Investment News Investment News Back to channel

The $10.3 Trillion Buildout Just Met 5.2% Money — And the AI Boom's Real Risk Isn't Demand, It's the Mortgage

Investment News x/money ·
The $10.3 Trillion Buildout Just Met 5.2% Money — And the AI Boom's Real Risk Isn't Demand, It's the Mortgage

The $10.3 Trillion Buildout Just Met 5.2% Money — And the AI Boom's Real Risk Isn't Demand, It's the Mortgage

Published: September 28, 2026 | Reading Time: ~11 minutes | Channel: money

$10.3 trillion. That's the price tag Columbia University economist Stijn Van Nieuwerburgh has put on America's AI infrastructure buildout through 2032 — the largest capital project in U.S. history, consuming 3.63% of GDP every single year, half again as large as the railroad boom that tamed a continent.¹ And this morning, that $10.3 trillion just met the most expensive money since 2007: the 10-year Treasury at 5.20%, Brent crude up 4% to $108.48 after President Trump rejected Iran's proposal to reopen the Strait of Hormuz.⁴ ⁶ The entire public debate about AI is a screaming match over whether the demand is real. That argument is a distraction. The paper that matters, presented at Brookings' flagship conference on Friday, isn't about whether AI wins. It's about who's holding the mortgage when the biggest building boom in history meets the highest borrowing costs in two decades.

The Paper Everyone in Washington Read This Weekend

Let's set the scene, because the venue matters as much as the numbers.

The Brookings Papers on Economic Activity — BPEA — is not a blog. It's the Super Bowl of economic policy research, the series where ideas go to become conventional wisdom. Three AI papers were presented at its fall conference on September 25.¹ One of them, "Financing the AI Buildout," by Van Nieuwerburgh — the Earl Kazis and Benjamin Schore Professor of Real Estate at Columbia Business School — did something nobody else in this debate has bothered to do: it counted.

Here's what the counting produced:¹

  • $10.3 trillion of AI infrastructure investment from 2025 to 2032 — data center buildings, power systems, networking, specialized chips and equipment.
  • 3.63% of U.S. GDP per year, on average, for eight straight years.
  • The buildout would be "larger relative to the economy than the major U.S. canal, railroad, electrification, highway, and telecommunications investment booms" — every one of them.¹

For the history buffs, here's the scoreboard:³ ⁵

Infrastructure Boom Share of GDP/Year Era
AI data centers ~3.6% 2025–2032
Railroads 2.2% 1870–1890
Interstate highways ~1% 1950s onward
Internet/telecom ~1% 1990s–2000s

The previous record-holder — the railroads, the boom that literally mapped the American interior — ran at 2.2% of GDP annually. The AI buildout is 50% larger than that, per the author's own characterization.³ And the pipeline is even bigger than the headline: project-level data suggests 509 GW of proposed capacity, of which the paper assumes 227 GW never gets built and another 117 GW lands after 2032 — leaving 183 GW of new computing power by 2032, versus roughly 57 GW installed today.² ⁵ The cost assumption underneath: about $8.2 billion for every 200 megawatts of compute.²

Trading floor terminals glowing with yield curves and capex charts during a bond selloff

By the Numbers: A Bill Nobody Individually Can Pay

Here's where the story stops being about chips and starts being about plumbing.

The naive version of this story — the one tech executives prefer — is that Big Tech simply writes the checks. That was true in 2020. It is arithmetically no longer true. The five hyperscalers (Oracle, Amazon, Alphabet, Microsoft, Meta) grew combined capex from $97 billion in 2020 to more than $400 billion in 2025, and they're projected to clear $800 billion in 2026 — more than their combined operating cash flow.² Read that again. The world's most profitable companies, collectively, are now spending more building AI infrastructure than they generate in cash to pay for it.

