The AI buildout is on track to absorb a larger share of US output than the railroads, electrification, the interstate highways or the telecoms boom, and the way it is being financed could create systemic risks. That is the conclusion of a paper by Stijn Van Nieuwerburgh, a finance and real estate professor at Columbia Business School, prepared for the Brookings Institution’s autumn economics conference. It estimates that investment in AI data centers, power systems, networking and chips will total $10.3 trillion between 2025 and 2032, an average of 3.63% of US gross domestic product a year.
The size is only half the story. The paper’s main warning is about where the risk now sits. The AI buildout was first paid for out of the cash piles of Amazon, Microsoft, Alphabet and Meta. It is now increasingly financed through leases, joint ventures, private credit, securitisations and special-purpose vehicles that keep debt off the big companies’ balance sheets. “This opacity of all these special purpose vehicles is somewhat reminiscent of what happened in the subprime mortgage crisis,” Van Nieuwerburgh told reporters, according to Reuters.
This article explains what the paper estimates and how the AI buildout compares with earlier infrastructure booms. It walks through what a campus costs and the Meta financing the paper uses as its case study, then sizes the off-balance-sheet exposure and the revenue the industry must earn to justify it. It also looks at the early warning signs already visible this month and what businesses buying AI services should take from the analysis. For the deal that put AI financing on the front pages this week, see our analysis of the Oracle and Blue Owl project delay.
Table of contents
- What the Brookings Paper Says About the AI Buildout
- How the AI Buildout Compares With America’s Great Infrastructure Booms
- What an AI Buildout Campus Actually Costs
- Where the Money Comes From: The AI Buildout Capital Stack
- Inside Hyperion: A Template for AI Buildout Financing
- How Big Is the Off-Balance-Sheet AI Buildout?
- Why the Structure Could Amplify an AI Buildout Shock
- The $3.7 Trillion Revenue Question Behind the AI Buildout
- Is the AI Buildout Like 2008?
- Early Warning Signs Already Visible in the AI Buildout
- What the Paper Asks Policymakers to Do About the AI Buildout
- What the AI Buildout Means for Businesses Buying AI Services
- What to Watch Next
- Frequently Asked Questions About AI Buildout Financing
- References
What the Brookings Paper Says About the AI Buildout
The paper, Financing the AI Buildout, is a conference draft dated 4 September and presented on Friday 25 September at the fall meeting of the Brookings Papers on Economic Activity. Reuters’ Howard Schneider reported its findings on 24 September after a briefing with the author.
$10.3 trillion, or 3.63% of GDP
Van Nieuwerburgh estimates that a representative AI campus costs about $41 million per megawatt of power capacity. Applying that cost, growing 4% a year, to the projects he expects to be built gives $10.279 trillion of investment between 2025 and 2032. Against nominal GDP growing at the same 4% rate, that is 3.63% of cumulative output. Spending peaks at 5.1% of GDP in 2032, on his central scenario.
183 gigawatts by 2032
The estimate starts from project-level data compiled by Cleanview in July 2026. It counts about 57 gigawatts of US data center capacity operating today and a planned pipeline of about 509 gigawatts. The paper does not assume all of it gets built. Its central scenario has 182.8 gigawatts coming online by the end of 2032, 117.2 gigawatts completed later, and 226.9 gigawatts never completed.
Premature to call it systemic, but not safe
The paper is careful. “It would be premature to conclude that AI infrastructure already poses systemic risk comparable to earlier credit booms,” it says. But off-balance-sheet structures matter “because they may make correlated exposures hard to observe before a downturn”. The relevant question, in the paper’s words, “is therefore not whether AI infrastructure is already systemically risky, but under what conditions project-level losses could become correlated and propagate across firms and financial institutions”.
| Measure | Paper’s central estimate |
|---|---|
| Investment, 2025 to 2032 | $10.3 trillion |
| Average share of GDP | 3.63% a year |
| Peak year | 2032, 5.1% of GDP |
| Capacity operating today | About 57GW |
| Planned pipeline | About 509GW |
| New capacity online by 2032 | 182.8GW |
| Cost per megawatt, 2025 | About $41 million |
| Revenue needed by 2032 | About $3.7 trillion a year |
How the AI Buildout Compares With America's Great Infrastructure Booms
The historical comparison is the paper’s headline, and it holds up on the author’s own numbers. Table 1 of the paper measures average annual capital spending as a share of GDP over each boom’s main building period.
