Financing the AI Boom From Cash Flows to Debt | BIS

The AI Investment Boom Reaches Macroeconomic Scale

The artificial intelligence investment surge has crossed a critical threshold. A BIS Bulletin by Aldasoro, Doerr, Gambacorta, and Shin reveals that AI-related capital expenditure has become macroeconomically significant, reaching a scale that demands attention from financial regulators, central bankers, and investors alike. By mid-2025, expenditures on IT manufacturing facilities and data centres alone equalled 1% of US GDP — a remarkable concentration of investment in a single technological wave.

When broader IT equipment and software spending are included, total IT-related investment has reached 5% of GDP, surpassing even the peak of the dot-com era in 2000. But the BIS analysis identifies a fundamental structural difference: unlike the dot-com boom, which was driven by firms using IT products, the current wave is propelled by firms producing AI infrastructure — semiconductor fabrication facilities, hyperscale data centres, and the computing hardware that powers large language models and other AI systems.

This distinction matters enormously for risk assessment. Infrastructure producers face different economics than technology consumers: longer investment horizons, higher capital intensity, greater exposure to construction and execution risk, and revenues that depend on sustained demand from downstream AI applications. The BIS Bulletin documents how this structural difference shapes the financing dynamics of the AI boom and its potential vulnerabilities.

How AI Spending Now Drives US GDP Growth

The macroeconomic significance of AI investment extends beyond its absolute scale to its role as a driver of economic growth. From a negligible contribution before 2022, expenditures on semiconductor manufacturing facilities and data centres have contributed an average of 0.4 percentage points per year to US GDP growth over the subsequent three years. In recent quarters, total IT investment has accounted for nearly half of US GDP growth, helping offset the negative effects of trade tariffs and other headwinds.

This concentration of growth dependence on a single technological wave creates both opportunity and vulnerability. On the positive side, AI investment is generating employment in construction, manufacturing, engineering, and operations — real economic activity that extends well beyond the technology sector. Analyst forecasts indicate annual data centre spending alone could increase by $100-225 billion over the next five years, potentially pushing data centre spending from 0.5% to 0.8-1.3% of GDP.

The vulnerability lies in the dependence itself. If the AI investment cycle were to slow or reverse — due to disappointing returns, regulatory constraints, energy supply limitations, or a shift in technological trajectory — the growth impact would be felt immediately and broadly. An economy deriving nearly half its growth momentum from a single investment category is exposed to a concentration risk that warrants careful monitoring by the Federal Reserve and other central banks.

The Financing Shift From Cash Flows to External Debt

The BIS Bulletin identifies the most consequential financial development of the AI boom: a structural transition in how leading technology firms fund their investments. Historically, companies like Alphabet, Amazon, Meta, Microsoft, and Oracle operated with substantially lower debt-to-asset ratios than other firms, relying on highly profitable operations to generate the cash flows needed for investment. This self-funding model limited the transmission of technology risk to the broader financial system.

That model is breaking down. The scale of current and anticipated AI capital expenditures is so vast that it is exhausting firms’ internal financing capacity. Free cash flows have recently fallen below capital expenditures in absolute terms — a pivotal moment that forces these companies to seek external funding. The BIS describes this as a qualitative change in the risk profile of the AI boom, not merely a quantitative one.

Equity financing is considered suboptimal at the current juncture. Volatile AI valuations create narrow issuance windows, while new stock sales are dilutive and costly for the long-dated, asset-heavy infrastructure projects that dominate AI investment. The result is a decisive pivot toward debt — through bond markets, where tech companies have reportedly achieved record issuance levels, and increasingly through private credit and other lending arrangements.

This transition fundamentally alters who bears the risk of the AI investment cycle. When firms funded investment from cash flows, risk was contained within the firm and borne primarily by equity holders. As firms shift to debt, risk distributes across the financial system — to banks, private credit funds, bond investors, and ultimately the institutional investors (pension funds, insurance companies, sovereign wealth funds) that allocate capital to these vehicles. If AI expectations prove unfounded, the consequences will reach far beyond technology company shareholders.

Private Credit Surges Into AI Infrastructure

Among the most striking findings in the BIS Bulletin is the explosive growth of private credit as a funding source for AI infrastructure. Outstanding private credit loans to AI-related sectors have surged from near zero to over $200 billion, with their share of total private credit volumes rising from less than 1% to nearly 8%. The BIS projects that outstanding private credit to AI firms could reach $300-600 billion by 2030, making it one of the most significant sectoral concentrations in private credit history.

The growth metrics are remarkable across multiple dimensions. In 2025, private credit funds originated over $40 billion in loans to AI-related companies, compared with roughly $3 billion in 2010 — more than a tenfold increase. Approximately 20% of all private credit funds now have some AI exposure, up from 5% in 2010. For the average fund with AI exposure, AI-related loans account for about 5% of total volumes, up from near zero in 2010.

