Abstract

This paper examines how investment in artificial intelligence infrastructure may translate into durable returns for companies and their investors. It evaluates bubble risk and the possibility that valuations, expected demand and project capacity move ahead of realised earnings, utilisation and cash generation, creating the conditions for an overinvestment cycle or a sharp market correction. It also examines the growing role of private credit and other debt structures in financing the build-out. Available evidence shows rapid expansion in private credit and technology lending. Public data on credit specifically funding AI infrastructure remain limited, so the paper separates broad technology lending from disclosed AI-infrastructure financing.

For the Americas, the paper examines the United States in detail and draws on selected examples from other markets in the region. The European analysis considers industrial suppliers, local infrastructure, policy initiatives and adoption conditions. Asia is included because it remains central to semiconductor production and other supply-chain dependencies. Investor positioning is assessed using disclosed transactions, financing research and portfolio channels. The central assessment is continued investment accompanied by increasing differences among companies and projects: scarce capabilities may retain value, while undifferentiated capacity may face pricing and utilisation pressure.

Keywords: artificial intelligence; data centres; semiconductors; electricity infrastructure; investment returns; investor positioning; valuation

1. Introduction and research question

Artificial intelligence is often presented to investors as a technology story, centred on better models, faster chips, more capable assistants and changes to how people work. Yet the investment requirements are also physical. Advanced chips must be manufactured, data centres constructed, electricity delivered through reliable networks, and heat removed from computing equipment. These activities form a substantial chain of capital expenditure.

The investment question extends beyond which AI applications will attract users. It asks which businesses can convert infrastructure spending into durable cash flows and adequate returns on invested capital. This paper analyses that question across the main infrastructure layers, regional markets, financing channels and portfolio applications.

A growing industry, a successful company and an attractive security are distinct propositions. The analysis consequently separates evidence of demand from evidence of profitability and separates business quality from valuation. It concludes with a conditional outlook, a set of monitoring indicators and implications for four model portfolios.

2. Research scope and analytical approach

The paper uses public primary sources available by 17 October 2026. These include company filings and results releases, government and energy-system publications, and research from the International Energy Agency, the Bank for International Settlements, the European Central Bank and the International Monetary Fund. Reported results, management expectations, public-policy targets and scenario analysis are identified separately because they provide different kinds of evidence.

The regional discussion considers the United States and Europe in detail, selected examples from elsewhere in the Americas, and Asia's role in the supply chain. Investor positioning is assessed through public-market exposures, disclosed private-capital programmes and financing research.

The company examples highlight distinct business models across the value chain. The operating and valuation scenarios show how changes in key assumptions affect returns, while the portfolio discussion considers strategic fit across the four mandates.

3. The investment question and the evidence on demand

The scale of demand is no longer just a hypothetical future opportunity.

NVIDIA reported $89 billion of Data Center revenue in its second quarter of fiscal 2027, ended 26 July 2026. Microsoft reported 43% growth in Azure and other cloud-services revenue in its quarter ended 30 June 2026. [1, 2]

The physical implications are substantial too. In its 2026 update, the International Energy Agency projects that global data-centre electricity consumption could rise from 485 terawatt-hours in 2025 to approximately 950 in 2030. The projection covers data centres across the economy, with AI among the sources of expected growth. [3]

These results confirm substantial spending; the investment question is which suppliers can convert that demand into durable cash flow.

A company's capital expenditure becomes revenue for its suppliers. For investors, however, the more important measure is the cash remaining after the company has paid the costs of competition and asset maintenance.

The analysis therefore focuses on durable cash generation as well as exposure to industry growth.

4. Industry structure and business models

The industry is best analysed as a set of businesses with distinct economic characteristics.

These categories overlap, but their economics differ. A data-centre owner that leases powered space faces different risks from an operator that also owns computing equipment, which may become technologically obsolete. A regulated electricity network also has a different revenue model from a merchant generator exposed to wholesale prices.

NVIDIA, TSMC, ASML, Microsoft, Schneider Electric and Eaton illustrate different positions in the value chain, alongside their other businesses. [1, 2, 4, 5, 6, 7]

Holding several such companies may reduce exposure to individual company risks while preserving substantial exposure to the same underlying investment cycle.

Market measurement and double counting

AI-infrastructure estimates can overstate activity when they combine spending and sales from different stages of the same supply chain.

Suppose a data-centre operator spends €100: €60 on computing equipment, €20 on electrical systems and €20 on construction. The €100 is the operator’s total project spending. The computing-equipment supplier may then pay a chip manufacturer for components used in its equipment. That manufacturer’s sale is part of the supply chain behind the operator’s €60 purchase. If an estimate adds the operator’s €100 to every supplier’s revenue, it counts some of the same economic activity more than once.

