---
id: "lib-2026-10-05-financing-the-ai-buildout-w35865"
type: "library"
title: "Financing the AI Buildout (w35865)"
lang: "EN"
org: "NBER (Columbia)"
source_type: "Paper"
date: "2026-10-05"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-05-financing-the-ai-buildout-w35865/"
original_url: "https://www.nber.org/papers/w35865"
tier: "S"
tier_name: "Shapes the field"
tier_reason: "Columbia finance professor Stijn Van Nieuwerburgh's paper, forthcoming in the Brookings Papers on Economic Activity, sizes the US AI infrastructure buildout against historic capex booms and maps where its financial risk sits."
summary_basis: "full-text"
original_minutes: "60"
tags: ["business transformation", "ai capex", "finance"]
---

# Financing the AI Buildout (w35865)

Tier S (Shapes the field): Columbia finance professor Stijn Van Nieuwerburgh's paper, forthcoming in the Brookings Papers on Economic Activity, sizes the US AI infrastructure buildout against historic capex booms and maps where its financial risk sits.

Summary based on the full text.

## In short

NBER Working Paper 35865 (October 2026) by Stijn Van Nieuwerburgh of Columbia Business School, forthcoming in the Fall 2026 Brookings Papers on Economic Activity. It treats AI as a physical capital boom and prices it: a 1 GW AI campus costs about $41 billion, and a 188 GW US buildout through 2032 requires nearly $9 trillion, 3.2% of GDP per year. As hyperscaler capex outgrows cash flow, financing moves into leases, joint ventures, private credit and special-purpose vehicles, which raise asset-level leverage and make exposures harder to see.

## Key numbers

- $41 billion: estimated cost of a 1 GW AI campus ($41 million per MW)
- $8.973 trillion: US data center investment 2025-2032 in the central scenario
- 3.22%: average annual share of GDP, above railroads 1870-1890 (2.24%) and telecom/fiber 1996-2003 (1.10%)
- $3.55 trillion: mature annual revenue needed for a 10% unlevered return on the buildout

## What it reports

- Cost per MW: about $11 million for the building, $2 million for power infrastructure, $28 million for GPU racks and IT; a 200 MW campus costs about $8.2 billion, two-thirds of it IT equipment.
- Cleanview data (accessed 12 September 2026): about 63 GW of operating US data center capacity and a 494 GW pipeline (47 GW under construction, 447 GW planned).
- Central scenario: 188.3 GW completed 2025-2032, 114.9 GW after 2032, 213.5 GW never completed (32.6% of projects cancelled).
- At 3.22% of GDP per year the buildout exceeds canals (0.66%), railroads (2.24%), electrification (0.50%), highways (1.13%) and telecom/fiber (1.10%).
- Capex by Oracle, Microsoft, Amazon, Meta and Alphabet rose from about $97 billion in 2020 to over $400 billion in 2025 and is projected above $800 billion in 2026, exceeding their combined operating cash flow for the first time.
- Meta's Hyperion project: about 2.0 GW and $30 billion; the Beignet joint venture raised $27 billion of debt, a debt-to-asset ratio of about 90% with debt service coverage of 1.12x.
- Moody's estimates about $970 billion of hyperscaler lease commitments, $660 billion of them not yet on balance sheets.
- A 10% unlevered return requires about $6.4 per billed GPU-hour at 80% utilization, close to frontier rental prices in August 2026. Data center REIT market beta rose from about 0.5 to around 1.

## How they got it

The author builds bottom-up campus costs from hardware and construction assumptions, applies them to project-level pipeline data with realisation rates by stage, and grows costs and GDP at 4% a year. The financing analysis uses public deal disclosures, rating agency estimates and a detailed case study of the Hyperion transaction. Four appendices document the calculations.

## Limits

- The buildout figure is a scenario, sensitive to large multi-phase projects and missing completion dates; the author says it is not a forecast.
- Revenue and return calculations rest on assumed margins, utilization and hardware lives, all uncertain.
- The paper states distress is not imminent; it identifies conditions under which losses could spread rather than measuring them.

## Why it matters for companies

The price companies pay for AI services depends on whether this capacity earns its required revenue, which the paper puts at about 8.8% of 2032 GDP. Buyers signing multi-year AI contracts should know the supply side is financed with high leverage and short-lived hardware, which pushes vendors to maximise utilization and price. Pension or treasury holdings in private credit and infrastructure funds may carry indirect exposure to a few tenants. Railroads and fiber both overbuilt before demand arrived; the paper uses those cycles as the comparison.

Original: https://www.nber.org/papers/w35865