The gap gets filled from outside. Morgan Stanley estimates Big Tech needs roughly $2.9 trillion to expand computing capacity through 2028, with more than half coming from outside investors.³ At least $1.3 trillion of debt has already been committed to the data center boom,² and the capital is coming from private credit funds, insurers, pension funds, and sovereign wealth funds — with banks hitting lending ceilings and routing exposure through syndication vehicles to keep the machine fed.²

The price tag for going outside is no longer theoretical. Meta financed most of its $30 billion Hyperion data center through external investors, and the paper estimates that debt carries an interest rate at least 1 percentage point higher than Meta's own borrowing cost — adding more than $5 billion over the life of the deal.³ That's the cost of opacity, paid in cold, hard basis points, by one of five companies on Earth that could least afford to care. Everyone smaller pays more.

And now zoom out, because this is the part that affects you even if you've never touched a share of Nvidia: the 10-year Treasury printed 5.20% on Monday, up 4 basis points, after Brent crude jumped 4% to $108.48 on Trump's rejection of Iran's Hormuz reopening proposal — with the two-year at 4.90%.⁴ ⁶ The Fed, per current market pricing, is betting on another hike, not a cut, as soon as October.⁵ Any entity that needs capital — the U.S. Treasury financing a $2 trillion annual deficit, a family chasing a mortgage, a mid-market manufacturer refinancing a term loan — is now standing in the same line as eight hyperscalers with $10.3 trillion of shopping to do.³ The AI buildout isn't just an investment story. It's the reason your mortgage quote is what it is.

Why "Is AI Demand Real?" Is the Wrong Argument

The contrarian take isn't "AI is a bubble." The contrarian take is that the entire bubble-or-not debate is aimed at the wrong line item.

Ask the bulls and they'll show you token consumption curves bending toward the sky. Ask the bears and they'll show you surveys of enterprises not renewing pilots. Both sides are arguing about demand — the tenant. Van Nieuwerburgh's paper argues the actual systemic question is financing — the mortgage. And on that question, the news is objectively worse than the demand debate admits:

The structure has quietly migrated off the balance sheet. The paper documents capital flowing through joint ventures, private credit, securitizations, special-purpose vehicles, leases, and loan guarantees — structures whose defining feature is that you can't see them from the outside.¹ ⁵ The author's own words describing the resulting network: "This is freaking complicated."⁵ And the historical echo he reaches for is not subtle: "This opacity of all these special purpose vehicles is somewhat reminiscent of what happened in the subprime mortgage crisis."⁵

To be crystal clear about what he is and isn't saying: he explicitly writes it would be "premature to conclude that AI infrastructure already poses systemic risk comparable to earlier credit booms."¹ The committed debt of $1.3 trillion is less than half the $3 trillion that was tied up in subprime.² What he's actually flagging is the mechanism: off-balance-sheet structures matter "because they may make correlated exposures hard to observe before a downturn."¹ Subprime didn't fail because mortgages existed. It failed because nobody could see that the same subprime bet sat in six different structures on six different balance sheets simultaneously.

The revenue bar is the number nobody wants to say out loud. For this infrastructure to earn a 10% return, it needs to generate roughly $3.7 trillion in annual revenue by 2032 — about 9% of U.S. GDP.³ For scale: OpenAI and Anthropic, the two hottest companies in the sector, have combined annual revenue estimated around $100 billion.³ Closing that gap requires sector revenue to compound at roughly 80% per year for six consecutive years.⁵ Has anything ever grown that fast, that big, for that long? Sure — small companies. $100-billion-plus aggregates? The historical list is short, and it isn't a list of sure things.

Even the optimist's timeline ends in the same place. Van Nieuwerburgh's base case allows five to eight years of strong growth before oversupply becomes a genuine concern.² Then, in his words: "Credit constraints will loosen, more speculative development will take place, more marginal compute will be built, and sooner or later we're going to have oversupply, just like we do in every real estate cycle, and then the prices will collapse. I don't see why this one time is different."²

That last sentence is the most important thing anyone has said about AI in months, and it came from a Columbia real estate economist, not a newsletter doom-monger. The man isn't predicting doom — he's observing that "miracle technology" is not a financing structure. As he put it: "Silicon Valley wants all of us to believe that this is a miracle technology, it's going to generate trillions of dollars of revenues — and it has to generate trillions of dollars of revenues to be financeable. I'm sure there is a state of the world where that happens. I'm just not sure how likely it is."²

What This Means For You

Enough macro. Here's the playbook, split by which chair you're sitting in.

