Disclaimer / Methodological Note: This report has been compiled exclusively on the basis of public sources only. It presents a synthesis of factual information, official data, and documented positions from EU institutions, international organisations, press releases and independent research bodies. It does not express an opinion, political judgement, or recommendation of its own.

Summary

The report deliberately avoids predicting either success or failure. Instead, it identifies a series of questions that deserve closer examination. Can AI infrastructure ultimately pay for itself? Who should own the physical infrastructure on which AI depends? Will specialised infrastructure companies gradually replace hyperscalers as owners of data centres? How much of today's AI wealth reflects productive assets and sustainable future cash flows, and how much reflects expectations embedded in market capitalisations? If AI investment disappoints, could financial institutions become the channel through which losses spread more widely through the economy?

These are not merely questions about artificial intelligence. They concern the financing of one of the largest industrial transformations of the twenty-first century.

The purpose of this report is therefore not to predict the future, but to document the evidence currently available, identify the questions that remain unanswered, and provide a framework for understanding one of the most significant economic transitions now under way.

Intelligence is no longer simply a software industry. It is rapidly becoming one of the largest private infrastructure programmes in modern economic history.

The world's leading technology companies are investing unprecedented sums in data centres, advanced processors, electricity generation, transmission networks and supporting infrastructure. Forecasts suggest that cumulative AI infrastructure investment could reach several trillion dollars over the coming decade. The central question is therefore no longer whether AI represents a technological revolution. It is whether this infrastructure can be financed—and ultimately generate sufficient economic returns to justify the capital committed.

Until now, much of the investment has been supported by exceptionally high market valuations, allowing companies to raise large amounts of equity capital. However, equity financing has natural limits. It dilutes existing shareholders and depends on investors continuing to assign increasingly high valuations to AI companies. As investment requirements continue to grow, banks, private-credit providers, infrastructure funds, insurers and pension funds are expected to play an increasingly important role in financing AI infrastructure.

This shift changes the nature of the debate. The key question is no longer simply whether companies can raise capital. It is whether AI infrastructure can generate sufficient long-term cash flows to service the capital—whether debt or equity—that finances its construction.

Electricity illustrates how fundamentally AI differs from previous software revolutions. Access to reliable and affordable power is becoming a strategic competitive advantage. The future leaders in artificial intelligence may not necessarily be those that develop the most powerful models, but those capable of securing electricity, computing capacity and financing on an industrial scale.

As AI infrastructure expands, a second question emerges: who ultimately pays? Companies may finance part of the investment directly. Governments are increasingly treating computing infrastructure as strategically important and are beginning to support sovereign computing capacity. Consumers may also bear part of the cost indirectly through higher prices, electricity infrastructure or taxation. The long-term sustainability of AI will therefore depend not only on technological progress, but also on whether the distribution of costs and benefits is regarded as economically and politically acceptable.

From a technology boom to a financing boom

AI is no longer financed only through research budgets or ordinary information-technology expenditure. It has become a physical infrastructure programme. The leading AI technology groups are building data centres, purchasing processors, signing long-term energy contracts and securing land, cooling systems and network connections on a scale previously associated with the energy, transport and telecommunications sectors.

Goldman Sachs estimates that annual AI-related capital expenditure could reach approximately $765 billion in 2026 under its baseline scenario and rise to $1.6 trillion by 2031.

The technological competition is therefore becoming a capital-expenditure problem. It cannot be financed indefinitely by raising additional billions of dollars on equity markets. New share issues provide capital but dilute existing shareholders. At the same time, high expectations surrounding AI can inflate company valuations—sometimes to bubble-like levels—and make further equity financing easier. Yet rising market capitalisations do not generate the long-term cash flows required to repay infrastructure investment.

As funding requirements continue to grow, the AI build-out will increasingly depend on banks, private-credit providers, infrastructure funds, insurers and other long-term financiers. Unlike equity investors, whose returns depend largely on future growth expectations, lenders are expected to apply traditional credit discipline. Their focus should remain on predictable cash flows, customer commitments, electricity availability, construction costs, collateral values and the project's ability to service debt under less favourable conditions.

The question is whether those lending standards will remain intact. Financial institutions face the same competitive pressure as technology companies: no one wants to miss what is widely portrayed as the defining investment cycle of the coming decade. Competitive pressure can encourage increasingly optimistic assumptions, particularly when projects are backed by well-known hyperscalers or when AI infrastructure is widely perceived as an almost inevitable source of future growth.

The comparison with the period preceding the mortgage-backed securities crisis should therefore not be dismissed too quickly. The underlying assets are clearly different—but whether the financing dynamics could eventually prove comparable remains an open question. Before the global financial crisis, lenders and investors continued financing an expanding market because remaining on the sidelines appeared commercially more risky than participating. Credit standards gradually weakened, risks were distributed through increasingly complex financing structures, and confidence in ever-rising demand and asset values often replaced conservative credit analysis.

A similar fear of missing out could emerge in AI infrastructure finance. Lending decisions may gradually become driven less by project-level cash-flow analysis than by the broader conviction that AI demand will continue to justify almost any level of investment. If that happens, the quality of the financing may become as important as the quality of the technology itself.