Other figures describe different stages of investment. A fundraising target records capital a manager intends to raise; an acquisition records a change in ownership; and a power contract supports a project. None measures the same thing as end-customer spending or newly operating capacity.

Market estimates should compare like with like: end-customer spending with end-customer spending, supplier revenue with supplier revenue, and operational capacity with operational capacity. Consistent definitions and time periods make it easier to track actual growth and avoid inflating totals with transactions counted at multiple stages of the supply chain.

Workload differences between training and inference

How AI is used affects the infrastructure customers need. Training is the process of building or updating a model. Inference is the use of a trained model to produce an answer, recommendation or action. Both require computing, although the hardware, locations and purchasing arrangements can differ.

Large training jobs may require many accelerators to work together efficiently. Inference services may place greater weight on response speed, reliability, energy cost per task and proximity to users. Some inference tasks can use smaller models or specialised processors; others still require more capable systems.

As AI use shifts between training and inference, the mix of hardware, computing locations and service costs may change. Software, networking, memory, utilisation and switching costs will all shape which suppliers benefit.

Chip shipments and token volumes are useful measures, alongside the cost of completing a task and the value it creates for a paying customer. A system earns a lasting place in the market when it solves a customer problem at a cost the customer is willing to pay.

Who owns each part of a data-centre project?

The businesses involved in a data-centre project earn different returns and carry different risks. A property owner may earn rent. A developer may earn a margin for constructing the facility. A cloud operator may buy servers and sell computing time. A lender earns interest but depends on the borrower’s cash flow and the value of its collateral.

These businesses also face different risks. The landlord depends on the tenant’s ability to pay and on whether the building can be reused. The computing operator must consider electricity costs, customer demand and the need to replace servers. The lender must consider whether the borrower can repay and whether collateral could be sold for enough if it defaults. A facility can remain standing while its computing equipment becomes outdated or its operator loses customers.

Before comparing investment returns, an analyst should therefore identify who owns the land, power connection, cooling system and computing equipment; who has signed the customer contracts; and who must pay for repairs, upgrades and replacement equipment. That ownership map shows which company receives each revenue stream and where the project's costs and risks ultimately sit.

5. The United States and the wider Americas

The United States provides the clearest starting point for analysing demand in the Americas. The IEA's 2025 study estimated that it accounted for 45% of global data-centre electricity consumption in 2024, compared with 25% for China and 15% for Europe. [8]

The model is strongly shaped by large technology platforms: they can combine existing customers, software distribution, cloud services and substantial financing capacity. That can make it easier to turn a new capability into a product customers already know how to purchase. It can also create pressure to invest defensively, because being short of capacity may mean losing customers to a rival.

Microsoft's cloud growth offers evidence of demand within one large platform. For investors, the key question is how much of that growth comes from AI services and what return the latest infrastructure earns. [2]

Electricity availability and regional constraints

Electricity availability varies materially across the United States; national averages can therefore obscure constraints at the project level. A project needs deliverable capacity at a particular site, on a particular timetable, under an acceptable contract. Access to fibre, customers, skilled labour, equipment and local approvals matters alongside the energy price.

Lawrence Berkeley National Laboratory's June 2026 update estimates a 649 TWh reference case for US data-centre electricity consumption in 2030, with a combined uncertainty range of 521–843 TWh. Its reference case corresponds to 11.8% of US electricity use. The estimate models electricity use across US data centres, including workloads beyond AI. [9]

The IEA and Lawrence Berkeley National Laboratory use different datasets and modelling approaches, so their estimates are best read as separate scenarios. Together, they show why project economics depend on a range of possible electricity demand.

For investors, a bottleneck can have opposing effects. A secured connection may make a site valuable; the same scarcity can delay construction, increase costs and undermine a customer's ability to occupy it. An announced megawatt pipeline is therefore weaker evidence of operating performance than capacity that is contracted, financed, connected and generating cash.

Selected markets elsewhere in the Americas

The analysis also considers selected markets elsewhere in the Americas. Aligned's disclosed footprint includes sites in Brazil, Mexico and Chile as well as US markets. Its acquisition shows that capital is reaching several national markets, each with distinct demand, financing conditions and return prospects. [16]

A project in Canada or Latin America requires a country- and site-specific assessment of power contracts, local demand, currency exposure, taxation, legal enforceability and cash repatriation. A low-cost site becomes a sound investment when workloads can be routed there economically and customers sign bankable contracts.

6. Europe and its distinct routes into the investment cycle

Europe's AI infrastructure market follows a different path from the US hyperscaler market.

The investment opportunity includes globally active equipment suppliers, local data-centre and energy assets, industrial adoption and efforts to develop more sovereign computing capacity. These channels can grow together, but they create different claims on future cash flows.

