Financing the AI buildout before the risks go off balance sheet

The AI boom is usually discussed as a contest over models, chips, and productivity. This column argues that it is also a large physical infrastructure and credit cycle. The policy priority is not to call the peak, but to measure who owns the assets, who owes the debt, and where contingent obligations will land if demand disappoints.

The policy debate around artificial intelligence has moved quickly from workplace exposure and productivity to a more immediate macroeconomic question: how large is the AI economy already, and where does it show up in the statistics and in the financial system? Korinek and McKelvey (2026) argue that conventional GDP accounts are missing much of the scale and speed of AI activity, estimating an AI sector that is already large but difficult to see in standard categories. Recent VoxEU columns make the same point from complementary angles: Carpinelli et al. (2026) show why data centres sit at the core of the AI economy, Coyle (2016) explains why digital activity can disrupt GDP measurement, and Filippucci et al. (2024) emphasise that aggregate productivity gains depend on diffusion well beyond a few frontier firms. These are measurement debates, but they are also policy debates. What cannot be measured in good times is hard to supervise in bad times.

In my recent paper (Van Nieuwerburgh 2026), I study a related but more tangible part of the AI transition: the financing of the physical buildout that makes frontier AI possible. AI is often described as an intangible revolution. In the near term, however, it is also an enormous physical capital cycle. Training and deploying frontier models requires specialised chips, high-density servers, cooling systems, substations, transmission access, fibre networks, and purpose-built data centres. The question is therefore not only whether AI will raise productivity. It is also who will supply the capital, who will own the infrastructure, and who will absorb losses if expectations about AI demand are revised.

The scale is unusual. A single, representative 200-megawatt AI training campus costs about $8.2 billion, with roughly two-thirds of the cost in IT equipment and one-third in the facility and associated power infrastructure. Scaling that benchmark to the US project pipeline, I estimate investment of about $10.3 trillion over 2025-2032, or 3.63% of GDP per year on average. The estimate starts from an announced data centre pipeline of over 500 GW and applies plausible horizon-dependent cancellation rates – reflecting supply chain constraints, mounting political opposition, and economic forces – to arrive at a 183 GW buildout between 2025 and 2032. This is a scenario, not a prophecy. But even a materially smaller buildout would rank with the largest investment waves in US history. Relative to GDP, the central estimate exceeds the canal, railroad, electrification, highway, and telecom and fibre booms.

Figure 1 Selected US capital-expenditure booms

Source: Van Nieuwerburgh (2016), Table 1.

The financing problem follows directly from the scale. For several years, the largest cloud platforms could largely fund data centres out of internal cash flow. That model is now under strain. Aggregate capital expenditures by Oracle, Microsoft, Amazon, Meta, and Alphabet rose from about $97 billion in 2020 to more than $400 billion in 2025 and are projected to exceed $800 billion in 2026. On current estimates, capex will exceed their combined operating cash flow for the first time in 2026. This does not mean these firms are weak credits. It means the marginal dollar of AI infrastructure increasingly comes from a broader capital stack.

Figure 2 Hyperscaler capital expenditure and operating cash flow

Note: Oracle, Microsoft, Amazon, Meta, and Alphabet.
Source:  Source: Van Nieuwerburgh (2016), Figure 1.

That capital stack is changing rapidly. Hyperscalers still own and finance many assets directly, especially IT equipment. But more investment is being financed through leases, joint ventures, project debt, private credit, securitisation, and special-purpose vehicles. Data-centre developers, REITs, infrastructure funds, and private equity investors provide equity capital. Banks, private credit funds, and structured-finance vehicles provide debt. The same logic is beginning to extend from buildings and power systems to GPUs and other hardware. The economic user of compute is increasingly separated from the legal owner of the asset and from the investor holding the credit exposure.

This separation has real benefits. It brings in deep pools of capital, allows risks to be shared, and can match long-duration contracted cash flows with investors that want them. It may also prevent the AI buildout from being limited by the balance-sheet capacity of a small set of technology firms. The problem is that risk is transformed, not eliminated. A facility or equipment vehicle may have high asset-level leverage relative to hyperscaler modest balance-sheet debt. Lease commitments, residual-value guarantees, purchase obligations, and support agreements can create contingent liabilities that are easier to see in footnotes than in headline leverage ratios.

The risk is not simply that AI demand might disappoint. It is that many exposures are correlated. Cash flows depend on a small number of tenants, on uncertain future demand for AI services, on whether today’s hardware remains valuable, on whether power and grid connections arrive on time, and on whether capital markets remain willing to refinance long-lived projects backed by fast-depreciating technology. The collateral is also unusual. Buildings and substations may have long lives, but GPUs and high-density servers can lose value quickly if models, chips, or computing architectures change.

This is why comparisons with past infrastructure and credit cycles are useful but should be handled carefully. The point is not that AI infrastructure is another subprime mortgage market or another telecom bubble. The point is that large investment booms become more fragile when leverage, maturity transformation, and opacity grow faster than the information available to investors and supervisors. The structured-finance literature has long shown that pooling and tranching can expand funding capacity while obscuring who bears tail risk (DeMarzo 2005, Coval et al. 2009). The AI buildout adds technological obsolescence and tenant concentration to that familiar financial logic.

What should policymakers do? The first answer is not to stop the buildout. There are plausible social returns from more compute, and the productivity upside discussed by Filippucci et al. (2024), Borri et al. (2026), and others may require substantial complementary capital. Nor is the right answer to declare systemic risk before the evidence warrants it. The right answer is to improve measurement and transparency while the capital structure is still forming.

Three steps would help. First, statistical agencies should develop better measures of AI infrastructure investment, separating domestic construction, imported IT equipment, power infrastructure, and cloud-service flows. Second, financial regulators should map the ownership and financing chain: hyperscaler obligations, leases, SPVs, private credit vehicles, bank exposures, securitisations, and guarantees. Third, disclosures should make contingent claims more comparable across firms, including uncommenced leases, residual-value guarantees, purchase commitments, construction guarantees, tenant concentrations, and refinancing needs and schedules. None of this requires deciding whether AI will fulfil its most optimistic forecasts. It requires recognising that the financing of the boom is now itself a macro-financial object.

AI may ultimately prove transformative. If it does, the infrastructure built today will be part of the capital stock that makes the transformation possible. But the path from investment to productivity is uncertain, and the path from productivity disappointment to financial loss runs through contracts, leases, and credit structures that are still poorly observed. The best time to map those exposures is before disappointment arrives.

Source : VOXeu