Private credit’s structural characteristics make it particularly well-suited to AI infrastructure financing. Data centre projects involve specific risks — construction timelines, power availability, tenant concentration, and technology obsolescence — that may not fit neatly into standardized bank lending or bond market frameworks. Private credit offers bespoke covenant structures, certainty of execution, speed of deployment, and flexible renegotiation capabilities that can be tailored to these idiosyncratic project risks.

The loan characteristics reveal important details. Average AI-related private credit loans are nearly twice the size of non-AI loans ($169 million vs. $90 million), reflecting the capital intensity of data centre projects. Maturities are similar at approximately 4.7 years, and secured shares are virtually identical at around 46-48%. The Financial Stability Board has highlighted the rapid growth of private credit more broadly as an area requiring enhanced monitoring.

The Equity-Debt Pricing Disconnect in AI Markets

Perhaps the BIS Bulletin’s most analytically important contribution is identifying a troubling disconnect between debt market pricing and equity valuations for AI companies. Loan spreads on private credit to AI firms are virtually identical to those charged to non-AI borrowers — 6.2 versus 6.1 percentage points — even after controlling for loan and borrower characteristics. This implies that lenders view AI-related loans as no riskier than average private credit exposures.

Yet equity markets simultaneously price AI companies at valuations that imply outsized future returns. This “schism” between two signals from two different markets assessing the same underlying economic phenomenon creates a logical tension. Either the expected returns will materialize — justifying equity valuations and suggesting debt is appropriately or even attractively priced for the quality of the credit — or they will not, suggesting debt is underpriced for the actual risk and equity is overvalued.

The BIS warns that this pattern echoes dynamics observed in previous credit cycles, where spread compression and risk underpricing preceded significant market dislocations. When lenders charge AI borrowers the same rate as non-AI borrowers despite the concentration risk, execution risk, and technology dependency inherent in data centre projects, it raises the question of whether market participants are systematically underpricing AI risk precisely as their exposures grow most rapidly.

Data Centre Economics and the Cost of AI Infrastructure

Understanding the economics of data centre construction and operation is essential for assessing the sustainability of the AI financing boom. The BIS cites a standard industry rule of thumb: the physical structure accounts for approximately one quarter of total data centre spending, with IT equipment — servers, networking gear, cooling systems, and power infrastructure — comprising the remaining three quarters.

This cost structure has important implications for financing risk. The physical infrastructure component represents a relatively durable asset with long useful life, making it suitable for longer-term debt financing. However, the IT equipment component is subject to rapid technological obsolescence as AI hardware evolves. A data centre built today with current-generation GPU servers may require significant equipment refresh within 3-5 years as next-generation chips offer dramatically improved performance per watt.

Power availability has emerged as a binding constraint on data centre expansion. AI training and inference workloads are extraordinarily power-intensive, and hyperscale data centres require reliable access to hundreds of megawatts of electricity. Securing power commitments, navigating grid interconnection processes, and increasingly investing in dedicated power generation (including nuclear and renewable sources) adds complexity, cost, and execution risk to data centre projects.

The geographic concentration of data centre investment creates regional economic effects and risks. Certain markets — Northern Virginia, Texas, the Pacific Northwest — attract disproportionate investment due to power availability, connectivity, and regulatory environment. This concentration means that a slowdown in AI investment would have outsized impact on specific regional economies and real estate markets.

Hidden Leverage and Off-Balance-Sheet AI Risks

The BIS Bulletin flags an often-overlooked dimension of AI financing risk: the use of off-balance-sheet structures and circular financing arrangements that can mask the true extent of leverage in the AI ecosystem. Some financing structures being established to support AI investments move debt off the balance sheets of the primary technology companies, creating the illusion of lower leverage than actually exists.

Joint ventures and special-purpose vehicles for data centre development allow technology companies to share costs and risks while keeping significant liabilities off their consolidated balance sheets. While these structures can serve legitimate risk-sharing purposes, they also reduce transparency and can make it difficult for investors and regulators to assess the total financial exposure of the AI ecosystem.

Circular financing within the AI ecosystem presents a more subtle concern. When AI companies invest in each other, purchase each other’s products, and establish reciprocal commercial relationships, the reported revenues and profitability that support current valuations and credit assessments may be less robust than they appear. If a significant portion of AI revenue is generated by transactions within a closed ecosystem, the apparent growth trajectory may overstate genuine economic value creation.