Railroads were the previous record
Railroad construction averaged 2.24% of GDP between 1870 and 1890, and the annual share exceeded 4% in 1870 and 1881. The AI buildout’s projected average is about 1.6 times the railroad figure. The telecom and fiber boom of 1996 to 2003, which ended in a wave of bankruptcies and a sharp collapse in communications investment, averaged 1.10%. Each of those earlier booms delivered lasting infrastructure. Each also left investors with heavy losses when demand arrived later, or smaller, than the financing assumed.
Shorter-lived assets
The comparison flatters the AI buildout in one way. Canals, railroads and fibre cables provided services for decades. GPUs, the paper notes, “may become economically obsolete within three to six years”. So each dollar of AI investment creates less lasting capital than a dollar spent on a railway. Buildings, power connections and cooling systems last longer, but they are the smaller share of the cost.
Refreshes are not counted
The $10.3 trillion also excludes hardware replacement. Capacity built in 2025 and 2026 may need most of its GPUs replaced by around 2030, while the buildings stay in use. Those refresh cycles would push cumulative spending higher than the headline figure.
What an AI Buildout Campus Actually Costs
The paper builds its estimate from the bottom up, using a representative 200-megawatt AI training campus. The breakdown shows where the money goes.
| Component | Assumption | Cost | Share |
|---|---|---|---|
| Data center building | 200MW at $11 million per MW | $2.2 billion | 26.8% |
| Extra power infrastructure | Campus allowance | $0.4 billion | 4.9% |
| Nvidia GB300 NVL72 racks | 1,170 racks at $4.0 million | $4.68 billion | 57.0% |
| Network fabric and fibre | 10% of rack cost | $0.468 billion | 5.7% |
| Shared storage | 5% of rack cost | $0.234 billion | 2.8% |
| Integration and commissioning | 5% of rack cost | $0.234 billion | 2.8% |
| Total | 200MW campus | $8.2 billion | 100% |
Two-thirds is equipment
About 68% of the cost is IT equipment and about 32% is the building and power. The paper says this means “the economic center of gravity has shifted toward the compute layer”, even though the physical constraints, above all power, still decide whether a campus can be built. That split matters for financing. Lenders are comfortable with buildings that last decades. They are less comfortable lending against chips that may be obsolete before the loan is repaid.
What 200 megawatts means
A 200-megawatt campus supports about 1,170 high-density racks and uses about as much electricity a year as 170,000 average US households, the paper estimates. The full AI buildout, it says, would double the electricity consumption of the entire US residential sector.
Where the Money Comes From: The AI Buildout Capital Stack
The paper’s central argument is about how the AI buildout is financed. The biggest technology companies used to pay for data centers from their own cash. That model is running out of road.
Capex is outrunning cash flow
Capital spending by Oracle, Microsoft, Amazon, Meta and Alphabet rose from about $97 billion in 2020 to more than $400 billion in 2025, and is projected to exceed $800 billion in 2026, the paper says. On its figures, 2026 is the first year in which the five companies’ combined capital spending exceeds their combined operating cash flow.
The paper’s figure puts 2026 operating cash flow for the five at about $707 billion, against $800.5 billion of capital spending. In 2020, the same companies spent about 38% of their operating cash flow on capital projects. In 2026, the projection is about 113%.
Leases, joint ventures and private credit
To close the gap, hyperscalers are combining direct ownership with leases, joint ventures, project finance, private credit and structured vehicles. Developers, infrastructure funds, real estate investment trusts and private equity firms supply equity. Banks, private credit funds and securitisation vehicles supply debt. Financing is also spreading from buildings to GPUs, which can be leased or financed through asset-backed structures.
