---
title: "The trillion dollars that isn't on any balance sheet: mapping hyperscaler shadow debt, LLC by LLC"
description: "Meta's $27.3B Beignet bond shows on Meta's books as a $2.9B investment. Microsoft's $66.6B of finance leases sit in 'other liabilities'. Alphabet's $43.8B of backstops carry at $815M. SpaceX's GPU leases got forced back on-book by its own auditor. And $1.09 trillion of signed data-center leases appear on no balance sheet at all. The full map, from filings."
canonical: https://evidinvest.com/blog/hyperscaler-shadow-debt-spv-map-2026-08-22
date: "2026-08-23"
last_updated: "2026-08-23"
author: "EvidInvest Team"
publisher: "EBD Sweden AB (EvidInvest)"
tags: ["META", "MSFT", "GOOGL", "AMZN", "ORCL", "NVDA", "SPCX", "SPV", "shadow debt", "private credit", "SEC filings"]
---

# The trillion dollars that isn't on any balance sheet: mapping hyperscaler shadow debt, LLC by LLC

- Published: 2026-08-23
- Author: EvidInvest Team, EvidInvest (published by EBD Sweden AB)
- Canonical URL: https://evidinvest.com/blog/hyperscaler-shadow-debt-spv-map-2026-08-22
- Research, not investment advice.

> Meta's $27.3B Beignet bond shows on Meta's books as a $2.9B investment. Microsoft's $66.6B of finance leases sit in 'other liabilities'. Alphabet's $43.8B of backstops carry at $815M. SpaceX's GPU leases got forced back on-book by its own auditor. And $1.09 trillion of signed data-center leases appear on no balance sheet at all. The full map, from filings.

*Figures trace to: Meta 10-K `0001628280-26-003942` and 10-Qs `0001628280-26-028526` / `0001628280-26-050705`; Microsoft 10-K `0001193125-26-323660`; Alphabet 10-Q `0001652044-26-000071`; Amazon 10-Q `0001018724-26-000026`; Oracle 10-K `0001193125-26-277521`; Blue Owl Real Estate Net Lease Trust 10-Qs (CIK 1944366); deal terms from IFR, Bloomberg, S&P and issuer releases as cited. The Silicon Valley 101 documentary on this topic prompted the dig; every number below was independently re-verified against the primary source, and several were corrected. Research, not investment advice.*

Last week we wrote that the hyperscalers [flipped from buybacks to
bonds](https://evidinvest.com/blog/hyperscaler-buybacks-to-bonds-2026-08-22) — roughly $210
billion of debt raised in their latest reported periods. That number is
real. It is also the *visible* layer. Underneath the bond tallies sits a
second architecture of special-purpose vehicles, residual-value
guarantees, credit backstops booked as derivatives, finance leases filed
under "other liabilities," and — the largest single item — **$1.09
trillion of signed data-center leases that have not commenced and
therefore appear on no balance sheet at all**.

The Bank for International Settlements has a name for this:
**shadow borrowing** (BIS Quarterly Review, March 2026). The name sounds
like an accusation. It isn't one — every structure below is legal, rated,
audited and disclosed. But it is *complex by design*, and the complexity
has a purpose. Below is the map, company by company, entity by entity,
with the actual lenders named where they are disclosed — and then the
honest question: is this hidden leverage, or is this simply what
financing looks like when an industry needs more capital than any
balance sheet was built to carry?

## The scoreboard first

| | Bonds raised, latest period | Recognized lease liabilities | Signed, NOT yet on balance sheet |
|---|---|---|---|
| Microsoft | $0 | $88.5B ($66.6B finance + $21.9B operating) | **$329.1B** |
| Meta | $24.9B (H1'26) + $39.8B via SPVs | (leases on book) | **$278.99B** |
| Oracle | $46.1B (FY26) | $37.9B | **$260B** |
| Amazon | $82.4B gross (H1'26) | ~$110B+ | **$137.2B** |
| Alphabet | $56.2B gross (H1'26) | $20.6B | **$85.2B** + $43.8B backstop notional |

Uncommenced leases across the five: **~$1.09 trillion**, against roughly
$285 billion of lease liabilities actually recognized. Under lease
accounting this is legal and disclosed — in the footnotes. A lease that
has not commenced books nothing, however binding the signature.

