Brookings: $10.3T AI buildout, financed off the books

Somebody has to pay for the data centers that serve your API calls. A new Brookings paper puts the bill at $10.3 trillion through 2032 and says a growing share of it is being financed in places nobody can see. If you buy tokens by the million, that is not macroeconomic trivia. It is the cost structure under your unit economics.

What actually happened

Stijn Van Nieuwerburgh, a real estate professor at Columbia, presented Financing the AI Buildout at the Brookings Papers on Economic Activity conference. The numbers:

• $10.3 trillion in projected AI investment from 2025 to 2032 — data center buildings, power systems, networking, chips and equipment.
• 3.63% of US GDP per year, on average, for eight years. The paper's comparison set is canals, railroads, electrification, highways and telecom. This one is bigger than all of them relative to the economy. The paper makes that comparison qualitatively and does not publish a GDP share for each historical boom, so neither will we.
• The financing is moving off balance sheets. Van Nieuwerburgh catalogs joint ventures, private credit facilities, securitization vehicles, special-purpose entities, lease obligations and loan guarantees, and warns that risk is "migrating from transparent on-balance sheet financing by major corporations to opaque off-balance sheet financing."
• The collateral is contingent on things nobody controls — AI demand that has not materialized yet, a technology stack that turns over every 18 months, power interconnects, hardware supply, and the creditworthiness of a very short list of tenants.

His policy ask is measurement and transparency now, while the capital structure is still malleable. Note what the paper does not claim: it does not forecast a crash or a date. We are not going to put one in its mouth.

Why AI infrastructure financing matters for your business

Here is the operator translation. Today's token prices are set by companies racing for share, subsidized by capital that expects a return later. That is a good deal for you right now and a fragile one to build a P&L on. We have already watched vendors on this beat cut prices with an expiry date printed on them and then double them on January 1. Those are not pricing mistakes. They are what happens when somebody's interest payment comes due.

So stop treating your model provider as infrastructure and start treating it as a supplier with a balance sheet. Three things we do on every build. Abstract the model layer — route through one internal interface so swapping a provider is a config change, not a quarter of work. Every client we have moved between vendors did it in days because the seam was already there. Know your cost per completed job, not per token. Tokens per dollar changes when a vendor reprices; jobs per dollar is the number that shows up in your margin, and it is the only number that tells you whether a 2x hike is survivable. Keep an open-weights fallback you have actually tested, on a workload you actually run. An untested fallback is a slide, not a plan.

None of this is a bet against AI. It is a bet that the price of it will move, in both directions, faster than your pricing page can.

Source: Financing the AI Buildout, Brookings Institution.