Morgan Stanley’s estimate
The paper cites Morgan Stanley Research’s estimate that more than half of the roughly $2.9 trillion needed for hyperscalers’ extra compute between 2025 and 2028 will come from outside capital, split about 60% equity and 40% debt. Within the debt, private credit accounts for about $800 billion, corporate debt about $200 billion and structured finance about $150 billion. The paper warns that the 60-40 split “can understate leverage at the asset level”, because the hyperscalers put their own cash into chips while buildings and power sit in separately financed vehicles with much more debt.
Inside Hyperion: A Template for AI Buildout Financing
To show how the structures work, the paper dissects Hyperion, Meta’s data center project in Louisiana. It calls the deal a possible “template for future data center financings”.
90% debt at the project level
Hyperion involves about 2 gigawatts of capacity and about $30 billion of investment, excluding the chips, which Meta finances separately. Meta sold an 80% equity stake to Blue Owl for about $2.5 billion. The resulting joint venture, called Beignet, raised $27 billion of debt in October 2025. S&P rated it A+, one notch below Meta. The paper says it was the largest individual investment-grade corporate debt issue in US history. About $27 billion of debt against a $30 billion asset means a debt-to-asset ratio of about 90%, with debt service covered only 1.12 times. Meta’s own leverage, by contrast, was about 25% using book equity at the end of 2025.
Leases that can end every four years
Meta leases the campus through five four-year leases starting in 2029 and running to the bonds’ 2049 maturity. At each renewal, Meta may walk away from some or all of the campus. If it does, the assets are sold and Meta must cover any shortfall below a contractually specified minimum value. That residual value guarantee gives Meta flexibility while giving lenders the long-duration support they need.
The cost of staying off the balance sheet
Beignet’s debt was issued at a yield of 6.58%, at least 100 basis points above what Meta would likely have paid on its own unsecured debt, the paper says. Over the life of the financing, that spread implies more than $5 billion of extra interest, ultimately paid through Meta’s rent. The benefit for Meta is not cheaper money but “preservation of an asset-light corporate profile”. Under current accounting rules, Meta does not record the future lease payments or the residual value guarantee on its balance sheet before they take effect, “even though one of the two commitments will ultimately materialize for sure”.
Already being copied
The paper notes that the 960-megawatt Sopaipilla project in El Paso, Texas, a joint venture of Meta and BlackRock, follows a nearly identical structure. Its investor vehicle issued $12.3 billion of A+-rated bonds in July 2026.
| Hyperion feature | Figure or term |
|---|---|
| Capacity | About 2GW |
| Investment, excluding chips | About $30 billion |
| Equity sold to Blue Owl | 80%, for about $2.5 billion |
| Debt raised by the joint venture | $27 billion, October 2025 |
| Rating | A+, one notch below Meta |
| Debt to asset value | About 90% |
| Debt service coverage | 1.12 times |
| Lease structure | Five four-year leases from 2029 to 2049 |
| Yield | 6.58%, at least 100 basis points above Meta’s likely cost |
| Extra interest over the life of the debt | More than $5 billion |
How Big Is the Off-Balance-Sheet AI Buildout?
The paper assembles several estimates of the AI buildout obligations that do not appear on the hyperscalers’ balance sheets. They differ in scope, but they all point to large numbers.
Moody’s: $660 billion of future leases
Moody’s estimates that hyperscalers have about $970 billion of lease commitments, of which about $660 billion relates to future leases not yet on their balance sheets, the paper says.
The Wall Street Journal’s $2.4 trillion tally
The paper also cites a Wall Street Journal analysis counting $2.4 trillion of off-balance-sheet liabilities at Microsoft, Alphabet, Amazon and Meta: $904 billion of future lease commitments and $1.52 trillion of future purchase commitments, mostly for chips. That compares with $604 billion of on-balance-sheet obligations, made up of $356 billion of long-term debt and $248 billion of current lease obligations. On that tally, only about one-fifth of roughly $3 trillion of data-center-related obligations is currently recognised.
