Brookings Says AI Capex Now Exceeds the Railroad Mania — The 1873 Lesson Is About Who Carries the Debt, Not Who Builds the Chips

Brookings's \"Financing the AI Buildout\" (Stijn Van Nieuwerburgh, Columbia; summary by David Skidmore, Sep 23, 2026) projects $10.3 trillion of U.S. AI infrastructure investment — data-center buildings, power systems, networking, and specialized chips — from 2025 through 2032. That works out to about 3.6% of GDP per year, a level higher than every major U.S. buildout the authors benchmark against. The historical ranking in Van Nieuwerburgh's underlying paper puts railroads (1865–1890) at roughly 2.4% of GDP, the Interstate Highway System at 1.6%, electrification at 1.1%, and the 1996–2003 telecom build at 0.8%.

The reason the gap matters: in the 1870s, railroad capex consumed roughly 15–20% of all U.S. net capital formation at its peak — enough to crowd out other investment and enough to break the financial system when the bond market turned. AI capex at 3.6% of GDP, layered on a much larger economy, is doing the same thing to private credit and the investment-grade debt market, but through off-balance-sheet structures the bond market can't see.

The Panic of 1873 did not start with a railroad accident or a demand collapse. It started with a bank — Jay Cooke & Company, the largest U.S. financier of the Union war effort — unable to roll Northern Pacific Railway mortgage bonds on Sep 18, 1873. Cooke had borrowed heavily to build the transcontinental line and could not sell the bonds abroad because German and Central European capital had withdrawn after the Vienna real-estate bust earlier that year. Within weeks, 55 U.S. railroads had failed; another 60 followed by the first anniversary, the New York Stock Exchange had closed for ten days from Sep 20, and the U.S. entered a depression that ran to 1877 and peaked in unemployment of 8.25% in 1878.

Van Nieuwerburgh's paper isolates the equivalent of Cooke's bond book — and shows it has already grown materially. Hyperion, Meta Platforms' 2 GW Louisiana data-center project, was financed through an SPV called Beignet that raised about $27 billion of external debt in October 2025 — the largest single investment-grade corporate debt issuance in U.S. history. Beignet's debt carries an A+ rating (one notch below Meta's), a 6.58% yield roughly 100 basis points above Meta's own corporate curve, and 90% asset-level leverage. The bond matures in 2049 and is repaid from five successive four-year lease payments by Meta — a structure that, under current GAAP, leaves both the lease and the residual-value guarantee off Meta's balance sheet.

The point is not that any one SPV will fail. The point is that the buildout is structurally dependent on the same opaque, off-balance-sheet, asset-level leverage that economists now flag as the systemic-risk channel — the same channel Cooke's bondholders held in 1873, with the same dependency on a small number of investment-grade tenants whose cash flows must remain intact for the structure to hold.

In the 1870s the upstream bottleneck was steel, ties, and locomotives — and the companies that supplied them captured outsized returns until the bust. The 2026 upstream bottleneck is power: grid equipment, transformers, switchgear, gas turbines, and behind-the-meter fuel cells. The order books at the equipment vendors are starting to look like the rail-equipment order books of the late 1860s.

If a Brookings-style scenario plays out — even a partial one — the upstream equipment vendors carry the most asymmetric upside in the buildout, with multi-year contracted backlog that survives the kind of demand wobble that would hit the SPV-financed projects first. They are the closest modern analog to a 19th-century rail-equipment supplier.

The freight rails — Union Pacific, CSX, Norfolk Southern — are the most direct operating analog to the 1870s railroads: capital-intensive, network-effect businesses with pricing power that strengthens in a downturn as truck-to-rail conversion accelerates. Q2 2026 results show the operating model is intact: Union Pacific reported operating revenue of $6.9 billion, up 12% year-over-year, with freight revenue of $6.5 billion. Class I volume rose across the board in Q1 2026 — CSX +6%, Norfolk Southern +4%, CN +5%, Union Pacific +2% on a revenue-ton-mile basis.

The data-center REITs — Equinix, Digital Realty — sit in a less comfortable position. They are the closest listed analog to the financiers, not the operators: their revenue depends on hyperscaler tenants whose own capex is the capex that drives the bubble. Equinix trades near 24x 2026 AFFO per share on guidance of 10–12% AFFO growth; Digital Realty trades around 22x core FFO with Q2 revenue up 29% to $1.9 billion. Both stocks have already drawn explicit AI-slowdown warnings from management — Equinix's CEO pushed back, but only after the share price moved first.

Three parallels are real. AI capex as a share of GDP now exceeds the railroad peak (3.6% vs 2.4%). The financing has migrated off-balance-sheet into SPVs, private credit, and securitizations — exactly the structural migration that turned the 1870s railroad bond market into the panic's ignition point. And the cash-flow claims of the off-balance-sheet vehicles depend on a small number of investment-grade tenants — Meta, Microsoft, Alphabet, Amazon, Oracle — whose ability to perform is the entire system's load-bearing assumption.

Three breaks matter. The 1873 financiers had no federal backstop; today's hyperscalers fund in markets where the U.S. Treasury is the marginal price-setter of risk-free rates and a coordinated Fed response is the baseline expectation. The 1873 railroads sold into a competitive freight market with no pricing power; modern rails and utilities operate under regulated or quasi-regulated pricing models that hold up in downturns. And the 1873 bankruptcies were equity-funded by bondholders with no recovery; today's SPV debt is bankruptcy-remote, with collateral pools, residual value guarantees, and credit enhancements that change the recovery math even if the operating thesis breaks.

Short-term (days to quarters). The first signal is hyperscaler capex guidance. Microsoft has guided roughly $175 billion of 2026 capex (per a FactSet compilation cited in the WSJ), Amazon around $220 billion, Alphabet $195–205 billion, Meta Platforms $130–145 billion. A single downward revision from any of them resets the financing math for roughly $300 billion of contracted SPV debt that depends on the tenant's lease commitments holding. Watch the next two earnings cycles and the Moody's commentary on lease off-balance-sheet treatment — that is the modern equivalent of the bond market failing to roll Cooke.

Long-term (1–3 years). The structural implication is who carries the buildout when the buildout slows. The 1873 analog says bondholders and construction suppliers take the loss; the operating rails recovered within a decade. Today's analog points the same way: GE Vernova, Eaton, Quanta Services, and Bloom Energy carry backlog that survives a 2027 demand air pocket; Blue Owl Capital and other private-credit lenders carry the asset-level concentration; Equinix and Digital Realty carry the multiple. At 22–24x forward FFO, a 200-basis-point rise in cap rates wipes out several years of AFFO growth. The investor question for the next 36 months is not whether the buildout slows — Brookings's framework implies it must, because 3.6% of GDP is unsustainable. The question is which listed names have already locked in enough contracted revenue to outlast the air pocket, and which are still priced for a buildout the bond market will not roll at the current coupon.