Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026)

"Financing the AI Buildout" is a paper by Stijn Van Nieuwerburgh of Columbia Business School, written for the fall 2026 edition of the Brookings Papers on Economic Activity (BPEA), the conference series and journal published by the Brookings Institution. The conference draft is dated September 4, 2026; Brookings posted it with a summary on September 23, 2026, and it was presented at the BPEA fall conference on September 25, 2026 (Source: brookings.edu). The paper estimates the physical and financial scale of the U.S. AI data-center buildout, compares it with earlier U.S. infrastructure booms, describes how the financing is shifting away from hyperscaler balance sheets, works through Meta's Hyperion data-center financing as an example, and sets out the risks and the policy and research questions the author sees in that financial structure. It is a conference draft and has not been through the journal's final revision.

Summary

The paper's argument has three parts. First, it treats the near-term economics of AI as a physical capital boom: specialized chips, electricity, cooling, network equipment and purpose-built facilities. It estimates the cost of a representative campus and aggregates project-level data into a national investment scenario. Second, it argues that hyperscaler capital expenditure has outgrown internal cash flow, so financing is moving into leases, joint ventures, project finance, private credit and structured-finance vehicles that separate the users of compute from the owners of the assets and from the investors who bear the risk. Third, it argues that these arrangements "transform rather than eliminate risk": they can raise leverage on the underlying assets while reported hyperscaler leverage stays low, and they write long-duration debt against assets whose demand, technological life and residual value are uncertain.

Cost of a representative campus

The paper's building block is a 200 MW AI training campus (Section II and Appendix A). At a power usage effectiveness of 1.2, 200 MW of total facility power supports about 166.7 MW of critical IT load, with the remaining about 33.3 MW used for cooling, electrical conversion, lighting and other facility operations. At about 142 kW per NVIDIA GB300 NVL72 rack (NVIDIA's reported figure), the campus holds about 1,170 high-density racks. The paper contrasts this with traditional CPU racks of 5–10 kW; current rack-scale AI systems "can draw well above 100 kW" and increasingly use direct-to-chip liquid cooling. A continuous 200 MW load would use about as much electricity a year as 170,000 average U.S. households.

The facility benchmark of $11 million per MW is close to JLL's 2026 global construction-cost forecast of $11.3 million per MW; the paper also cites Turner & Townsend's finding that liquid-cooled U.S. data centers cost 7–10% more than comparable air-cooled ones. The $4.0 million rack price includes GPUs, high-bandwidth memory, Grace CPUs, NVLink switching, power systems and the rack itself; installed IT cost comes to about $4.8 million per rack, or about $34 million per MW of critical IT capacity. The power allowance is described as campus-related incremental infrastructure, not a measure of all induced grid investment. The result is about $8.2 billion per 200 MW campus, about one-third facility and power and about two-thirds IT (32% and 68% in the appendix), which scales to about $41 billion per GW. The paper concludes that "the economic center of gravity has shifted toward the compute layer," while physical infrastructure still constrains feasibility.

Capacity scenario

Using Cleanview project-level data accessed on July 23, 2026, the paper estimates about 57 GW of operating U.S. data-center capacity and a planned pipeline of about 509 GW, which it states "is not a completion forecast." The database held 2,972 projects; the author excluded 39 canceled projects, leaving 2,933. Capacity was reported for 2,501 and missing for 432, which were imputed with a log-capacity regression on land area, building area and investment, with completion-year fixed effects and ridge-penalized county and developer effects (17 named-developer groups; penalty λ = 1 chosen by five-fold cross-validation; 2,501 observations; R² of 0.846 in logs; Duan smearing factor 1.1513; imputed values bounded between 1 and 2,500 MW, which affected 10 predictions). Capacity was imputed for 61 operational and 371 planned projects. Cleanview's capacity field does not consistently separate total facility power from critical IT load, so the paper calls the figures "reported total project power capacity."