European suppliers and global demand

ASML and Schneider Electric illustrate why headquarters, end-demand and investment exposure must be kept separate. A European-listed supplier can benefit from a US project, giving investors exposure to the global investment cycle through a European company. [4, 6]

For a euro-based investor, these suppliers broaden the research universe. Their earnings can still move together when both depend on the same customers expanding capacity.

A euro quotation leaves the investor exposed to the currencies in which the company earns revenue, pays costs and finances its operations. The analysis should identify where the company earns revenue, incurs costs and finances its operations, and how it manages those exposures.

Sovereign computing and the commercial test

On 30 July 2026, the European Commission launched a call for up to seven AI Gigafactories. It described up to €10 billion of EU and national support intended to unlock at least €20 billion of private investment. The €30 billion figure combines planned public support with the private investment the programme aims to mobilise. [10]

The policy rationale includes access to computing and greater technological autonomy. The investment question remains narrower: who receives the funding, what must be built, who pays to use it, and what commercial return remains after obligations and operating costs?

Subsidies can make a project viable, but they can also encourage capacity that earns weak returns without continuing support. Procurement conditions, ownership, pricing, access obligations and the allocation of construction risk deserve as much attention as the headline budget.

Sovereignty also has several meanings: data location, operating control, legal jurisdiction, access to hardware and the ability to keep a service running. Microsoft's European sovereign-cloud initiatives show that US suppliers can participate in this demand too. A European facility may still rely on American technology and global supply chains. [12]

National differences in infrastructure conditions

France and Ireland illustrate why a single European average is insufficient. RTE reports that more than 95% of French electricity generation was low-carbon in 2025. Ireland's CSO reports that data centres accounted for 23% of metered electricity consumption in 2025. Those facts describe different opportunities and system pressures; neither establishes the return on a particular project. [13, 14]

For France, the analysis should examine whether a project's apparent energy advantage is actually available at the relevant site under a durable contract. For Ireland, particular attention should be paid to connection requirements, energy availability and the ability to expand. Other European locations require their own assessment of supply, connectivity, latency, water, planning and customers.

The United Kingdom, Norway and Switzerland follow separate frameworks from the EU, so each project's jurisdiction matters for legal and commercial analysis.

Adoption and productivity beyond infrastructure construction

ECB analysis in March 2026 described a widening digital-investment gap with the United States and identified financing and adoption obstacles in Europe. The ECB's digital-investment measure includes technologies beyond AI, making it a broader indicator of Europe's technology investment. [11]

This suggests that European opportunity may arise both from supplying infrastructure and from using AI to improve existing businesses. A manufacturer that reduces downtime or a software provider that improves customer productivity could capture benefits without owning a new data centre.

Adoption across manufacturers and software businesses also matters: when AI produces measurable economic value for those users, they are more likely to pay for the infrastructure supporting it.

7. Asia and the international supply chain

A two-region story would miss crucial dependencies. TSMC's manufacturing role and ASML's equipment role demonstrate how the value chain crosses borders before a server reaches an American or European facility. NVIDIA's disclosures also identify reliance on third parties and export restrictions among its risks. [1, 4, 5]

Four locations should be mapped separately: the customer's location, the company's domicile, the production footprint and the physical infrastructure. They can all differ. A portfolio divided between US and European stock exchanges can still depend on the same manufacturing and packaging chain in Asia.

This creates an additional valuation question. Geographic diversification is only useful if it reduces the risks the investor actually owns. Adding a new listing country while keeping the same customers, suppliers and financing exposure may change the appearance of the portfolio more than its economics.

Export rules, tariffs or interruptions to advanced manufacturing can affect availability, cost and market access. The analysis should test how reliably each company can turn component supply into finished, usable capacity, including how quickly it can source alternatives.

The United States combines deep hyperscaler demand with substantial infrastructure investment, while Europe offers industrial suppliers, local assets and policy-backed computing. Investors need to trace demand, scarce capabilities and risk across both markets.

8. Competitive advantage and pricing power

The analysis centres on businesses with capabilities that remain difficult to replace as the construction cycle matures.

That could be a technology advantage, a software ecosystem, a qualified component, a reliable service network or a scarce, usable connection to the grid.

A shortage becomes a durable advantage when competitors face long lead times and customers have few practical alternatives.

If a product becomes scarce and profitable, suppliers have an incentive to expand. Customers have an incentive to redesign systems, negotiate harder or develop alternatives. The question is whether the original supplier can protect its economics when those responses arrive.

