The comparison has been running for months: AI capital spending today, the housing bubble of 2006. Usually it stops at the gesture — lots of money, lots of debt, this will end badly.

Put the numbers side by side and it gets more interesting. One of the most frequently cited parallels is off by a factor of twenty-one. Another, which almost nobody mentions, is larger than the entire on-balance-sheet debt of the companies involved. And the mechanism that made 2006 work in the first place cannot operate in the AI case at all.

How the 2006 loop actually worked

The dominant product in the US subprime market was the 2/28: two years at a fixed rate, then twenty-eight years floating, resetting every six months against six-month LIBOR plus a margin of typically five to six percent. Between 2004 and 2006, more than half of all newly written subprime mortgages were built this way, with another twenty percent as 3/27s. The gap between the initial rate and the fully indexed rate at origination ran between 300 and 600 basis points.

These loans were never meant to survive their own reset.

According to work by the Federal Reserve Bank of Boston, for most annual cohorts a large share of these loans — typically over seventy percent — had already been repaid before the rate adjusted. Repaid here means refinanced. The house had gained value over two years, that gain became equity, and the equity carried the next loan. Another 2/28.

That is the crux: the risk was never in any single loan. It was in the assumption that there would be a next one.

And that assumption had three conditions that had to hold simultaneously: house prices rise, rates stay low, credit stays available. Remove one and the refinancing fails — leaving a borrower facing a payment jump he was never underwritten for.

The engine ran smoothly until roughly 2004. The OFHEO repeat-transactions index rose about ten percent in both 2004 and 2005. By the third quarter of 2006 it was running at one and a half percent annualised.

How big that market really was

The scale of 2006 is routinely underestimated.

And one figure that explains much of the rest: roughly forty percent of all net private-sector job creation between 2001 and 2005 came from housing-related industries.

And how big the AI buildout is

The parallel most people draw — and why it fails

The comparison almost always runs through securitisation. Opaque structures, pooled loans, risk passed along.

Run the arithmetic. Subprime MBS outstanding in 2007: $1.3 trillion. Data-centre ABS outstanding in 2026: $61 billion. A factor of twenty-one.

The growth curve genuinely does rhyme — from $4bn to $61bn in six years, from three to twelve percent of the esoteric ABS market. Private-label subprime did the same thing between 2003 and 2006, going from $37bn to $127bn per quarter.

But a similar curve at one twentieth the level is not a similar danger. Name securitisation as the main parallel and you have the right shape at the wrong scale.

The parallel that does hold

It sits where the real problem sat in 2006 too: off the balance sheet.

Moody's puts the data-centre leases signed but not yet commenced by the major providers at roughly $662 billion. Under the prevailing accounting standard these appear only at lease commencement, so they never enter the capex figures analysts examine. That sum exceeds the entire on-balance-sheet debt of the same companies.

Structurally this is the pattern of the 2006 special-purpose vehicles: the published number is not wrong. It is incomplete. And the rule that makes it incomplete is publicly documented and still not read alongside it.

Two further parallels survive:

A valuation assumption props up reported profit. In 2006 it was model pricing of CDO tranches. Today it is the depreciation schedule: five to six years on the books, against a hardware life that may run two to three. Michael Burry and others estimate the understatement at roughly $176 billion across 2026 to 2028. In both cases profit rests on an assumption nobody outside can verify, and whose correction works backwards.

A cohort with a known maturity date runs into an unknown environment. 2/28 loans written in 2005 and 2006 reset in 2007 and 2008, into falling prices. GPU-backed debt from 2023 to 2025 matures in 2026 and 2027, coinciding with peak depreciation on exactly those purchases. The date was fixed at signing. The environment at that date was not.

Where the analogy inverts

Here is where the comparison flips — in favour of the AI case and against it at the same time.

The 2/28 loop worked because the collateral appreciated. The rising house price was the engine: it manufactured the equity for the next round. No appreciation, no loop.

A GPU cannot do that. It loses value because the next generation arrives — by design, not by cycle. CoreWeave's $7.5 billion facility led by Blackstone used GPUs and customer contracts as security, at roughly eleven percent floating, with repayment beginning in January 2026. Precisely as the collateral's market value was falling sharply.

This cuts both ways. There is no self-reinforcing upswing of the 2004-to-2006 kind, where the collateral finances the next round. There is also no rescue by appreciation. The AI case has the reset without the carousel that preceded it.

The actual loop sits elsewhere: not in the collateral but in demand. Interlocking commitments among chipmakers, cloud providers and AI labs, vendor stakes in customers, take-or-pay contracts. That makes demand look more independent than it is — but the precedent for it is Lucent and Nortel around 1999, not subprime in 2006.

The difference that matters most

In 2006 the leverage sat with households. Home ownership at sixty-nine percent, mortgage debt at ninety-seven percent of GDP, a negative saving rate. A price decline hit the consumption of millions directly — and forty percent of the preceding years' job creation hung on the same industry.

The AI buildout is corporate investment. The leverage sits with a handful of firms carrying the largest cash flows on earth. That is changing — Alphabet's long-term debt more than doubled to $98 billion over the first half of 2026, Amazon's jumped 81 percent to $119 billion in the first quarter alone, and Alphabet's free cash flow turned negative in the second quarter. But it is not an over-extended household facing a rate reset.

So the transmission channels differ. In 2006 the path ran from default to foreclosure to a collapse in consumption. Today it would run through credit markets and equity valuations. The first is slow and broad. The second is fast and narrow.

A note on the sources

While writing this we hit a contradiction that shows how unsettled even well-documented history remains.

In 2006 testimony to the US Senate, the Center for Responsible Lending described the 2/28's low initial rate as a teaser used to draw debt-strapped families into loans they could not later service. A working paper from the Federal Reserve Bank of Boston contradicts this directly: there was never anything like a low teaser rate on the typical subprime ARM.

Both are serious sources on the same product. Cite one without knowing the other and you have written nothing false while telling a story that is disputed.

What holds up

Scale decides, not shape. Securitisation looks identical and is twenty times smaller. Off-balance-sheet commitments look unremarkable and exceed total balance-sheet debt.

A loop needs an engine. In 2006 it was the rising house price. The AI case has no equivalent — the collateral declines on schedule. That makes the upswing less self-reinforcing and the downswing less cushioned.

And whoever carries the leverage sets the speed. Millions of households break slowly. A credit market breaks fast.

None of this says whether or when anything happens. It says which numbers to watch if you want to know: depreciation schedules, free cash flow, maturities from the 2023-to-2025 vintages — and not the headlines about new spending commitments.

---

This piece compares historical and current data. It is not investment advice, not a recommendation, and not a forecast.

Sources: Federal Reserve History, "The Great Recession and Its Aftermath" · Federal Reserve Board, Monetary Policy Report, February 14, 2007 · Federal Reserve Bank of Boston, Public Policy Discussion Paper 08-2 · Federal Reserve Bank of New York, Staff Report 318 · Inside Mortgage Finance, via Corvid Partners and Robert Stowe England · Center for Responsible Lending, US Senate testimony, September 20, 2006 · Conference of State Bank Supervisors, joint statement on subprime lending · deRitis, Kuo, Liang, "Payment shock and mortgage performance", Journal of Housing Economics, 2010 · Adam Tooze, Crashed · Moody's Ratings, early 2026 · Morgan Stanley Research · Barclays Research · PIMCO · Quinn Emanuel, client alert on AI data-centre financing risk, March 2026.