## Meta: the Beignet machine

The cleanest case study in the group, because the bond documents leaked
into the rating agencies' and one lender's own filings.

![The Beignet stack: where Meta's $27.3B lives](https://evidinvest.com/blog/beignet-stack-2026.png)

Meta's Hyperion campus in Richland Parish, Louisiana is financed by
**$27.294 billion of privately placed senior secured notes — 6.581%,
fully amortizing, due May 30, 2049** — issued in October 2025 by an
entity called **Beignet Investor LLC**. None of that debt is on Meta's
balance sheet. What Meta carries instead is an equity-method investment
of **$2.92 billion** (June 30, 2026 — it was $1.83B at year-end, $2.37B
in March; the venture keeps calling capital).

The chain, reconstructed from S&P's presale, the Blue Owl Real Estate
Net Lease Trust's own 10-Q, and Meta's Note 5:

- **Laidley LLC** owns the campus and holds 15-year power agreements
  with Entergy Louisiana.
- **Project Beignet Holdings LLC** owns Laidley. This is "the Venture."
- **Beignet Investor LLC** owns 80% of Holdings and issued the $27.3B.
- **Beignet Pledgor LLC** owns the issuer and pledged that equity to the
  trustee — the bondholders' collateral is the SPV stack itself.
- Above Pledgor: funds affiliated with **Blue Owl Capital** — its
  non-traded net-lease REIT (ORENT, which disclosed a 63% indirect
  stake in the issuer at close for a $1.533B commitment), its $7B
  Digital Infrastructure Fund III, its Digital Infrastructure Trust,
  plus undisclosed co-investors.
- Meta's 20% sits in a Meta subsidiary named **Iris Crossing LLC** —
  the entity that also posts Meta's guarantee.
- A Meta subsidiary named **Pelican Leap LLC** leases the eleven
  buildings back on four-year triple-net terms renewable to twenty
  years, starting 2029. Initial lease commitment: **$12.31 billion**.

Why doesn't Meta consolidate a venture whose buildings it designed,
constructed (it provides construction management, administrative and
property management services), leases entirely, and guarantees? Because
under the VIE rules the "activity that most significantly impacts the
Venture's economic performance" was determined to be **remarketing the
campus** — deciding what happens if Meta ever leaves — and that power
belongs to Blue Owl. Meta is therefore "not the primary beneficiary."
Its auditor EY flagged exactly this consolidation judgment as a
**Critical Audit Matter** in the FY2025 10-K.

And if Meta leaves? The leases run four years against 24-year debt. The
bridge is a **residual value guarantee with an aggregate threshold of
approximately $28 billion**, declining over time, covering the first 16
years (a termination fee covers the rest). Meta's filing then delivers
the sentence that makes the whole structure work: *"RVG payments are not
probable, and therefore no liability has been recorded."* Meta's own
disclosed **maximum exposure to loss: $46.03 billion** — against a $2.92
billion carrying value.

The lenders: **PIMCO anchored ~$18 billion** of the notes (after
beating Apollo and KKR for the mandate, per Bloomberg); BlackRock took
$3B+; per S&P Capital IQ data cited by the FT, **116 managers across
384 funds** hold the paper — PGIM, Natixis, T. Rowe, Invesco, Fidelity,
JPMorgan — with insurance and retirement money throughout. Morgan
Stanley was sole bookrunner. S&P rates the bonds **A+**, one notch below
Meta's AA– — the market prices it as Meta risk with a haircut, whatever
the org chart says.

**Deal two, "Sopaipilla"** (El Paso, ~1GW, ~$14B campus): **$12.547
billion** priced July 27, 2026 at **7.534%, due 2048** — a full 62.5bp
wider than Beignet priced, and demand was soft. Same 80/20 anatomy, but
the 80% is **BlackRock** — specifically its GIP infrastructure-equity
and HPS private-credit arms — with J.P. Morgan joining Morgan Stanley
on the books. Meta's RVG this time: **~$13 billion**, disclosed as a
subsequent event in the Q2 10-Q. Two SPV bonds in nine months:
**$39.8 billion** of debt that services Meta data centers and appears
on no Meta balance sheet. Meanwhile Hyperion itself has been expanded
to a 5GW, $50B+ program — the obvious candidate for pastry number three.

For calibration: Meta paid roughly 6.58% via Beignet when its own
public curve was near 5.5%. On $27.3B that is close to **$300 million a
year of extra interest, paid for balance-sheet cosmetics**.