The AI investment race may consequently produce two connected forms of excess. Technology companies may construct more capacity than future demand can support, while financial institutions may extend more credit than the underlying projects can safely repay. At that point, what began as a technological competition becomes a financing boom—and the quality of lending decisions becomes as important as the quality of the technology itself.

Capital expenditure is beginning to exceed internal funding capacity

The AI infrastructure race is entering a new phase. The question is no longer simply how much technology companies are willing to invest, but how unprecedented levels of capital expenditure can continue to be financed over many years.

Reuters estimates that by 2027 the combined capital expenditure of the largest hyperscalers could exceed the growth in their operating cash flow by almost $200 billion, implying that investment is growing faster than internally generated cash. Whatever the precise figures, the direction is clear: the scale of AI infrastructure is becoming too large to rely indefinitely on internally generated funds.

One obvious source of financing is the equity market. Rising share prices allow companies to issue additional shares and raise fresh capital. Yet this solution has limits. Equity financing dilutes existing shareholders and depends on investors continuing to assign increasingly high market valuations to AI companies. The financing of physical infrastructure therefore becomes increasingly linked to stock-market expectations rather than to the cash flows generated by the infrastructure itself.

This raises a broader question that extends well beyond corporate finance. To what extent are today's AI market capitalisations reflecting productive assets and sustainable future earnings, and to what extent are they reflecting expectations that may continue to inflate as long as investors remain convinced that AI growth is virtually unlimited? Put differently, is part of today's AI wealth real economic wealth, or is it largely virtual wealth created through market valuations?

As financing requirements continue to grow, equity markets alone are unlikely to satisfy the enormous capital needs of AI infrastructure. This raises a second question. Will hyperscalers continue to own the physical infrastructure themselves, or will ownership increasingly pass to specialised infrastructure developers and long-term investors while technology companies become tenants purchasing computing capacity?

These questions are likely to shape not only the future financing of AI but also the future valuation of AI companies themselves. They will be examined in the following chapters.

Electricity: the Forgotten Cost of Artificial Intelligence

Artificial intelligence is often presented as a software revolution. In reality, it is rapidly becoming one of the largest private infrastructure programmes of the modern economy. Behind every AI model lies an enormous physical system of data centres, processors, cooling equipment, fibre-optic networks and, above all, electricity.

The International Energy Agency estimates that electricity consumption by data centres could almost double by 2030, largely because of AI. While this remains only a modest percentage of global electricity production, the challenge is highly concentrated. AI data centres require enormous amounts of power in specific locations where electricity, land and network capacity are simultaneously available. In many regions, securing sufficient electricity has become more difficult than obtaining processors.

Electricity is therefore no longer simply an operating expense. It is becoming a strategic asset that may determine which companies can expand and which cannot.

The difference can already be seen among today's AI leaders.

Amazon has gradually transformed electricity into a competitive advantage. Through AWS, long-term power purchase agreements, investments in renewable energy, partnerships in nuclear energy and the acquisition of strategic sites for future data centres, it has built an ecosystem in which energy planning forms part of its AI strategy. Amazon is no longer simply purchasing electricity; it is helping shape the infrastructure on which future AI services will depend.

Other AI developers face a different reality. OpenAI, despite developing some of the world's most advanced models, remains largely dependent on infrastructure provided by partners. Similarly, xAI demonstrated extraordinary speed in building its "Colossus" supercomputer in Memphis, yet the discussion rapidly shifted away from processors and algorithms towards electricity supply, gas turbines, environmental permits and transmission capacity. In both cases, the limiting factor was no longer artificial intelligence itself but the infrastructure required to power it.

These examples illustrate a broader shift. The future leaders in AI may not necessarily be those that develop the most powerful models, but those capable of securing computing capacity, electricity and financing on an industrial scale.

The financial implications are equally significant. Building a modern hyperscale AI data centre can require investments of between $5 and $15 billion, while large AI campuses incorporating dedicated energy infrastructure may require tens of billions of dollars. Collectively, the largest technology companies are now investing several hundred billion dollars every year in AI infrastructure, and cumulative investment over the coming decade is expected to reach several trillion dollars.

Processors attract most public attention, but they represent only part of the investment. Land acquisition, buildings, cooling systems, substations, transmission lines, fibre networks, backup generation and electricity connections represent a substantial share of the total capital required. AI is therefore evolving from a software industry into one that increasingly resembles electricity generation, telecommunications and transport infrastructure.

This evolution brings the report back to its central financial question. Whether these investments are financed through equity or debt, they must ultimately generate sufficient cash flow to justify the capital committed. Electricity is therefore no longer merely an engineering challenge. It has become one of the key financial variables that will influence the future economics of artificial intelligence.

Who Will Ultimately Pay for AI: Companies, Consumers or Governments?

The expansion of artificial intelligence is often discussed as a technological challenge. Increasingly, however, it is becoming a question of political economy: who will ultimately pay for the enormous infrastructure on which AI depends?