ASML's October results illustrate the distinction between current strength and future assumptions. The company reported €11.7 billion of third-quarter sales and said customers were raising equipment-demand forecasts, primarily because of their expectations for AI investment. Management's outlook reflects stronger customer expectations for AI investment, while future returns will depend on orders translating into delivered equipment and cash flow. [4]

Schneider Electric's first-half results offer a different angle: the company reported strong demand led by data centres, alongside growth across its other end-markets. Eaton's second-quarter release also reported accelerating orders and backlog. [6, 7]

Investors can gain exposure through several parts of the build-out. The analytical challenge is to establish how much of a company's profit depends on this activity and how much future growth is already reflected in its valuation.

Illustrative company research agenda

The following table turns the sector thesis into company-level questions across different parts of the value chain.

The company positioning is informed by the primary releases cited earlier. The investment tests are the author’s analysis. A company can pass the business-quality test and still fail the valuation test. [1, 2, 4, 5, 6, 7]

9. Revenue, cash flow and reinvestment

For an equipment supplier, the analysis should trace orders into deliveries, margins and cash receipts. A growing backlog is encouraging only if projects proceed, customers pay and the company can deliver profitably.

For a cloud platform, the analysis should trace investment into usable capacity, customer demand and cash generation. High utilisation with firm pricing is very different from filling a new facility through aggressive discounts.

For a data-centre owner, relevant variables include the tenant, contract length, power availability, maintenance obligations and refinancing schedule. A long lease can improve revenue visibility, while tenant default and upgrade costs remain important risks.

For a utility, more electricity demand can require more investment before shareholders receive additional cash. Regulated returns, cost recovery and financing matter as much as the growth in consumption.

Across these businesses, the central question is:

Does the company earn an adequate return on the next euro it invests?

Accounting earnings and cash flow can tell different stories. A capital-intensive company may report growing profits while much of its operating cash is absorbed by new equipment. Depreciation spreads an asset's cost over its accounting life; the cash may already have left the business.

Expansion spending may recur for years when frequent upgrades are needed to remain competitive.

A robust assessment should examine returns on invested capital, cash conversion, dilution and debt together. Margins should be assessed alongside returns on capital, cash conversion, dilution and debt.

10. Illustrative project economics and sensitivity analysis

To make the economics concrete, consider a simplified fictional computing operator.

Assume €100 million of equipment and associated infrastructure, eight million standardised computing-unit hours of annual saleable capacity before utilisation, €15 million of annual cash operating costs and a four-year economic equipment life with no residual value.

The example uses a €25 million annual capital-recovery allowance to reflect that assumed economic life. This allowance represents the capital that must be recovered over the equipment's useful life. The cash leaves the business when equipment is purchased and again when it is replaced, so the allowance helps estimate how much operating cash can be distributed over time.

The weaker case assumes largely fixed cash operating costs to isolate operating leverage; a real model would separate variable power costs from fixed costs. It also excludes working capital, construction delays, ramp-up, inflation, taxes and financing. This simplified case shows how sensitive project economics can be to operating leverage; a full valuation would also model these omitted factors.

At the stronger case's €8 price, covering €15 million of cash costs and the €25 million recovery allowance requires 62.5% utilisation. At €6, the same test requires about 83.3%. A lower selling price and lower occupancy can therefore interact much more severely than either assumption alone suggests.

Equipment life is equally important. If useful life is only three years, the annual allowance becomes about €33.3 million and the stronger case's surplus falls to approximately €2.9 million. If life is six years, the allowance falls to €16.7 million and the surplus rises to approximately €19.5 million. Neither assumption should be chosen merely to make the investment work.

Now add €60 million of debt at 8%. Annual interest alone is €4.8 million before principal repayment. The strong four-year case's remaining surplus falls to €6.4 million; the weaker case is already inadequate on this economic test. Loan maturity, amortisation, covenants and refinancing can create a cash problem earlier than the accounting results suggest.

A contract reduces uncertainty when the customer can pay, the volume commitment is enforceable and termination clauses allocate risk clearly. Backlog should therefore be tested against customer credit, contract terms and the likelihood of delivery.

A project-level assessment should build a dated cash-flow model, distinguish the building's life from the servers' life, model replacement expenditure explicitly and calculate both unlevered and equity returns. Equipment replacement is a recurring lifecycle cost and belongs in the return model alongside initial construction spending.

11. Investor positioning and capital allocation

Investor exposure to AI infrastructure takes several forms. Broad equity-index investors, active growth funds, pension infrastructure teams and private lenders participate through different instruments and seek different sources of return.

The discussion distinguishes current investment exposure, completed transactions and announced intentions to raise or deploy capital. It maps observable activity across public markets, private capital and lending, using disclosed information on each channel.

Public-market exposure

Market-cap-weighted equity investors gain exposure as successful companies grow into larger index weights. Active investors may add specific positions through semiconductor suppliers, cloud platforms and electrical equipment. A dedicated semiconductor or AI ETF may simplify implementation. Its holdings reveal how much it overlaps with existing funds, where it is concentrated and which part of the value chain it represents.

