---
id: "lib-2026-10-08-the-10-trillion-question-financing-the-ai"
type: "library"
title: "The $10 trillion question: Financing the AI buildout (Brookings Podcast on Economic Activity transcript)"
lang: "EN"
org: "Brookings (Van Nieuwerburgh, Klein, Eberly, Steinsson)"
source_type: "Podcast"
date: "2026-10-08"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-08-the-10-trillion-question-financing-the-ai/"
original_url: "https://www.brookings.edu/wp-content/uploads/2026/10/BPEA-S9E1-financing-the-ai-buildout.pdf"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Brookings podcast in which Columbia's Stijn Van Nieuwerburgh explains his BPEA paper on financing the AI buildout; specific figures from a strong source, but a conversation about the paper rather than the analysis itself."
summary_basis: "transcript"
original_minutes: "40"
tags: ["business transformation", "ai capex", "macro finance"]
frames: ["solow-paradox"]
---

# The $10 trillion question: Financing the AI buildout (Brookings Podcast on Economic Activity transcript)

Tier B (Useful, limited): Brookings podcast in which Columbia's Stijn Van Nieuwerburgh explains his BPEA paper on financing the AI buildout; specific figures from a strong source, but a conversation about the paper rather than the analysis itself.

Summary based on the full transcript.

## In short

Transcript of the Brookings Podcast on Economic Activity (8 October 2026). Aaron Klein of Brookings interviews Stijn Van Nieuwerburgh of Columbia Business School about his fall 2026 BPEA paper "Financing the AI Build-Out". The paper projects US AI data center investment at $10.3 trillion from 2025 to 2032, an average of 3.63% of GDP a year, larger relative to the economy than railroads, highways or telecom. The main concern is financing: hyperscalers increasingly use joint ventures, leases and private credit that keep leverage off their balance sheets.

## Key numbers

- $10.3 trillion: projected US AI data center buildout, 2025-2032
- 3.6% of GDP: average annual investment, versus about 2.2% for the 1870-1890 railroad boom and 1.1% for the telecom boom
- $27 billion: debt issued by Meta's Hyperion joint venture against an asset valued at roughly $30 billion, about 90% debt-to-asset
- $1.4 trillion: debt issued so far on AI data center structures, about half ($700-800 billion) held by banks

## What it reports

- Historical comparison: railroads about 2.2% of GDP (1870-1890 peak), highways about 1.1% (1950s), telecom/fiber about 1.1%, canals about 0.66%, electrification about 0.5%. AI is about 50% above railroads and more than three times the internet boom.
- Cost basis: about $4.1 billion per 100 MW facility in 2025 ($41 million per MW), rising 4% a year; about two-thirds IT equipment, one-third building and power.
- Sensitivity: 5% nominal GDP growth lowers 3.6% to 3.5%; very strong growth brings it to about 3%, hard to go lower.
- Meta's Hyperion (Richland Parish, Louisiana, about 2 GW in the initial phase, up to 5 GW): Meta sold a stake to private credit firm Blue Owl for $2.5 billion and kept 20% of the equity. The joint vehicle issued $27 billion of debt on an asset valued at roughly $30 billion, with debt service coverage around 1.15.
- Meta signed five consecutive 4-year leases plus a residual value guarantee. Under GAAP, obligations less than 70% likely need not be booked, so neither the future leases nor the guarantee appear on Meta's balance sheet, although one of the two is certain to occur.
- At 90% leverage, a 10% fall in cash flows wipes out the equity.
- Hyperscaler capex is about $800 billion in 2026, mostly IT equipment they still own.
- About $1.4 trillion of debt issued on these structures to date; $700-800 billion with banks. Private credit leverage is about 50%, versus about 90% for banks.

## How they got it

The paper builds a bottom-up cost model in megawatts of planned capacity and divides by projected GDP, then studies one deal in detail as a template. The podcast is an interview about that work; the author says his mapping of who ultimately holds the risk is still in progress.

## Limits

- Projections depend on announced capacity; a recession or demand shortfall would shrink the numerator.
- The transcript relays figures; the full paper holds the method. The Hyperion ownership split is stated loosely in the conversation.
- No complete map yet of where losses would land; the $1.4 trillion figure is "by some tally".

## Why it matters for companies

Data centers are being financed like property and infrastructure, with long leases and project debt. For companies buying AI services, supplier pricing and capacity depend on that financing holding up. For finance teams, the case shows how lease options and guarantees can carry commitments that GAAP does not require on the balance sheet. The author asks for disclosure of who owns each asset, who owes the debt and who bears the loss if demand disappoints.

Original: https://www.brookings.edu/wp-content/uploads/2026/10/BPEA-S9E1-financing-the-ai-buildout.pdf