## Microsoft: the lease ledger

Microsoft has issued no new public bonds since 2021, and its reported
debt fell to ~$40 billion. Our last piece called it the only hyperscaler
funding the buildout from operating cash. That was the visible layer.

![Microsoft: the debt line is the smallest number in the filing](https://evidinvest.com/blog/msft-obligations-2026.png)

From the FY2026 10-K: **finance lease liabilities of $66.6 billion**,
up from $46.2 billion a year earlier — recorded not on the debt line
but inside "other current liabilities" ($4.3B) and "other long-term
liabilities" ($62.3B). Total lease payments committed under finance
leases: $89.7 billion. Weighted average term: 13 years. These are
mortgages in everything but name — and they are one and a half times
the bond debt.

Then the footnote that resets the scale: *"As of June 30, 2026, we had
additional leases, primarily for datacenters, that had not yet
commenced of $329.1 billion."* Commencing fiscal 2027 onward. Not a
dollar of it recognized yet — over $130 billion of it added in the June
quarter alone, per Bloomberg's tally.

Who owns the buildings? Microsoft doesn't say. Publicly known
counterparties: **$60B+ of neocloud capacity commitments** (Nscale
~$23B, Nebius $17–19B, IREN $9.7B, CoreWeave, Lambda), the traditional
landlords (Blackstone's QTS, KKR/GIP's CyrusOne, Brookfield's Compass,
DigitalBridge/Silver Lake's Vantage), and its own **AI Infrastructure
Partnership** with BlackRock/GIP and Abu Dhabi's MGX — a $30B equity /
$100B-with-debt vehicle that has already acquired Aligned Data Centers
for ~$40 billion.

Two more FY26 details complete the picture. Commercial paper is back on
the balance sheet ($6.7B short-term debt). And in July Microsoft
announced that from FY2027 datacenter depreciation stretches from 15 to
**25 years** and more future leases will classify as *operating* rather
than finance leases — which pulls them out of reported capex, trimming
the FY2027 plan from ~$190B to ~$175B without changing a dollar of
actual spending.

## Alphabet: selling insurance instead of borrowing

![Alphabet: $43.8B guaranteed, $815M carried](https://evidinvest.com/blog/googl-backstops-2026.png)

Alphabet's method is the most creative: it doesn't borrow for third-party
capacity, and mostly doesn't own it — it **guarantees other people's
debt** so that they can borrow to build capacity Alphabet's ecosystem
needs. In the filings these backstops are booked as **credit
derivatives**: notional up from $16.9 billion at year-end to **$43.8
billion at June 30, 2026** — carried on the balance sheet at **$815
million**, under 2% of the exposure. Terms run up to 15 years, and on a
default Alphabet can step into the underlying lease. On top: $7.6B of
supplier backstops for long-lead power equipment and another $24.1B of
backstops agreed but not yet finalized.

Where the guarantees point (all disclosed via the counterparties, not
by Alphabet): the **Fluidstack** web serving Anthropic's TPU capacity —
TeraWulf's Lake Mariner buildout (a **$3.2B Morgan Stanley construction
bond** exists because Google backstops the leases; Google took warrants
for ~14% of TeraWulf), Cipher Mining's Barber Lake ($1.4B backstop,
~5.4% warrants), TeraWulf Abernathy (~$1.3B), Fluidstack directly
(~$1.8B). And in June 2026 the structure went industrial-scale: a **$35
billion financing for Anthropic's five-site TPU expansion, led by
Apollo with Blackstone**, underwritten not against Anthropic but
against Google's backstop of the lease payments. As one project-finance
analyst put it: "The lenders are not underwriting Anthropic. They are
underwriting Google."

## Amazon: the honest borrower, mostly

Amazon does it the old way: in its own name. The March 2026 offering —
$37B in dollars plus €14.5B — was **the largest corporate bond sale in
history**, followed by C$14B in June and $25B more in July: over **$72
billion of 2026 issuance**, all on balance sheet, with the 10-Q stating
plainly that more financing should be expected. Alongside that:
**$137.2 billion of not-yet-commenced leases** (it was $106B one
quarter earlier) and $130B of unconditional purchase obligations. Even
the "honest" one carries a quarter-trillion of committed spend beyond
its bonds.