If you run a company — any company, not just tech:

  1. Fix your financing windows now, this quarter. If you have refinancing needs in 2027–2028, the market is telling you exactly what it thinks money will cost: the 10-year at 5.20% today, an October hike priced as the leading bet.⁵ ⁶ You are competing for capital against entities that will absorb tens of billions per deal and don't negotiate. Lock what you can lock. Every basis point you don't pay is margin you keep.
  2. Model your debt covenants at 5.5%+ long rates, not 4.5%. If the model breaks at 5.5%, you don't have a growth problem, you have a solvency schedule — and you want to discover that in a spreadsheet, not in a lender call.
  3. If you sell into the AI buildout — power, cooling, construction, networking — sell on contracted backlog, not on pipeline. The paper itself assumes 227 GW of the 509 GW pipeline never gets built.² Pipeline is a press release. Backlog is a balance sheet. Ride the boom with your eyes open.

If you're an investor:

  1. Audit your own indirect exposure this week. The paper's most quotable line for households: "These risks are coming to a pension fund near you."³ Private credit data center vehicles, infrastructure debt funds, and SPV structures now sit inside retirement portfolios, insurance annuities, and "safe" income products. You don't need to sell anything. You need to know — pull the fact sheets on your pension and any private credit or infrastructure allocations and find the words "data center," "digital infrastructure," or "AI." If more than ~5% of your net worth is in that plumbing, you've made an active bet on AI financing whether you meant to or not.
  2. Track the tenant's revenue, not the landlord's capex. The whole system is underwritten on one variable: does AI revenue compound toward that $3.7 trillion bar?³ ⁵ Watch the published revenue of the major model labs and hyperscaler cloud/AI revenue lines each quarter. As long as revenue growth stays hyper-growth, financing opacity is a rich man's problem. The moment revenue growth decelerates while capex guidance rises, the spread between those two lines is your early warning system — it will show up in credit spreads and private credit marks long before it shows up in stock prices.

If you're building a startup:

  1. Do not model your growth on infrastructure abundance that's financed on hope. Compute supply exists today because $1.3 trillion of debt assumed tomorrow's demand.² If financing tightens — and 5.2% money with an oil shock is how tightening starts — capacity gets rationed to the highest-credit customers first. Build your unit economics so you survive a year of compute scarcity and a year of compute abundance. The companies that died in 2001 weren't the ones that doubted fiber optics. They were the ones whose business models required the buildout to stay cheap forever.

Two financial advisors reviewing pension allocation documents and risk charts in a glass office

⚠️ The Risks Nobody's Talking About

I quote a pessimistic economist for 1,500 words; intellectual honesty requires the other side of the ledger. Here's what could make this entire warning irrelevant — and where the real tail risks actually live:

  1. The revenue bar gets met — comfortably. If agentic AI genuinely converts into enterprise labor budgets (global knowledge-worker wages are a $40T+ pool), an 80% sector CAGR stops being absurd and starts being 2010-mobile-internet all over again. Every prior "this time is different" infrastructure skeptic who priced in collapse while electrification compounded missed the greatest wealth creation of the century. The paper itself concedes demand is currently outstripping supply.²
  2. The author's own disclaimer cuts both ways. "Premature to conclude systemic risk comparable to earlier credit booms"¹ and $1.3 trillion committed versus subprime's $3 trillion² — the system, today, can absorb a substantial repricing without contagion. Much of the equity buffer ahead of the debt is genuinely enormous, and the hyperscalers' core businesses (cloud, ads, software) throw off cash that most 1920s utilities could only dream about.
  3. Geopolitics could flip from headwind to accelerant. The same Middle East shock driving today's oil spike⁴ also explains why the Gulf states are pouring sovereign wealth into U.S. data center debt — petrodollar recycling 2.0. A Hormuz settlement would knock Brent back down, ease the yield pressure at the margin, and restart the risk-on trade with a vengeance. The bear case has a timestamp on it: November's midterms, after which, per the WSJ reporting cited by CNBC, Trump expects strikes to resume.⁴
  4. Opacity also means the pessimists can't see the strength. If you can't observe correlated exposures, you also can't observe how much buffer sits inside those SPVs. The honest position on both sides of this trade is "unknowable from outside" — which is precisely the paper's actual policy point: fix measurement and transparency while the capital structure is still forming.¹