Before any realization adjustment, the database holds 566.0 GW across the 2,933 projects: 39.0 GW in 971 projects operational before 2025 (1.6 GW imputed), 12.7 GW in 170 projects with a 2025 completion year, 33.8 GW in 285 projects dated 2026 (80 operational projects with 6.3 GW and 205 planned with 27.5 GW), and 353.6 GW in 662 planned projects with no completion date. The paper then assigns planned capacity to completion by 2032, completion after 2032, or never, with the share completed by 2032 falling with the reported horizon (from 90% for observed 2026 capacity to 30% for observed capacity dated after 2031) and lower where capacity is imputed. For undated projects it assumes 20% by 2032, 30% after and 50% never for reported capacity, and 10%, 25% and 65% for imputed capacity, so that of the 353.6 GW, 59.7 GW is completed by 2032, 100.6 GW after 2032 and 193.3 GW never. The paper identifies this undated category as "the largest source of uncertainty in the central scenario."

The central scenario divides the 526.9 GW gross post-2024 pipeline into 182.7 GW operational by the end of 2032 (given as 182.8 GW in Section II and in one sentence of Appendix B), 117.2 GW completed after 2032 and 226.9 GW never completed. The 182.7 GW comprises 19.0 GW associated with projects operational in 2025 or 2026, 141.7 GW of planned capacity with observed megawatts and 22.1 GW of planned capacity with imputed megawatts. The paper calls this "a scenario rather than a forecast," sensitive to large multiphase projects and projects with no completion date.

Investment and historical comparison

Applying $41 million per MW in 2025, rising 4% a year, and spreading each project's cost over its completion year and the three preceding years (10% in year t−3, 25% in t−2, 40% in t−1 and 25% in t), the scenario implies $10.3 trillion ($10,279 billion) of investment during 2025–2032, including pre-completion spending on projects that come online after 2032. With nominal GDP of $30.762 trillion in 2025 growing 4% a year, that is 3.63% of cumulative 2025–2032 GDP; the arithmetic average of the eight annual ratios is also 3.63%.

The paper compares this with average annual capital expenditure as a share of GDP in earlier U.S. booms (Table 1, Appendix C):

On this basis the author writes that the buildout "would be larger relative to the economy than the major U.S. canal, railroad, electrification, highway, and telecommunications investment booms," and that it "would double the electricity consumption of the entire U.S. residential sector." He qualifies the comparison: the historical series differ in scope and accounting, canals, railroads and fiber served for decades while GPUs "may become economically obsolete within three to six years," so the buildout generates less net capital formation per dollar of gross investment, and the estimate excludes hardware refreshes, so capacity commissioned in 2025–26 that needs new GPUs around 2030 would add to cumulative gross investment. The paper cites Soto, Thieu and Allen's July 17, 2026 FEDS Note for the finding that AI infrastructure investment contributed meaningfully to U.S. GDP growth in 2025 and the first quarter of 2026, and notes that McKinsey, Morgan Stanley and Moody's also project a multi-trillion-dollar buildout.

Hyperscaler capex and the move to external finance

Aggregate capital expenditures of Oracle, Microsoft, Amazon, Meta and Alphabet rose from roughly $97 billion in 2020 to more than $400 billion in 2025 and are projected to exceed $800 billion in 2026, when, on current estimates, capex will exceed the five firms' combined operating cash flow for the first time (Figure 1; historical values from 10-K and 10-Q filings, 2026 values combining reported results, company guidance and Wall Street estimates; the series includes non-AI capex and excludes capex on data centers the hyperscalers lease). The extracted text layer of Figure 1 carries aggregate values of 96.8 (2020), 130 (2021), 158.3 (2022), 153.5 (2023), 240.4 (2024), 415.8 (2025) and 800.5 (2026) billion dollars under a garbled column label; the per-company values in the extraction are chart readings and are not reproduced here.

The paper describes the investment-grade credit of Amazon, Microsoft, Meta, Alphabet and Oracle as what makes outside capital available: long-term agreements with these tenants support real estate, infrastructure and project financing. Frontier AI firms such as OpenAI and Anthropic "generally lack comparable standalone credit profiles" and obtain compute largely through hyperscaler contracts, "effectively relying on the hyperscalers' balance sheets and creditworthiness." Developers, REITs, infrastructure funds and private equity increasingly build or buy facilities and lease them to hyperscalers; bank mortgage loans and syndicated credit lines remain important, while private credit and structured finance supply a growing share of project debt. Financing is extending to GPUs and other IT equipment through leases and asset-backed structures, collateral the paper describes as shorter-lived and more exposed to obsolescence than real estate or aircraft.

