offer prices its competitors cannot reach, which generates revenues that finance subsequent investments.

This circle is not inevitable. It is so in the absence of intervention. What makes the current decade different from those that preceded it in the history of technology is that the regulatory tools exist — antitrust, efficiency standards, access markets, price transparency — and that regulators, particularly in Europe, have the political will to use them. The question is not whether we know how to do it. The question is whether we will act fast enough that the structures of 2035 are not simply an amplification of those of 2025.

Europe has already shown it can transform its regulation into an instrument of economic sovereignty. Computing infrastructure is the next terrain where this capacity will be tested — and where inaction would have lasting consequences, not for technology companies, but for the hundreds of millions of users, businesses, and administrations who will need equitable access to artificial intelligence in the decades to come.

Sources

  1. Carbon Direct, AI Scale and Climate Commitments: A 2026 Outlook — https://www.carbon-direct.com/insights/ai-scale-and-climate-commitments-a-2026-outlook
  2. International Energy Agency (IEA), Electricity 2024: Analysis and Forecast to 2026 — https://www.iea.org/reports/electricity-2024
  3. Synergy Research Group, Cloud Infrastructure Market Shares Q4 2024 — no guaranteed link, source cited without URL
  4. Daron Acemoglu and Simon Johnson, Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity, PublicAffairs, 2023
  5. Wood Mackenzie, analyses of Power Purchase Agreements markets for hyperscalers, 2024-2025 — no guaranteed link, source cited without URL
  6. European Commission, AI Act and Data Act — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
  7. IEA — Key Questions on Energy and AI (April 2026) — https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
  8. IEA — Energy and AI (April 2025) — https://www.iea.org/reports/energy-and-ai/executive-summary
  9. Synergy Research Group — Cloud market share Q3 2025 — https://www.srgresearch.com/articles/cloud-market-share-trends-big-three-together-hold-63-while-oracle-and-the-neoclouds-inch-higher
  10. RTE — French Electricity Review 2023 — https://analysesetdonnees.rte-france.com/en/annual-review-2023/keyfindings
  11. Epoch AI / Stanford — AI Training Costs (arXiv 2024) — https://arxiv.org/html/2405.21015v1
  12. The Planetary Society — Cost of the Apollo Program — https://www.planetary.org/space-policy/cost-of-apollo
  13. Visual Capitalist / Epoch AI — Big Tech Capex 2022–2025 — https://www.visualcapitalist.com/visualized-big-tech-ai-spending/