Oracle’s unstarted leases
Oracle, which the paper counts among the hyperscalers, is a live example. Its latest quarterly report, as we covered in our Oracle and Blue Owl analysis, disclosed $288 billion of lease commitments that have not yet started, “substantially all related to data center arrangements”, with terms of 15 to 19 years.
Why the Structure Could Amplify an AI Buildout Shock
The paper’s core point is that financing can transmit and amplify risks that already exist in the business. It names four vulnerabilities.
Tenant concentration
Many large AI campuses are effectively single-tenant facilities. Their cash flows depend on one hyperscaler’s willingness to keep leasing, and ultimately on demand from a small number of model developers. If demand disappoints, the paper warns, hyperscalers may protect their own facilities by moving workloads away from leased capacity, “concentrating the adjustment on third-party assets”.
Technological obsolescence
Campuses built around today’s chips may become uncompetitive before their buildings wear out. Rising rack densities, new interconnect standards and new cooling requirements can force costly retrofits. More efficient models, or a shift to smaller and more distributed inference, could reduce demand for giant campuses altogether. That creates a mismatch between long-dated debt and short-lived assets.
Execution: power and hardware
Large campuses need electricity, transmission, permits and specialised hardware on schedule. Grid interconnection delays, local opposition and shortages of GPUs, high-bandwidth memory and networking can postpone completion. “For investors, the resulting risk is not only that project costs escalate, but that revenue begins later or never reaches the levels assumed at underwriting,” the paper says. Oracle’s force majeure notice at its New Mexico campus this week, citing delays in securing power, is exactly that risk arriving.
Circularity
The paper highlights the $500 billion financing platform that Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR set up in August 2026 to finance Nvidia chips, with Nvidia providing a 25% residual value guarantee. That exposes Nvidia to as much as $125 billion of backstops, and “illustrates the circularity created when a dominant equipment supplier helps finance customers’ purchases of infrastructure built around its own products”. We examined the same pattern in our analysis of circular financing.
Markets are already treating data centers as risky
One piece of evidence stands out. The market beta of publicly traded data center REITs, a measure of how closely their returns track the stock market, has risen from about 0.5 to about 1, the paper finds. Data centers “appear less like defensive infrastructure assets or a hedge against AI-related risk and more like part of the same systematic risk complex”.
The $3.7 Trillion Revenue Question Behind the AI Buildout
The most striking number in the paper is what the industry must earn to justify the spending. It is not a forecast of AI revenue. It is the revenue needed to earn a 10% unlevered return, assuming a 50% operating cash-flow margin, six-year lives for IT equipment and 20-year lives for everything else.
$3.7 trillion a year by 2032
On those assumptions, the 182.7 gigawatts completed between 2025 and 2032 must generate about $3.725 trillion of annual revenue once mature, about 9.2% of projected 2032 GDP. “Given current estimates of annual combined revenues of OpenAI and Anthropic of around $100 billion, revenues would need to grow at roughly 80% per year until 2032,” the paper says.
What that means per GPU-hour
In market terms, the requirement works out at about $5.50 per installed GPU-hour at full utilisation, $6.90 per billed hour at 80% utilisation and $7.90 at 70%. The paper notes that current on-demand prices for high-end Nvidia capacity range from about $6 to more than $10 per GPU-hour. So the requirement is not implausible today. The question is whether prices stay there once 180 gigawatts of new capacity is competing for customers.
| Required unlevered return | 33% cash-flow margin | 50% cash-flow margin | 67% cash-flow margin |
|---|---|---|---|
| 5% | $4.65 trillion | $3.07 trillion | $2.29 trillion |
| 10% | $5.64 trillion | $3.73 trillion | $2.78 trillion |
| 15% | $6.72 trillion | $4.44 trillion | $3.31 trillion |
If the chips wear out faster
The six-year life for IT equipment is the assumption that matters most. If chips last three years instead, the annual revenue requirement rises from about $3.7 trillion to about $6.0 trillion, or from 9.2% to 14.8% of 2032 GDP. The required price rises from about $5.50 to $8.80 per installed GPU-hour.
Is the AI Buildout Like 2008?
Van Nieuwerburgh’s subprime comparison drew the headlines. It is worth being precise about where it holds.