Citing Morgan Stanley Research, the paper reports that more than half of the roughly $2.9 trillion required for hyperscalers' incremental compute needs over 2025–2028 is expected to come from outside capital, with an approximate 60–40 equity-debt split; within debt, private credit is about $800 billion, corporate debt roughly $200 billion and structured finance approximately $150 billion. The author argues that the aggregate split understates asset-level leverage, because hyperscalers direct internal capital toward IT equipment while shells, power and real estate are financed through project debt and leases, which can raise loan-to-value and debt-to-equity on the infrastructure even when the hyperscalers' reported leverage remains low.

Hyperion (Beignet) as the worked example

Section IV describes Meta's Hyperion data-center financing in Louisiana: about 2.0 GW of capacity and a total investment of about $30 billion (about $15 million per MW excluding IT equipment, which Meta finances separately as tenant). Meta initially owned the project outright, then sold an 80% equity stake to Blue Owl for approximately $2.5 billion. The joint venture, Beignet, raised $27 billion of debt in October 2025, rated A+ by Standard & Poor's, one notch below Meta's corporate rating; the paper describes it as "the largest individual investment-grade corporate debt issuance in U.S. history."

• Leverage. About $27 billion of debt against $30 billion of asset value, a debt-to-asset ratio of about 90%, with a debt service coverage ratio of 1.12×. By comparison Meta's own year-end 2025 leverage ratio was about 25% on book equity and 4% on market equity, and "Meta's operating cash flow of $130 billion is at least 18× its 2026 debt service."
• Structure. Bankruptcy-remote special-purpose vehicles issue debt backed by future lease payments. Per the paper's Figure 2 (drawn from S&P's pre-sale report and Entergy's 2025 Form 10-K), Laidley LLC owns the campus, funded by its subsidiary Project Beignet Holdings LLC, which is owned 20% by Meta through Iris Crossing LLC and 80% by Blue Owl through Beignet Investor LLC; Laidley has a power purchase agreement with Entergy Louisiana, which will develop three gas-fired generating units.
• Lease and guarantee. Meta holds a sequence of five four-year leases beginning in 2029 and running to the bond's 2049 maturity. At each renewal it may terminate for some or all of the campus; the assets are then offered for sale and Meta covers any shortfall below a contractually specified minimum value, a residual value guarantee.
• Accounting. Under current GAAP, Meta does not recognize the future lease obligations or the residual value guarantee before they become effective, "even though one of the two commitments will ultimately materialize for sure," so reported leverage understates the economic obligation.
• Cost. Beignet's debt yielded 6.58%, at least 100 basis points above what Meta likely would have paid on comparable unsecured debt, implying more than $5 billion of additional interest over the life of the financing, which is passed through in Meta's lease payments. The author states that the rationale is "not cheaper financing, but preservation of an asset-light corporate profile."

The paper suggests Hyperion "may therefore provide a template," noting that the 960 MW Sopaipilla project in El Paso, Texas, a joint venture of Meta and BlackRock, follows a nearly identical structure and that its investor vehicle issued $12.3 billion of A+-rated bonds in July 2026. It adds that publicly rated debt is a relatively small share of data-center financing; much larger volumes come through syndicated bank loans and private-credit facilities "for which little information is publicly available."

Off-balance-sheet obligations

The paper cites Moody's (February 2026) estimate that hyperscalers have roughly $970 billion of lease commitments, of which about $660 billion relates to future leases not yet on balance sheet. It also cites a Wall Street Journal tally of $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 chips, against $604 billion of on-balance-sheet obligations ($356 billion of long-term debt and $248 billion of current lease obligations), so that only one-fifth of about $3 trillion of data-center-related obligations is currently recognized.

Risks

Section V argues that "the sector's financing structure can transmit and amplify" technological and operating risk, and names three vulnerabilities:

1. Tenant and demand concentration. Many large campuses are effectively single-tenant; hyperscalers' commitments depend on AI profitability and on a small set of model developers and enterprise customers, and if demand disappoints hyperscalers may shift workloads away from leased capacity to protect utilization of owned facilities, concentrating the adjustment on third-party assets.
2. Technological obsolescence. Campuses built around current chip generations may lose competitiveness before their structures wear out; rising rack densities, new interconnect standards and cooling requirements can force retrofits, and more efficient models, edge computing or decentralized inference could reduce demand for large centralized campuses, creating a mismatch between long financing maturities and shorter asset lives.
3. Execution. Power, transmission, permits and hardware (GPUs, high-bandwidth memory, networking) are capacity-constrained and often outside developers' control; concentration of advanced semiconductor production means geopolitical disruption could amplify bottlenecks.