The October IMF report documents growing leverage around AI-related exposures and greater cross-border holdings of US equities. In its leveraged ETF data, single stocks made up 28% of exposures, up from 16% in 2024. [20]

AI exposure now warrants portfolio-wide monitoring. Public disclosures offer useful indicators, though institution-wide positioning is difficult to summarise in a single figure.

Private equity and infrastructure capital

BlackRock's AI Infrastructure Partnership originally described a target to mobilise $30 billion of equity capital and up to $100 billion including debt, with a primary focus on the US and selected partner markets. These figures describe the partnership's initial fundraising ambition. [15]

There is also completed activity. In July 2026, AIP, MGX and GIP completed the Aligned acquisition, with an announced enterprise value of approximately $40 billion. The consortium separately committed $5 billion of growth capital. The enterprise value reflects the acquired business, while the separate $5 billion growth-capital commitment is intended to fund expansion. [16]

Brookfield's November 2025 programme announced a $10 billion equity-fund target within a broader potential $100 billion AI-infrastructure programme. The programme shows the scale of institutional ambition, combining an equity-fund target with a broader financing plan. [17]

These programmes combine operating assets, development capability, long-term capital and relationships with technology customers; the returns will depend on execution and the price paid.

Private credit and the financing of the build-out

BIS research describes a broad shift towards debt as AI-related investment needs grow, with private credit playing an increasing role. Public data have yet to isolate the amount of private credit funding AI infrastructure. Several financing channels support the expansion, including company cash flow, public bonds, bank lending, private placements, asset-backed securities, leases and private credit. Project funding can include several of these channels, and public announcements often describe different stages of the same financing. [18, 21, 22]

The scale of private credit overall has grown rapidly. BIS estimates that global outstanding direct loans rose from around $100 billion in 2010 to more than $450 billion in 2019 and nearly $2.5 trillion by 2025. A separate BIS analysis of PitchBook records for US-based borrowers finds that loans to technology firms increased from about $22 billion in 2010 to $127 billion in 2019 and more than $1 trillion in 2025. Technology lending represented close to 45% of the direct-loan amounts in that dataset by 2025, compared with about 22% in 2010. The technology category includes software and other technology businesses alongside AI-related firms. The global direct-loan and US technology-loan figures use different coverage, so they describe separate trends. [21]

Private credit and technology lending have expanded very rapidly. Public data have yet to quantify private-credit funding specifically for AI infrastructure, so claims of exponential growth in that segment remain unverified. The most detailed technology-loan data cover US borrowers; comparable Europe-wide figures remain limited. [21, 23]

Some AI infrastructure is financed through dedicated joint ventures or special-purpose vehicles. These entities can acquire or develop data centres, raise equity from sponsors and borrow through private placements. A hyperscaler may take a minority stake, sign a long-term lease or capacity-purchase agreement, and sometimes provide guarantees. Private credit funds and other institutions hold some of the debt, while banks may provide funding lines. This structure can move borrowing away from the hyperscaler's reported balance sheet, but the lease payments and guarantees still create economic commitments. BIS describes this as “shadow borrowing” and notes the links it creates among technology firms, private-credit funds, insurers and banks. [22]

A lender's return is usually contractual and capped, while its downside depends on the borrower's ability to pay and the recovery value of collateral. Lending against a building, a diversified platform or rapidly depreciating chips can produce very different outcomes. Long-duration loans may suit insurers and pension funds with similar liabilities, but a loan that remains outstanding beyond the useful life of the financed equipment can create a maturity mismatch. The credit assessment turns on borrower cash flow, collateral quality and the loan's position in the capital structure. The Financial Stability Board also identifies leverage, concentration, liquidity mismatches and links with banks, insurers and private-equity firms as areas where data and stress testing remain difficult. [23]

Summary of investor channels

Investors often use several of these channels at once; a sovereign fund or large asset manager may operate across the full set.

Investor positioning matters when it changes the valuation, liquidity or operating risk of a portfolio exposure.

12. Valuation and shareholder returns

Valuation is the point at which an attractive industry narrative must be tested against the price paid for a security.

Consider a hypothetical company that earns €1 per share and trades at €40. Its price-to-earnings multiple is 40.

If earnings double to €2 per share over five years but the valuation falls to 20 times earnings, the implied share price remains €40.

Thus, substantial earnings growth can coincide with no price appreciation when the valuation multiple contracts sufficiently.

Here is the same simplified investment under three outcomes:

The scenario starts at €40 per share and calculates annualised price returns over five years, excluding dividends, fees, taxes and currency movements.

The example shows how the entry valuation determines the growth needed to generate a given shareholder return.