Where the comparison holds
Before 2008, mortgage risk was sliced into securities, moved into off-balance-sheet vehicles and spread across investors, so that nobody could see where losses would land until they did. The AI buildout shares the opacity. Much of its debt sits in bankruptcy-remote special-purpose vehicles, private credit funds and syndicated loans “for which little information is publicly available”, the paper says. It also cites research on how originate-to-distribute lending can weaken monitoring, a mechanism central to the mortgage crisis.
Where it does not
There are real differences. Today’s main tenants are investment-grade companies with large cash flows, not subprime borrowers. Much of the financing is equity rather than debt. Publicly rated deals like Hyperion come with rating agency analysis. And the paper states plainly that the developments “do not imply that financial distress is imminent”. Strong growth in AI applications, high utilisation and better models could support the projected infrastructure and generate stable cash flows.
The common thread
The honest summary is that the AI buildout is not a mortgage bubble, but it is building the same kind of blind spot. When exposures are layered, correlated and hard to see, a shock that should be absorbed can spread. The paper’s worry is not a particular loss. It is that nobody can yet say who would take it.
| Feature | Mortgage crisis before 2008 | AI buildout today |
|---|---|---|
| Underlying borrowers | Subprime households | Investment-grade hyperscalers and unrated AI labs |
| Vehicles | Off-balance-sheet conduits and securitisations | Joint ventures, SPVs, private credit, leases |
| Transparency | Poor, exposures hard to locate | Poor for private and bank loans; better for rated deals |
| Asset life | Houses, decades | Chips, three to six years; buildings, decades |
| Leverage | High at the asset level | About 90% at Hyperion’s project level |
| Current stress | Not applicable | Loan discounts and delays at some projects |
Early Warning Signs Already Visible in the AI Buildout
The paper was written before this week’s news, but several of the AI buildout risks it describes are already visible in markets.
Project debt below par
On 18 September, the Financial Times reported that the roughly $18 billion of loans for Oracle’s Project Jupiter campus in New Mexico were being quoted at 89 to 91 cents on the dollar. Efforts to sell the debt to a wider pool of investors had stalled, leaving banks holding more than planned.
A force majeure notice over power
On 24 September, Oracle sent a force majeure notice to the Blue Owl unit developing Jupiter, citing potential delays in securing power, Reuters reported. The notice lets Oracle seek to pay lower development-stage rent for longer. That is the paper’s execution risk turning into a payment schedule.
The cost of money
The 10-year US Treasury yield closed at 5.11% on 23 September, according to Federal Reserve data, its highest close since July 2007. Every point on long-term rates raises the cost of financing long-lived AI infrastructure, and makes the 6.58% Hyperion yield look less generous.
Politics and inflation
Reuters noted that data center construction has become a central issue in US politics, with some localities increasingly reluctant to host the facilities and Federal Reserve officials considering whether the construction boom is adding to inflation. Some AI executives, it added, have suggested that a slower pace of development might be safer.
What the Paper Asks Policymakers to Do About the AI Buildout
The paper does not call for limits on the AI buildout or on off-balance-sheet financing, which it says “can serve legitimate economic purposes”. Its recommendations are about seeing the risk clearly.
Measure who owns what
First, measurement. Reported capital spending captures only part of the buildout, because assets are spread across corporate balance sheets, joint ventures, guarantees and SPVs. The paper wants measures that identify not just who legally owns each asset, but who bears the demand, refinancing and residual-value risk.
Better disclosure
Second, disclosure. More informative reporting would map project-level debt, joint venture stakes, future lease commitments, residual-value guarantees, construction guarantees, tenant concentration, renewal options and the assumed recovery values of specialised assets.
Valuation and financial stability
Third, valuation: whether current accounting captures the economic cost of capital when long-lived buildings are combined with short-lived chips. Fourth, financial stability: how much exposure sits with banks, insurers, private credit funds and securitisation vehicles, how concentrated it is in the same tenants, and how losses would spread if utilisation, collateral values and refinancing all fell together. “The most important policy contribution at this stage may therefore be to improve measurement and transparency while the capital structure of the industry is still evolving,” the paper concludes.
