The Housing Bubble
The most recent system-wide bubble, and the one that reached furthest beyond markets. Its lesson for AI is about plumbing: how financing structure turns a price decline into a crisis.
Updated
inflation="Late 1990s-2002" mania="2002-2006" peak="Mid-2006 (national prices)" trough="2011-2012" recovery="2016-2017" drawdown="27-35% national; >50% in some cities" peakToTrough="~5-6 years" recoveryTime="~10 years" />
What happened
US home prices rose relentlessly from the late 1990s to mid-2006, powered by ever-easier mortgage credit. Price-to-rent and price-to-income ratios hit multi-decade highs while lending standards collapsed into no-documentation loans and teaser rates. National prices (Case-Shiller) fell roughly 27-35% from the 2006 peak to the 2012 trough, with some cities down more than half; nominal recovery took until 2016-17 (overview (opens in a new window)). The financing chain built on those prices, mortgage securities, and the institutions holding them, turned the decline into the 2008 global financial crisis.
What inflated it
Credit growth, in industrial quantities. Subprime and near-prime mortgages proliferated from 2002 to 2006; household debt-to-income set records; and securitization let lenders sell the risk onward, removing their reason to care whether loans could be repaid. The credit-to-GDP gapHow far total lending in an economy has risen above its long-term trend. Readings above about +10 points have preceded many banking crises.Full definition in the glossary, total lending racing above its long-run trend, is precisely the indicator central-bank researchers later found flashing hardest before this crisis, and one of the best-tested early-warning signals across all crises (IMF (opens in a new window)).
What popped it
The Federal Reserve raised ratesWhen a central bank raises interest rates or drains money from the system to slow the economy. Most historical bubbles peaked during or within 18 months of tightening.Full definition in the glossary from 1% in 2004 to 5.25% by 2006, resetting the adjustable mortgages that had made peak prices affordable. Prices topped out mid-2006; delinquencies rose through 2006-07; credit spreadsThe extra interest a risky borrower pays compared to the safest borrower (the US government). Widening spreads mean lenders are getting nervous.Full definition in the glossary on mortgage securities began widening well before the headlines; and the failures cascaded from mortgage lenders to Bear Stearns' funds (2007) to Lehman Brothers (September 2008) (NBER (opens in a new window)). Warnings had been public since 2002-2004, from Robert Shiller among others, three to four years before the peak: correct, early, and expensive to act on, a pattern examined on the Case Against page.
How far it fell, and how long recovery took
Houses themselves fell by a third, severe but survivable. The leverage stacked on top is what turned it systemic: global losses on mortgage-linked securities are commonly estimated in the trillions of dollars, and the recession's cost far exceeded the housing losses themselves. National prices took about a decade to recover nominally.
Who captured the value
The houses stayed useful; the ownership changed. Investors who bought foreclosed homes and distressed mortgage bonds in 2009-2012 earned some of the era's great returns, while the households and institutions that bought at the top absorbed the loss. Once again: the asset endured, and the value migrated to whoever bought after the reset.
Echoes in today's AI boom
- Plumbing determines blast radius. Housing's fall was moderate; the opaque credit chain made it catastrophic. The AI equivalent to watch is the debt layer: private creditLoans made by investment funds instead of banks, with little public disclosure. Tens of billions of it now finances AI data centers.Full definition in the glossary, off-balance-sheet vehiclesA separate legal company created to hold one project and its debt, keeping both off the parent company's books.Full definition in the glossary, and GPU-backed loans, mapped on The AI Money Loop.
- The best indicators are credit indicators. The credit-to-GDP gap that flagged 2008 is, notably, not flashing for the US in 2026, one of the genuinely reassuring readings on the Warning Lights page, with the caveat that AI leverage sits in private structures the aggregate data sees poorly.
- Collateral that can fall. Housing finance assumed home prices only rose; GPU-backed lending assumes chips hold value. The depreciation evidence suggests they hold it for 3-5 years at best.
- Early warnings, again. Shiller was right in 2003 and prices rose for three more years. Bubble calls being early is the norm, not the exception, the central caution on Conclusions.