A company-level valuation should therefore work backwards from the current price to identify the revenue growth, margins and reinvestment it implies, and then compare those assumptions with a range of plausible outcomes.

Company-level valuation remains the next step after identifying promising parts of the value chain.

13. Trends and conditional outlook

The base case for the next 12–24 months is continued expansion accompanied by greater differentiation among companies and projects. Over a longer horizon, the central question is whether demand, productivity and cash generation can justify the accumulated capital stock.

From announced capacity to operating returns

The IEA's 2026 update expects a further sharp increase in major technology-company capital expenditure during 2026; those figures are estimates and include spending beyond narrowly defined AI infrastructure. [3]

As initial capacity is delivered, investors are likely to focus more closely on occupancy, paid usage, unit pricing and returns on new assets. A company may still grow rapidly while its valuation falls if the incremental economics become less attractive than expected.

The decisive evidence will be whether customers pay enough, often enough, for the capacity to earn its cost of capital. Revenue commitments supported by durable customer cash flow are stronger evidence than demand that depends on the customer repeatedly raising fresh financing.

Shifting constraints across the infrastructure system

A shortage of accelerators can encourage attention to chips. When chips arrive, the constraint may move to networking, memory, power, cooling or the grid connection. A supplier can benefit while its customers are delayed elsewhere.

As bottlenecks move across the system, the research focus will shift with them. Some suppliers may lose pricing power as capacity expands, especially when valuations already assume prolonged scarcity.

A durable investment advantage depends on constraints that retain economic value after competitors respond.

Regional expansion and commercial models

In the United States, the central question is whether the scale of committed investment translates into attractive incremental returns. In Europe, the relevant indicators are both the conversion of policy programmes into operational capacity and the order-to-cash performance of suppliers serving global customers.

Sovereignty and customer requirements can support local facilities even when a remote alternative appears cheaper. Regional duplication can also reduce scale economies and increase the capital required to serve the same demand. Global suppliers serving several systems may capture more value than owners of every duplicated facility.

European opportunities span public funding, local demand and suppliers with global revenues, each on a different timetable. Strategic value to Europe and shareholder returns are related but distinct questions.

Financing conditions and company differentiation

The expected implication is that balance-sheet quality, customer credit and refinancing schedules will matter more as the required capital grows. That is consistent with the BIS's analysis of an expanding role for debt and private credit. [18]

During favourable conditions, external capital can allow projects to proceed without requiring technology companies to own every asset. Under weaker conditions, the same financing structures may expose fragile relationships among borrowers, customers and parties supporting revenue.

That suggests a possible divergence: established platforms and proven suppliers may continue to fund themselves while some smaller, more leveraged operators face a much higher hurdle.

Outlook across three time horizons

The cycle is likely to vary across chips, buildings, electrical equipment and software, with each layer facing its own demand and capacity pressures.

14. Scenario analysis

Scenario analysis helps identify which developments would strengthen or weaken the investment case across the AI ecosystem.

Efficiency complicates a simple bullish or bearish interpretation of demand.

The IEA notes that energy use per AI task has fallen sharply, while more demanding applications and growing adoption can increase total demand. [3]

Cheaper computing could make AI useful to more customers. It could also reduce the revenue earned from a particular unit of computing capacity. The outcome depends on how much additional demand the lower cost creates, and who passes the savings to whom.

As AI use grows, profits will depend on which infrastructure owners can turn additional demand into durable cash flow.

15. Monitoring priorities and further research

The most informative developments will be evidence about project economics and cash generation.

Monitoring should include cloud growth alongside capital spending and cash flow; supplier orders alongside deliveries and cash conversion; and data-centre contracts alongside actual power availability and occupancy.

Further indicators include new competing architectures, changes in export restrictions and the financing terms available to operators.

Schneider Electric is a suitable subject for the next company paper because it would allow analysis of AI-related demand within a broader industrial business, including whether end-market diversification supports cash generation. The company's results calendar lists third-quarter revenue reporting for 29 October 2026, providing a timely update for a November research note. [6]

A focused research question for that paper is: “Schneider Electric: Can the AI Power Boom Become Durable Shareholder Returns?”

16. Principal risks and transmission channels

Customer concentration comes first. Several suppliers may depend on investment decisions made by the same small group of buyers, even when they appear in different stock-market sectors.

Technological change comes next. An asset can be physically operational yet economically less competitive than a newer alternative. This matters especially where equipment is financed with debt that outlasts its strongest earning years.

Execution matters too. A planned data centre must still be built, connected to power and occupied before it can generate revenue; delays leave capital tied up in the meantime.

Financing creates another transmission channel. A project may have sensible long-term economics but insufficient cash to survive refinancing or a tenant problem. The analysis should examine guarantees, collateral, covenants and the relationship between the parties financing, selling and buying capacity.