The author argues that high asset-level leverage and layered claims turn modest deteriorations into larger equity losses and possible creditor impairment, and that SPVs, guarantees, securitization and originate-to-distribute lending can weaken monitoring. He describes cross-dependencies among chip designers, cloud providers, model developers, operators and investors as a "circularity" that makes exposures more correlated than individual transactions suggest. His example is the $500 billion credit facility that Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR set up in August 2026 to finance NVIDIA chips, for which, he writes, NVIDIA "will provide a 25% residual value guarantee, exposing it to as much as $125 billion of backstops"; he adds that Google and Broadcom entered a similar arrangement for TPU chips. See Circular Financing in AI.

As asset-pricing evidence, Figure 3 reports 36-month rolling betas of U.S. data-center REIT excess returns (FTSE Nareit series, monthly January 2018 to March 2026, controlling for Treasury returns and the Fama-French five factors; AI stocks defined as constituents of the Global X AIQ ETF on March 18, 2026):

The author reads the rise from roughly 0.5 to around 1 as data centers appearing "less like defensive infrastructure assets or a hedge against AI-related risk and more like part of the same systematic risk complex." He states that these developments "do not imply that financial distress is imminent," and that the relevant question is "under what conditions project-level losses could become correlated and propagate across firms and financial institutions."

Revenue required

Appendix D computes the mature annual revenue the 2025–2032 cohorts would need, stated as a required-return calculation "not a forecast of AI revenues." The 182.7 GW has an initial cost of about $8.54 trillion; adding a return on capital committed during construction (a factor of 1.126 at a 10% cost of capital) raises the capital base to about $9.61 trillion. Assuming IT equipment is 68% of cost with a six-year life and the remaining 32% has a 20-year life, the capital-recovery factors at a 10% unlevered return are 22.96% and 11.75%, a weighted annual charge of 19.37%. The capacity must then generate $1.86 trillion a year of operating cash flow, which at a 50% operating cash-flow margin is about $3.725 trillion of annual revenue by 2032, 9.2% of projected 2032 GDP of $40.481 trillion.

In market units the central requirement is about $20.4 million per MW-year of total facility power ($1.70 million per MW-month), about $24.5 million per MW-year of critical IT load, or $2.79 per kWh of GPU power. The 182.7 GW supports roughly 1.07 million GB300 NVL72 racks, or 77 million GPUs, implying about $5.5 per installed GPU-hour at full utilization, $6.9 per billed GPU-hour at 80% utilization and $7.9 at 70%. The paper compares this with on-demand or short-term prices for high-end NVIDIA capacity of roughly $6 to more than $10 per GPU-hour, concluding that the requirement "does not appear implausible relative to current pricing" and that the question is whether demand can absorb roughly 180 GW at high utilization and whether GPU-hour prices hold as capacity and competition grow. Against current estimates of about $100 billion of combined annual revenue at OpenAI and Anthropic, revenue would need to grow roughly 80% a year until 2032 to reach $3.725 trillion.

If IT equipment has a three-year rather than six-year life, the weighted capital charge rises from 19.37% to 31.10%, required revenue from $3.7 trillion to about $6.0 trillion (9.2% to 14.8% of 2032 GDP), and the required price from about $5.5 to $8.8 per installed GPU-hour.