## Oracle: the biggest bet on the thinnest base

![Oracle: books vs signed vs promised](https://evidinvest.com/blog/orcl-promises-2026.png)

Oracle signed **$260 billion of data-center lease commitments** —
fifteen-to-nineteen-year terms, substantially all commencing fiscal
2027–2029, none of it on the balance sheet (recognized lease
liabilities: $37.9B). Buried in the same note: Oracle **guaranteed up
to $3.3 billion of one lessor's own borrowing** — the Alphabet
technique, appearing at the company least able to afford it. Remaining
performance obligations jumped from $138B to **$638 billion** in one
year — overwhelmingly OpenAI's Stargate commitments. S&P cut Oracle to
**BBB–** in July; its five-year CDS traded at ~203bp, the widest since
the financial crisis.

Who actually owns Oracle's Stargate shells: Crusoe with **Blue Owl** and
Primary Digital at Abilene ($15B JV); **Vantage** (DigitalBridge/Silver
Lake) at the Texas "Frontier" and Wisconsin campuses, financed by a
**$38 billion JPMorgan + MUFG debt package** — the largest data-center
debt raise on record; **Related Digital with Blackstone equity and
~$10B of PIMCO anchor debt** in Michigan — a deal Blue Owl declined on
underwriting standards and PIMCO took anyway; Blue Owl again in New
Mexico ($18B).

## Why none of this counts as debt: the five doctrines

Every structure in this piece leans on one of five accounting rules.
None of this is fraud; all of it is disclosed — in footnotes. The rules:

**1. The lease-commencement rule (ASC 842).** A signed lease books
nothing until the asset is delivered. Sign $329 billion of 20-year
datacenter leases that start next year and your balance sheet today
records zero. This single rule carries the $1.09 trillion.

**2. The VIE primary-beneficiary test (ASC 810).** You consolidate a
special-purpose entity only if you have "power over the activities that
most significantly impact its economic performance." Meta designs,
builds, operates, fully leases and guarantees Hyperion — but "remarketing
the campus" was designated the most significant activity, Blue Owl holds
that power, so $27.3B of debt lands on nobody's consolidated books but
the SPV's own.

**3. Guarantee fair-value accounting (ASC 460/815).** A guarantee is
carried at what it is "worth" today — the probability-weighted expected
payment — not at what you would owe if called. Alphabet's $43.8B of
backstop notional carries at $815M. Meta's $28B RVG carries at zero,
because payment is "not probable."

**4. Lease classification (ASC 842 again).** A finance lease is debt in
substance but is presented in "other liabilities," not the debt line —
Microsoft's $66.6B. And an operating lease books only the present value
of committed rent — Meta's 4-year lease terms against 24-year SPV debt
book $12B against $27B, with the RVG (carried at zero) bridging the gap.
Reclassify future leases from finance to operating, as Microsoft will
from FY2027, and reported capex falls with no change in cash spent.

**5. The failed-sale-leaseback doctrine (ASC 842-40) — the one that
bites back.** If a "lease" transfers the asset's economics to the
lessee, auditors must recharacterize it as a loan. This is the rule
SpaceX just collided with, below — proof that the perimeter of these
structures is a judgment call an auditor can reverse.

## The Musk complex: where the structure got recharacterized

The same architecture reached Elon Musk's companies with less
polish and more leverage — and produced the one case where an auditor
pushed the debt back on-book.

**The GPU SPV.** xAI's Colossus buildout is financed through **Valor
Compute Infrastructure L.P.** — a fund run by Valor Equity Partners
(Antonio Gracias, SpaceX director and, through Valor entities, holder
of 500M+ SpaceX Class A shares). The design mirrors Beignet, with GPUs
instead of buildings: the SPV raises equity and debt (~$7.5B/$12.5B
program target), buys NVIDIA GB200/GB300 systems, and leases them to
xAI's datacenter subsidiary **CTC Property LLC** on 5-year triple-net
terms with a purchase option. Three lease tranches (October 2025,
January 2026, April 2026) total roughly **$20 billion of undiscounted
obligations**, guaranteed by SpaceX.

The financing chain is a closed circle worth drawing: **NVIDIA is the
anchor equity investor (~$1.9–2B) in the SPV that buys NVIDIA's chips**
— the vendor capitalizing its own customer and booking the revenue.
**Apollo provided $3.5B of the debt** in the January tranche, reportedly
routed onward into Athene, its annuity insurer — retirement money, again.
Diameter Capital took part of the debt. Michael Burry called the
structure "fugazi"; Apollo called it "downside-protected."