Finally, industrial opportunity should be distinguished from investment enthusiasm. An electricity-equipment supplier can be a strong business and still become a disappointing investment if bought at a price that assumes exceptional growth indefinitely.

The central risk is that market prices assume exceptional execution and leave little room for disappointment.

Bubble risk and the possibility of overinvestment

Bubble risk depends on whether current prices and construction plans require unusually high returns or demand that paying customers have yet to demonstrate.

The IMF's October 2026 Global Financial Stability Report identifies weaker or less durable AI investment and profitability as potential triggers for a correction in AI-related asset prices, with leverage and concentrated exposures amplifying the effects. [20] BIS research also describes elevated valuations, investment spending outpacing the cash flows of some large firms, and greater reliance on debt and private credit. [24] Together, these conditions make overinvestment and market correction important scenarios for investors to assess.

A practical assessment should compare four things: the earnings investors expect with cash flow already produced; announced capacity with connected and occupied capacity; contracted demand with end-user ability to pay; and debt obligations with the useful life and refinancing needs of the assets. A widening gap between plans and realised economics would strengthen the overinvestment concern. Continued paid adoption, high utilisation and cash returns that cover replacement spending would weaken it.

A correction in listed shares and a credit loss are different outcomes. Share prices can fall sharply while a strong borrower continues to repay its loans. Credit losses arise when weaker demand reduces cash flow, refinancing fails or collateral recovery falls below the loan balance. Listed prices adjust quickly; private-loan impairments and valuations often emerge more gradually.

Financial interconnections and the transmission of stress

The BIS's October 2026 study finds that 46.4% by value of AI-to-AI investment deals in its 2021–2025 sample also involved commercial supply-chain relationships between investor and investee. The finding points to links between financing and commercial relationships within part of the AI ecosystem. [19]

Such relationships can secure inputs and align incentives. They also require careful analysis. If a supplier finances a customer that promises to purchase its products, the central question is how much demand would survive without further support and whether the same economic cash flow is being described multiple times across the ecosystem.

A customer's distress could weaken the supplier's investment and its expected sales simultaneously. A guarantee can protect one party while transferring the contingent liability to another. The October IMF report highlights these interdependencies and the potential amplification from leverage. [20]

A risk assessment should map the major counterparties and stress a common event: the customer delays spending, asset values fall, lenders demand more protection and financing becomes more expensive. These risks can interact across legal entities and compound under stress.

Inflation, interest rates and the cost of capital

An asset can meet its revenue plan and still be worth less if investors demand a higher return. A long-lived building, a power project and a high-growth equity all depend on the discount rate, although their cash-flow profiles differ.

This connects the industry paper to Macro \#06: higher construction costs, more expensive debt and lower terminal valuations can occur together as financial conditions tighten.

Environmental and community constraints

Water requirements, noise, land use and the distribution of grid costs can affect local acceptance. Annual low-carbon energy purchases can leave some hours unmatched with a facility's actual electricity needs.

For a project claiming an energy advantage, the assessment should examine the actual contract, connection timetable, backup arrangements and obligations to the local system. The contract and connection timetable determine whether the promised energy service can be delivered at the stated cost.

17. Conditions that would change the assessment

The investment case for a company would strengthen if growing customer demand produced durable cash flow, returns on new capital remained attractive, customer dependence declined and the valuation left room for slower growth.

The investment case would weaken if orders rose while cash conversion deteriorated, customer financing became essential to maintaining sales, or capacity expansion consistently outran evidence of profitable end-user demand.

Portfolio decisions should follow evidence about the investor's objectives, how the holdings behave and the returns expected for the risks taken.

A rigorous investment process should specify in advance which evidence would invalidate the thesis. Without such criteria, disappointing results may be repeatedly rationalised as temporary, while new announcements are treated as confirmation.

Indicators for ongoing review

These indicators should be reviewed together. Order books, customer financing, cash balances and project returns should be assessed together.

18. Implications for the four model portfolios

Capital Preservation

A dedicated AI-infrastructure position would require an unusually strong justification in this portfolio, whose primary objectives are capital preservation and liquidity.

The global-equity allocation already provides exposure to economic growth, so its holdings should be reviewed before adding a thematic position. Utilities, equipment suppliers and data-centre owners can still carry substantial technology-cycle and valuation risk despite owning physical assets or providing essential services.

Early-Career Growth

For this investor, broad diversification and regular contributions remain the preferred foundation.

A global equity fund can participate as businesses grow and index weights evolve. A separate AI fund could duplicate existing holdings and concentrate risk without solving an unmet portfolio need. Before adding a position, its incremental contribution should be demonstrated relative to the current global-equity exposure.