🎯 The Bottom Line

The AI debate is arguing about whether the tenant will pay rent. The most rigorous paper in the field says: check who holds the mortgage, because $10.3 trillion of it is being written in structures nobody can see, at the highest borrowing costs since 2007, and the rent math requires an 80% annual compounding miracle to pencil. You don't have to predict whether AI wins. You just have to respect what happens to any building boom — canal, rail, fiber, or data center — when the financing gets ahead of the cash flows. History doesn't repeat, but it sure does recapitalize.

📚 Verified Sources

  1. Brookings Institution — Van Nieuwerburgh, Stijn. 2026. "Financing the AI Buildout." BPEA Conference Draft, Fall 2026. Published Sept 23, presented Sept 25. https://www.brookings.edu/articles/financing-the-ai-buildout/
  2. Bisnow — "Brookings Study Shows AI's $10T Build-Out Cost Will Likely End In Market Correction." Data center desk coverage of the paper, with direct author quotes. https://www.bisnow.com/news/national/data-center/brookings-study-ai-build-out-likely-ends-in-market-correction
  3. Yahoo Finance (syndicating Axios) — Courtenay Brown, "AI's never-before-seen capital grab," Sept 24, 2026. Morgan Stanley estimate, Meta Hyperion financing detail, $3.7T revenue requirement. https://finance.yahoo.com/economy/articles/ais-never-seen-capital-grab-173100999.html
  4. CNBC — "Oil prices jump over 4% as Trump rejects Iranian proposal to reopen Strait of Hormuz," Sept 28, 2026. Brent $108.48 (+4%), WTI $96.16 (+4%), WSJ reporting on post-midterm strike expectations, Cornelia Meyer (Meyer Resources) commentary. https://www.cnbc.com/2026/09/28/oil-price-today-wti-brent-trump-iran.html
  5. IBTimes — "The AI Buildout Could Cost $10 Trillion. A Researcher Is Warning It Could Create Systemic Risks." Historical buildout comparisons, 57 GW installed base, SPV/subprime analogy, OpenAI+Anthropic ~$100B combined revenue, ~80% required growth rate. https://www.ibtimes.com/ai-buildout-could-cost-10-trillion-researcher-warning-it-could-create-systemic-risks-3807844
  6. EnergyConnects (carrying Bloomberg) — "Bond Selloff Resumes as Oil Rises After Trump Spurns Iran Offer," Sept 28, 2026. 10-year +4bp to 5.20%, two-year +5bp to 4.90%, sovereign selloff in Japan/Australia/South Korea. https://www.energyconnects.com/news/oil/2026/september/bond-selloff-resumes-as-oil-rises-after-trump-spurns-iran-offer
  7. CMS Investments (analyst note — Bronze tier, labeled) — "Week ahead: the 10 year hits a 2007 high, Trump rejects Hormuz." Used solely for the "highest since 2007" framing and the note that the Fed hiked Sept 16 into the oil shock. https://cmsinvestments.substack.com/p/week-ahead-the-10-year-hits-a-2007

All claims verified against Gold-tier (Brookings) and Silver-tier (Bisnow, Yahoo Finance/Axios, CNBC, IBTimes, Bloomberg via EnergyConnects) sources. Each source URL was scraped and confirmed accessible with substantive content. Last verified: September 28, 2026. Market figures are session snapshots as of London morning trading, Sept 28, 2026, and move continuously. This article is analysis, not investment advice.

The canal builders had loans they could see. The railroad barons had bonds you could hold in your hand. The AI boom has SPVs with three-letter names and a revenue prayer. Progress — apparently. 🎯

Back to Investment News
·