What the AI Buildout Means for Businesses Buying AI Services
Most businesses are customers of the AI buildout, not investors in it. The financing structures the paper describes still reach them, through prices, supplier health and contract terms.
Supplier credit risk is now an AI risk
The companies that provide AI capacity range from cash-rich hyperscalers to highly leveraged developers and GPU clouds. A vendor management review should ask, alongside the usual cybersecurity questions, how a supplier funds its capacity, and what happens to your service if its financing tightens.
Prices could move either way
If the industry reaches its revenue requirement, AI compute prices stay high. If capacity outruns demand, prices could fall sharply, which would be good for buyers but painful for lenders. Businesses signing long AI contracts should avoid locking in today’s prices for too long, and should build price reviews into multi-year deals.
Keep flexibility
The paper’s warning about short-lived hardware applies to buyers too. Committing to one provider’s infrastructure for many years ties you to its technology choices. An AI strategy that keeps workloads portable across providers is cheaper insurance than a stranded contract.
| Question for an AI supplier | Why the paper makes it relevant |
|---|---|
| How is your capacity financed? | Project debt and leases can transmit stress to customers |
| How concentrated are your customers? | Single-tenant campuses depend on a few buyers |
| Are prices reviewable during the term? | Prices could fall or rise sharply as capacity grows |
| What happens if a site is delayed? | Execution risk is rising as power and hardware tighten |
| Can we move workloads elsewhere? | Hardware ages in three to six years |
What to Watch Next
The paper will be discussed at the Brookings conference on 25 September, and a revised version will appear in the fall 2026 Brookings Papers on Economic Activity. Several other developments will test its warnings over the coming months.
The debt markets
Watch the prices of existing data center loans, starting with Jupiter’s, and the terms of the next large Hyperion-style bond. Wider spreads or failed syndications would show investors pricing in the risks the paper describes.
Hyperscaler cash flow
Quarterly results in October and November will show whether capital spending is still outrunning operating cash flow, and how much new financing is off the balance sheet.
The follow-up research
Van Nieuwerburgh cites forthcoming work with Z. Sun, due in November, that maps “the financial web behind AI compute”. If it identifies where exposures sit, it may answer the question the Brookings paper leaves open: who would take the losses if the AI buildout slows.
Frequently Asked Questions About AI Buildout Financing
How much will the AI buildout cost?
The Brookings paper estimates $10.3 trillion of US investment in AI data centers, power, networking and chips between 2025 and 2032, or 3.63% of GDP a year.
Is it bigger than past infrastructure booms?
Yes, relative to the economy. Railroads averaged 2.24% of GDP from 1870 to 1890, highways 1.13% and telecom and fibre 1.10%.
Why does the paper warn about systemic risk?
Because financing is shifting from company balance sheets to leases, joint ventures, private credit and special-purpose vehicles, which raise leverage at the project level and make exposures harder to see.
Does the paper say a crisis is coming?
No. It says it would be premature to call AI infrastructure a systemic risk comparable to earlier credit booms, and that distress is not imminent.
How much revenue does the industry need?
About $3.7 trillion a year by 2032 to earn a 10% return, on the paper’s central assumptions. That is roughly 80% annual growth from OpenAI’s and Anthropic’s combined revenue of about $100 billion.
What should businesses do?
Check how AI suppliers finance their capacity, avoid long price locks, keep workloads portable, and plan for delays at new sites.
References
Financing the AI buildout (Brookings Institution)
Financing the AI Buildout, conference draft (Brookings Papers on Economic Activity)
Financing of historic AI buildout raises systemic risks in US, researcher says (Reuters, via MSN)
The AI buildout could cost $10 trillion (International Business Times)
Oracle’s $18 billion data center debt under pressure, FT reports (Reuters)
Oracle triggers force majeure on data center project over power delays (Reuters, via MSN)
10-year Treasury constant maturity rate (FRED)
Oracle Form 10-Q for the quarter ended 31 August 2026 (SEC EDGAR)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