Policy and research questions

Section VI names four areas:

• Measurement. Reported hyperscaler capex captures only part of the buildout; the author calls for measures of who bears demand, refinancing and residual-value risk, not only who legally owns each asset, and cites ongoing work by Van Nieuwerburgh and Z. Sun ("Clouds of credit: Mapping the financial web behind AI compute," Columbia Business School working paper dated November 2026) and the private data provider Atrium.ai.
• Disclosure and accounting. More informative disclosure would map project-level debt, joint-venture equity, future lease commitments, residual-value and construction guarantees, tenant concentration, renewal options and assumed recovery values; the objective "need not be to eliminate off-balance-sheet financing, which can serve legitimate economic purposes," but to make commitments and risk transfers transparent.
• Valuation and capital-structure design. Depreciation, collateral values, lease duration, refinancing risk and debt maturity for assets mixing long-lived buildings with short-lived hardware, and the incentive to separate infrastructure-intensive assets from high-growth operating companies to preserve reported capital intensity and valuation multiples.
• Financial stability. The author writes that "it would be premature to conclude that AI infrastructure already poses systemic risk comparable to earlier credit booms," and that the questions are how much exposure banks, insurers, private credit funds and securitization vehicles hold, how concentrated it is in the same tenants, and how losses would propagate under a joint fall in utilization, collateral values and refinancing capacity. Off-balance-sheet structures matter "because they may make correlated exposures harder to observe before a downturn."

He concludes that "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 states that the author has no conflicts of interest to disclose.

Key claims

Discrepancies noted

• Section II and one sentence of Appendix B give 182.8 GW completed by 2032; Table 4, Table 6 and the following sentence of Appendix B give 182.7 GW. The abstract and introduction round to 183 GW, and the introduction gives the post-2032 figure as 118 GW against 117.2 GW in Table 4.
• Appendix B spreads the 59.7 GW of undated capacity completed by 2032 evenly over 2028–2032 for the investment estimate, while Appendix D allocates it across 2029–2032 for the revenue estimate.
• Section II reports about 57 GW operating and a planned pipeline of about 509 GW; Appendix B reports 566.0 GW in the database, 39.0 GW operating before 2025, and a 526.9 GW gross post-2024 pipeline. These are different aggregates of the same data (the 57 GW includes projects operational in 2025 and 2026), not stated as a correction.
• The Brookings summary page quotes the paper as saying off-balance-sheet structures "may make correlated exposures hard to observe before a downturn"; the draft reads "harder to observe." The summary describes the author as "of Columbia University" (his title there is Earl Kazis and Benjamin Schore Professor of Real Estate); the draft gives Columbia Business School.
• NVIDIA financing facility: the paper states NVIDIA "will provide a 25% residual value guarantee" on the $500 billion facility. Contemporary reporting described the August 10, 2026 arrangements as memorandums of understanding with an uncommitted $500 billion target and an option for Nvidia to backstop up to 25% of any project's financing (see Nvidia & TSMC — AI Compute Infrastructure).
• Meta El Paso: the paper gives 960 MW and $12.3 billion of A+-rated bonds issued in July 2026; Meta's July 28, 2026 announcement gave 1 gigawatt and a $12.5 billion debt financing, and the Wall Street Journal reported a $12 billion-plus debt sale led by BlackRock (see Meta AI).

Provenance

The raw file is a text-layer extraction of the 29-page conference draft PDF from brookings.edu, fetched on September 24, 2026 during the developments-log fold of the September 24 afternoon digest, which had carried only the Brookings summary (and attributed it to Reuters in its source list). Some words are run together by extraction, the figures are rendered as extracted text, and the Figure 1 data table's column labels are garbled. The Brookings landing page, whose summary language is credited to David Skidmore, identifies the BPEA editors as Janice Eberly and Jón Steinsson and lists the paper with two others presented on September 25 on the risks and rewards of the AI boom, "Why is AI so contentious" and "The vanishing advantage of specialization" (Source: brookings.edu). Canonical PDF: brookings.edu.

Relationships

• supports: AI Bubble Debate — quantifies the buildout and the revenue it requires
• supports: Private Credit & AI Infrastructure — documents the shift to leases, JVs, project debt, private credit and SPVs
• supports: AI Infrastructure Capex — hyperscaler capex exceeding operating cash flow in 2026
• supports: Circular Financing in AI — vendor-backed financing as a source of correlated exposure
• related: AI Data Centers — campus cost, capacity pipeline, obsolescence
• related: Meta AI — Hyperion (Beignet) and El Paso (Sopaipilla) financings
• related: Nvidia & TSMC — AI Compute Infrastructure — $500 billion financing platform and residual value guarantee
• related: AI Bubble vs. Buildout — Synthesis
• related: Stijn Van Nieuwerburgh, Brookings Institution AI Initiative