The BIS pointedly observes that “leverage does not disappear by being out of sight.” This warning carries particular weight given the historical precedent of off-balance-sheet vehicles playing central roles in financial crises — from Enron’s special purpose entities to the structured investment vehicles that amplified the 2008 global financial crisis. The opacity of current AI financing structures warrants proactive supervisory attention before potential problems become systemic.

Historical Investment Booms and Lessons for AI

The BIS contextualizes the current AI investment boom within a comparative framework of previous investment surges, providing essential perspective on both the scale of current activity and the likely consequences of its eventual unwinding. At approximately 1% of GDP for data centres and semiconductor facilities, the AI boom is similar in scale to the US shale boom of the mid-2010s, roughly half the size of the dot-com era IT investment surge, and less than one-fifth the size of Australia’s mining boom or Japan’s 1980s commercial property boom.

While the relatively modest scale might seem reassuring, the historical evidence on investment boom unwindings is sobering. The end of previous investment booms was associated with GDP growth slowdowns averaging more than 1 percentage point. The largest contraction occurred after the US dot-com boom — notably, despite its comparatively modest absolute size — suggesting that technology investment booms may have disproportionate unwinding effects due to the interconnection between technology investment, technology company valuations, and broader financial market sentiment.

Perhaps the most challenging finding for AI optimists is the BIS observation that there is “little evidence of investment booms translating into a sustained increase in GDP growth over the medium term, even if, like the US dot-com boom, they are driven by technological advances.” This does not mean that AI technology is without value — the internet, born of the dot-com era, ultimately proved transformative. But it suggests that the investment phase and the productivity payoff phase may be separated by a potentially painful adjustment period.

The historical pattern for technology investment cycles follows a recognizable sequence: initial euphoria, escalating investment, diminishing returns on marginal projects, a correction that punishes overleveraged participants, and eventually a more sustainable growth path for the underlying technology. The question is not whether AI is valuable, but whether current investment levels and financing structures are calibrated to the realistic timeline for that value to materialize.

Financial Stability Implications and Regulatory Response

The BIS Bulletin generates several urgent recommendations for financial regulators and central banks. The rapid growth of private credit as a funding source for AI infrastructure, combined with its relative opacity compared to public markets, creates an immediate need for enhanced monitoring. Regulators require better data on private credit fund exposures, leverage, interconnectedness with the banking system, and the specific terms and covenants of AI-related loans.

Stress testing frameworks need updating to incorporate AI investment scenarios. Given that total IT investment now accounts for nearly half of US GDP growth, scenarios modeling a sharp pullback in AI capital expenditure should be incorporated into standard financial stability assessments. These scenarios should model effects on GDP growth, equity and credit markets, and contagion pathways through institutional investor portfolios.

The equity-debt pricing disconnect warrants active supervisory attention. If lending standards — including collateral valuation, covenant protections, and concentration limits — are not appropriately calibrated for the unique risks of AI infrastructure financing, the consequences of an AI investment downturn could be amplified through the credit channel. The Basel Committee on Banking Supervision may need to consider whether existing prudential frameworks adequately capture AI-related concentration risks.

Cross-border coordination is essential. While the BIS Bulletin focuses primarily on US data due to availability, the AI investment boom has global dimensions through semiconductor supply chains, international investor flows, and the cross-border operations of major technology companies. Central banks and financial regulators internationally need to assess their indirect exposures and coordinate monitoring efforts.

Assessing the Sustainability of the AI Investment Cycle

The fundamental question raised by the BIS Bulletin is whether the AI investment cycle is building toward sustainable productivity gains or inflating a financial bubble that will eventually deflate painfully. The honest answer is that both outcomes remain possible, and the financing dynamics documented in the Bulletin will determine which path materializes.

On the optimistic side, AI technology is demonstrably capable — large language models, computer vision systems, and autonomous agents are already delivering measurable productivity improvements in specific applications. If these gains scale broadly across the economy, the investment in AI infrastructure could prove justified, much as investment in telecommunications infrastructure during the 1990s eventually supported the mobile revolution of the 2010s.

On the cautionary side, the gap between AI investment and realized AI revenue is widening. The financing shift from cash flows to debt means that the consequences of disappointed expectations will be more broadly distributed across the financial system. The pricing disconnect between equity and debt markets suggests that at least one set of market participants is misjudging the risk-return profile. And the historical evidence on investment booms, even technology-driven ones, provides little comfort about smooth transitions from investment to productivity.

The BIS Bulletin serves as a measured but unmistakable warning: the AI investment boom has reached a scale and financing complexity that demands regulatory vigilance. The transition from internally funded investment to externally financed debt creates systemic transmission channels that did not exist when AI was funded from tech company cash flows. Whether those channels transmit prosperity or crisis depends on whether AI delivers returns commensurate with the capital being deployed — and on whether regulators act now to ensure the financial system can withstand the scenario where it does not.