Balanced Entrepreneur

The €3 million venture reserve has a clear purpose and remains outside this investment question.

Within the €22 million strategic portfolio, any proposed position should be assessed to determine whether it adds a distinct source of return or simply increases the same technology and investment-cycle risks already present elsewhere. Any direct equity would also have to respect the project's single-stock limit and the investor's wider business exposure.

Family Stewardship

The family's long horizon and substantial capital allow it to investigate both listed and private infrastructure opportunities.

Any private project would require separate analysis of tenant quality, power supply, technology exposure, leverage, capital calls and exit assumptions. It would sit within the future private-market programme, separate from the liquid transition portfolio.

Across all four mandates, existing exposure should be measured before recommending more.

19. Conclusion

The analysis began with the proposition that rising AI adoption would increase demand for infrastructure and benefit its suppliers. The evidence reviewed here indicates that this relationship is conditional. The relevant investment questions are which companies can convert spending into shareholder cash flows, whether those cash flows can persist through a changing investment cycle, and whether the current valuation compensates investors for the risks involved.

The next steps are to assess existing portfolio exposure, research individual companies in greater depth and test valuations across a range of outcomes before allocating capital. AI may transform economic activity, but the investment consequences will depend on how that transformation affects the cash flows and valuations of the securities held.

20. Research scope and disclosure

Research cut-off: 17 October 2026. It uses public primary sources available by that date; company outlooks and IEA projections are identified as forward-looking. Business-model comparisons, scenarios and portfolio implications are the author's analysis.

This paper focuses on sector economics and strategic portfolio fit. Follow-up company research and live portfolio reviews will add current share-price valuations and portfolio look-through.

Interpreting the evidence

Company releases provide reported results, while government and agency documents describe programme terms and energy estimates. BIS, Financial Stability Board and IMF research informs the financing and financial-stability analysis. Targets, commitments, acquisition values and deployed capital measure different stages of investment and are treated separately.

The case studies show selected regional examples. Private-credit totals and technology-loan figures use different datasets; technology lending also includes businesses beyond AI, so the AI-specific amount remains unclear. Each forecast and historical figure is labelled with its source date.

[1] NVIDIA, Second Quarter Fiscal 2027 results, quarter ended 26 July 2026. Source ↗

[2] Microsoft, Fourth Quarter and Fiscal Year 2026 results, year ended 30 June 2026. Source ↗

[3] International Energy Agency, Key Questions on Energy and AI, 2026, executive summary. Source ↗

[4] ASML, Third Quarter 2026 results, 14 October 2026. Source ↗

[5] TSMC, Third Quarter 2026 results and supporting materials, 15 October 2026. Source ↗

[6] Schneider Electric, Half-Year 2026 results, 30 July 2026, and financial-results calendar. Source ↗

[7] Eaton, Second Quarter 2026 results. Source ↗

[8] International Energy Agency, Energy and AI, 2025, executive summary; historical regional electricity-consumption shares. Source ↗

[9] Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, published June 2026. Source ↗

[10] European Commission, AI Gigafactories call, 30 July 2026. Source ↗

[11] European Central Bank, AI and the euro area economy, 23 March 2026. Source ↗

[12] Microsoft, One year on: Progress on our European digital commitments, 29 April 2026. Source ↗

[13] RTE, Annual Electricity Review 2025, generation. Source ↗

[14] Central Statistics Office Ireland, Data Centres Metered Electricity Consumption 2025, 7 July 2026. Source ↗

[15] BlackRock, AI Infrastructure Partnership expansion, 19 March 2025. Source ↗

[16] Aligned Data Centers, acquisition completion and growth-capital commitment, 21 July 2026. Source ↗

[17] Brookfield Asset Management, AI Infrastructure Program launch, 19 November 2025. Source ↗

[18] Aldasoro, Doerr and Rees, Financing the AI boom: from cash flows to debt, BIS Bulletin 120, 7 January 2026. Source ↗

[19] Bank for International Settlements, Circular relationships among AI firms, BIS Bulletin 137, 1 October 2026. Source ↗

[20] International Monetary Fund, Global Financial Stability Report, October 2026, Chapter 1, especially pages 3–4 on leverage and AI financing. Source ↗

[21] Bank for International Settlements, “Financing the digital economy: the role of private credit,” BIS Quarterly Review, September 2026. Source ↗

[22] Bank for International Settlements, “Financing the AI infrastructure boom: on- and off-balance sheet borrowing,” BIS Quarterly Review, March 2026. Source ↗

[23] Financial Stability Board, “Report on Vulnerabilities in Private Credit,” 6 May 2026. Source ↗

[24] Bank for International Settlements, “Artificial intelligence, growth and financial stability: challenges for central banks,” speech, 10 September 2026. Source ↗