Other People's Money: Who Is Really Funding the AI Buildout

A chart from Michael Burry's newsletter has been going around: Other People's Money, part VI of his "Heretic's Guide to AI's Stars" (Scion Asset Management, Cassandra Unchained). It is dense: nine boxes, six numbered failure points and a lot of abbreviations. I had trouble reading it, so I spent an afternoon tracing it end to end. I wanted to know how the chain works, which numbers hold up, how big the exposure really is, and what I'd watch to know whether he's right.

Short version: the plumbing is real and well documented. The scary headline number rests mostly on his assumptions. And the risk sits less with the tech giants than with the people who lent them the money.

The chain


flowchart TD
  S["Retirees buy annuities"] --> I["Life insurers<br>Athene, Security Benefit"]
  I --> R["Affiliated Bermuda reinsurers"]
  R --> P["Asset managers / sponsors"]
  P --> V["SPV bonds<br>Beignet, Sopaipilla"]
  K["Banks<br>construction loans"] --> L
  V --> L["Data-centre landlord"]
  L -- "lease" --> T["Hyperscaler tenant<br>Oracle, Meta, Microsoft..."]
  T -- "rent services the debt" --> V
  N["Nvidia<br>residual value guarantees"] -.-> V

Read it like a call stack:

Retirees buy annuities.

Life insurers invest the premiums, increasingly in private credit. They also move some of the liabilities to affiliated Bermuda reinsurers, where capital rules are lighter.

Asset managers originate the deals and steer that money into special-purpose vehicles (SPVs), which issue bonds to build data centres.

Banks provide short-term construction loans, which are meant to be refinanced into those bonds later.

The hyperscaler signs a long lease and pays rent, and the rent services the debt.

The point of the structure is that the debt sits on the SPV's balance sheet, not the tenant's. A lease that hasn't started yet doesn't appear as a liability at all.

What checks out

I checked his main claims against filings and press coverage.

Claim	Status	Notes
Beignet: $27.3B bond for Meta's Hyperion at 6.58%	✅	Priced Oct 2025, anchored by PIMCO. Traded at 94.4 in late September 2026, its lowest since issue.
Sopaipilla: bond for Meta's El Paso campus	✅	$12.3B, July 2026. I couldn't confirm his 7.53% yield.
Oracle's force majeure notice at Jupiter	✅	Could defer rent by up to three years. The cause is power and permit delays. It is not a default.
Oracle-linked bank loans under water	✅	The $18B Jupiter loan is quoted at 89–91.
NAIC capital charge: 6.8% → 30% base	✅	Takes effect at year-end 2027. Falls to 15–24% with enough collateral. Covers all collateral loans, not just AI debt.
Nvidia's $500B financing platform	✅	Announced Aug 10 with six firms, three of which own life insurers.
Nvidia's $105B residual value guarantee	✅	A capped guarantee on OpenAI's Ohio campus.
Guarantees booked at $0	✅	Payment is judged "not probable" under GAAP, so no liability is recorded.
$1.19T of leases not yet started	✅	Adds up correctly from the latest filings.
About $65B of Oracle-linked bank loans	⚠️	Reported figures add to about $56B.
Insurer ratios (379% of surplus, 82×)	❓	From statutory filings I couldn't access.
Loss rates (Oracle 80%, Microsoft 65%, Meta 30%)	❌	His own assumptions, not reported figures.

The last row matters most. Box 6 multiplies each company's not-yet-started leases by a loss rate: the share he expects to end up written off as dead weight. That multiplication produces the frightening total. Change Microsoft's rate from 65% to 10% and about $180B disappears. The commitments are facts, but the percentages are a forecast about AI demand over the next decade.

The $1.19T is rent, not bonds

This is easy to misread. The $1.19T is the total future rent the five hyperscalers have committed to for buildings that aren't operating yet. The leases run 15–20 years and start between 2026 and 2036. The bonds and loans that finance the buildings are a separate, much smaller pile: what I could confirm adds up to under $200B. The rent is what pays them off.

Spread over 15–20 years, the rent is roughly $60–80B a year once all the leases have started. Here it is against current numbers:

Company	Leases not started	Revenue, last 12 months	Leases ÷ revenue	Est. rent as % of operating cash flow
Oracle	$288B	$72B	4.0×	31–41%
Meta	$347B	$228B	1.5×	13–18%
Microsoft	$329B	$332B	1.0×	9–12%
Amazon	$137B	$776B	0.2×	4–6%
Alphabet	$85B	$446B	0.2×	2–3%
Total	$1.19T	$1.85T	0.64×	8–11%

Rent is estimated as commitments ÷ 15–20 years. Amazon's figure includes warehouses and offices, not just data centres. Part of Meta's total is leases from its own joint-venture SPVs.

So is it manageable?

For four of the five, the rent on its own is affordable. Three things still keep me from relaxing.