The first answer appears straightforward: technology companies themselves. They invest hundreds of billions of dollars in processors, data centres and electricity infrastructure because AI has become central to their future growth. Increasingly, they are also securing long-term electricity supplies, investing in renewable energy and even exploring partnerships around nuclear power. These investments are no longer peripheral to AI; they are becoming part of the business model itself.

Yet the scale of the required investment raises an obvious question. Can companies alone finance the physical infrastructure needed for the next generation of AI? As investment programmes move from tens to hundreds of billions of dollars, the answer is becoming less certain.

A second possibility is that part of the cost is ultimately transferred to consumers. This need not happen directly through AI subscriptions alone. It may occur indirectly through higher electricity prices, increased prices for AI-enabled services, higher taxes to support public infrastructure or through the wider costs of expanding electricity generation and transmission networks.

This question has already become politically sensitive.

In the United States, concerns have grown that electricity infrastructure built primarily for AI data centres could increase costs for ordinary households. In response, several of the largest AI and cloud companies have publicly committed to covering the cost of electricity infrastructure required by their own projects rather than transferring those costs to existing ratepayers. The debate itself illustrates that the question is no longer theoretical. It has already entered public policy.

Governments represent a third source of financing. Across Europe, publicly supported AI Factories and Gigafactories are being developed to provide sovereign computing capacity for researchers, industry and public institutions. Similar initiatives exist in the United Kingdom and India. These programmes do not amount to the nationalisation of commercial data centres, but they do suggest that governments increasingly regard large-scale computing infrastructure as strategically important, much like electricity networks, railways or nuclear facilities.

Whether this trend eventually leads to greater public ownership remains uncertain. For the moment, governments appear more interested in guaranteeing national access to computing capacity than in owning all AI infrastructure themselves. Nevertheless, the boundary between private investment and public infrastructure is becoming increasingly blurred.

The broader question is whether society will continue to support this investment if the economic benefits of AI are not widely shared. Economists Daron Acemoglu and Simon Johnson argue in Power and Progress that technological revolutions have rarely produced shared prosperity automatically. Throughout history, productivity gains have often benefited owners of capital long before they improved the living standards of the wider population. Only political institutions, regulation and countervailing social forces eventually broadened those benefits.

Their argument is particularly relevant to artificial intelligence. If AI substantially increases productivity while simultaneously reducing employment opportunities, weakening wage growth or concentrating wealth in a relatively small number of companies and shareholders, governments may find it increasingly difficult to justify asking consumers and taxpayers to finance the infrastructure that makes AI possible.

This does not imply that such an outcome is inevitable. AI may ultimately generate sufficient economic growth and new forms of employment to offset these concerns. But it does highlight an important question that receives surprisingly little attention. The long-term success of AI may depend not only on technological progress, but also on whether the costs and benefits of the AI revolution are perceived to be distributed fairly across society.

For investors, this may become as important as advances in AI models themselves. Public acceptance, political legitimacy and access to affordable electricity could ultimately prove to be as valuable as technological leadership. The future winners in artificial intelligence may therefore be determined not only by superior algorithms, but also by their ability to build an economically and politically sustainable infrastructure.

---

Selected sources

International Organisations

Bank for International Settlements (BIS)

Annual Economic Report 2026

Quarterly Review, March 2026 (Financing AI infrastructure)

BIS Working Paper: The AI Investment Race (2026)

BIS research on AI financing, private credit and financial stability

International Energy Agency (IEA)

Energy and AI (2025/2026)

Key Questions on Energy and AI

Data Centre Electricity Demand Outlook (2026)

Financial Institutions

Goldman Sachs Research

Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out (2026)

BloombergNEF

AI Data Center Build Advances at Full Speed

Global Data Centre Investment Outlook (2026)

Chicago Federal Reserve

Tail Risk for Banks Posed by Investments in Generative AI (2026)

Public Policy and Government Sources

European Commission

AI Factories Initiative

AI Gigafactories Programme

EuroHPC Joint Undertaking

United Kingdom Government

UK Compute Roadmap

AI Research Resource

Critical National Infrastructure designation for data centres

IndiaAI Mission

IndiaAI Compute Capacity Programme

Government GPU infrastructure initiatives

News and Financial Press

Reuters

AI investment boom and hyperscaler capital expenditure

AI financing and operating cash-flow analysis

US electricity infrastructure and ratepayer protection

Hyperscaler debt issuance

Financial Times

Amazon investment in OpenAI

AI infrastructure financing

Data-centre investment trends

The Wall Street Journal

AI monetisation

Capital expenditure by hyperscalers

Cloud economics

Bloomberg

Data-centre financing

Electricity demand

AI infrastructure investments

Academic and Research Literature

Daron Acemoglu & Simon Johnson

Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity (2023)

(Referenced in the discussion on the distribution of technological gains and public acceptance of AI infrastructure.)

OECD

Research on AI governance, agentic AI and AI infrastructure.

Companies

Annual reports, investor presentations and earnings calls from:

Amazon

Microsoft

Alphabet

Meta

OpenAI (where available)

Oracle

NVIDIA

xAI