The cash is already spoken for. Rent at 8–11% of operating cash flow sounds small, but most of that cash already goes on chips and construction. The better comparison is rent against free cash flow, which is what's left after capital spending:

Company	Free cash flow, last 12 months	Est. rent per year	Rent as % of free cash flow
Alphabet	+$53B	$4–6B	8–11%
Microsoft	About +$67B*	$16–22B	25–33%
Meta	+$38B	$17–23B	46–61%
Amazon	−$7.6B	$7–9B	Already negative
Oracle	−$23.7B (FY26)	$14–19B	Already negative

*My calculation: operating cash flow minus capital spending. This excludes finance leases, so Microsoft's true free cash flow is lower.

Alphabet is comfortable and Microsoft manageable. Meta's rent would take about half its current free cash flow, and its free cash flow is falling fast: $0.8B in Q2, down from $8.5B a year earlier. The leases also cover only buildings and power. The GPUs inside are a separate bill, and they wear out in five or six years, not twenty. That's fine at 20–30% revenue growth and tight if growth slows.

Oracle is the outlier. Its commitments equal four years of revenue, and the rent would take a third or more of its operating cash flow. Its free cash flow is negative and its credit rating is one notch above junk. Oracle's answer is $664B of contracted customer revenue, much of it from OpenAI. So Oracle's ability to pay is mostly a bet on OpenAI's ability to pay Oracle.

The risk is designed to sit with the lenders. The tenant gets flexibility, and the lender carries the fixed obligation. Jupiter is the example: Oracle can push rent out by up to three years while the banks carry the loan. The closest analogy I have is a cloud reserved-instance contract with a cheap exit clause. The customer is fine either way. The provider, who borrowed to buy the hardware, is the one left holding idle capacity. In this chain, the provider's balance sheet belongs to insurers and banks, and behind them to annuity holders.

So "the hyperscalers can afford the rent" and "the financing chain is fragile" are both true. They aren't in tension. That's how the structure was designed.

Reading the bond prices

The most useful real-time signal is what these loans and bonds trade at. But a price mixes two things, and you have to separate them.

A bond's price falls when investors demand a higher yield. That happens for one of two reasons. Interest rates may have risen everywhere, which says nothing about AI. Or this particular borrower may look riskier, so its spread over Treasuries widens. It's like a latency metric that mixes overall network load with the health of one service: subtract the baseline before blaming the service.

That's why the two instruments in this story say different things:

Beignet pays a fixed 6.58%. Its drop to the mid-90s is partly rising rates and partly worry about the deal. Burry himself calls it "not a credit mark".

The Oracle-linked loans pay a floating rate of SOFR + 2.5%. Their interest resets as rates change, so rising rates barely move the price. A floating-rate loan at 90 cents means the market doubts it will be repaid in full or on time. That makes it the cleaner warning sign.

A price drop doesn't hurt the borrower on existing debt, because its rate is locked in. It hurts the holders, who take paper losses, and it jams the pipeline. Banks lent short-term planning to sell the loans on to long-term investors. At 90 cents, selling means taking a 10% loss, so the banks hold them. That's step 2 of Burry's unwind.

Signals to watch

🔭

This is the checklist I'm keeping. No single signal settles anything. A cluster of them moving the same way would.

Signal	Worry if	Relax if
Oracle-linked loan prices	They fall into the low 80s	They recover toward par (100)
Beignet vs Meta's own corporate bonds	Beignet falls while Meta's regular bonds hold: the market is pricing the off-balance-sheet structure, not Meta	The gap narrows
Spreads vs Treasury yields	Spreads widen while Treasury yields are flat	Prices move only with Treasuries, which is the general economy rather than AI
New deal pricing	New SPV bonds price at wider spreads or fail to sell	Spreads tighten and banks sell their loans on at par
Jupiter force majeure	It's accepted and other tenants follow	It's rejected or the project stays on schedule
OpenAI revenue vs its commitments	Revenue lags what it has contracted to spend	Revenue keeps pace
Oracle free cash flow	It stays deeply negative as leases start	It turns positive
Leases not yet started	They keep growing faster than revenue	They start and earn their keep
Guarantee accounting	Any company books a liability against a guarantee ("not probable" becomes "probable")	Guarantees stay at $0 while the assets perform
Insurers' appetite	Life insurers cut private-credit buying ahead of the 2027 NAIC rule	They keep buying

Dates on my calendar (based on the usual reporting schedule):

Late October 2026: Q3 earnings from Microsoft, Alphabet, Meta and Amazon. Watch their not-yet-started lease totals and free cash flow.

Mid-December 2026: Oracle's Q2 FY27 results. Watch free cash flow, contracted revenue (RPO) and any update on Jupiter.

1 February 2027: the delayed completion date for the gas pipeline that will power Jupiter.

Early 2027: insurers' annual statutory filings, where figures like "379% of surplus" come from.

31 December 2027: NAIC's new capital charges take effect.

Takeaway

Burry's chart is a bear case, and he doesn't hide it. Most of its plumbing is accurate. Its headline loss number rests on percentages he chose.

What I find most worth remembering isn't the size of the exposure but where it sits. The AI buildout has been financed so that the companies with the strongest balance sheets keep the flexibility, including the option to delay paying. The people with the least flexibility, retirees holding fixed annuities, sit at the far end of the chain.

Whether that ever matters depends on one thing: whether AI revenue grows into the rent. The prices above are where the market will tell us first.

Not financial advice. Figures as reported through early October 2026.