**Then the auditor arrived.** In SpaceX's IPO filings, PwC concluded the
Valor arrangements were **failed sale-leasebacks — "loans in substance,
not leases"** — and roughly **$9.0 billion of related-party debt to
Valor went ON SpaceX's balance sheet**, with the GPUs staying in
property and equipment. It is the only structure in this piece that an
auditor has so far refused to leave outside the perimeter — a useful
benchmark for what Beignet would look like if the judgment call ever
went the other way. The related-party geometry is also the sharpest in
the industry: SpaceX's GPU lessor is one of its largest shareholders,
sits on its board — and, as we showed in our
[SpaceX 13F census](https://evidinvest.com/blog/who-owns-spacex-q2-2026-13f-deadline-2026-08-14),
is the second-largest institutional holder of the stock at $86 billion.

**The rest of the Musk stack**, briefly: legacy xAI corporate debt cost
**12.5%** (since refinanced through a $20B Goldman bridge and ~$25B of
post-IPO SpaceX senior notes); the Memphis turbine fleet is financed at
**9.85%** through Stonebriar and housed in **Stateline Power LLC**, a
50.1/49.9 JV with Solaris that Solaris consolidates and SpaceX
equity-methods — 900MW of power plant off SpaceX's balance sheet, the
Meta trick at utility scale. Against all this sit the compute-rental
contracts that service the leases: **Anthropic at $1.25B/month through
May 2029 (~$45B)**, Google at $920M/month, Reflection AI at $150M/month.
For price calibration across the piece: Meta's SPV money costs ~6.6%,
Oracle's public money ~5–6%, Musk's stack ran 9.85–12.5% before the IPO
laundered it into investment-grade curves. Same machine, three price
tiers of the same collateral idea.

## The lender web, and who holds the bag

Rearrange the deals by financier and the names repeat: **Blue Owl**
(Meta Hyperion, Stargate Abilene and New Mexico), **PIMCO** (Meta
anchor, Oracle Michigan), **Blackstone** (QTS, Michigan equity, the
Anthropic loan), **BlackRock/GIP/HPS** (Sopaipilla, Microsoft's AIP,
Aligned), **Apollo** (Anthropic lead), **KKR** (CyrusOne), **Brookfield**
(Compass), with JPMorgan, MUFG and Morgan Stanley arranging. NVIDIA appears on both sides: anchor equity in the xAI GPU SPV, and in August
2026 it formalized the club: a compute-financing platform with
Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR
targeting **$500 billion of third-party capital**, with NVIDIA offering
residual-value support of up to 25% on select deals — the chip vendor
now underwriting the resale value of its own chips, exactly as Meta
underwrites the resale value of its own data centers.

Behind the asset managers, the money is insurance float, pensions, and
retail semi-liquid funds. Which is why the stress signals matter: Blue
Owl's non-traded credit funds took redemption requests of 19–41% of NAV
in the first half of 2026 against 5% quarterly gates, and the firm sold
$1.4B of assets to fund withdrawals. The data-center ABS market —
securitized rents — has grown from $4B in 2020 to **~$61 billion**.
The US Office of Financial Research and the FSB have both flagged the
interconnection; Senators have asked FSOC to examine the Meta structure
specifically.

## The other side: why they can't just issue bonds

It is tempting to read all of this as concealment. The fuller reading is
that the ordinary tool — a corporate bond on the issuer's own balance
sheet — simply cannot carry this buildout, for four reasons the deal
flow makes visible.

**The market has a ceiling.** Order books are already thinning: Amazon's
oversubscription fell from 5.3x to 1.6x in nine months; new-issue
concessions across the sector went from ~2 basis points to ~12; Meta's
second SPV deal needed a 62.5bp wider spread than its first. Morgan
Stanley puts data-center capex at ~$2.9 trillion through 2028, of which
~$1.5 trillion needs external money. The investment-grade bond market
digests a few hundred billion of *net* new tech paper a year at best.
The money has to come from somewhere else — and "somewhere else" is
insurance and pension capital that wants exactly what these structures
manufacture: long, amortizing, asset-backed, investment-grade paper.

**The tenor has to match the asset.** A data center earns over 15–25
years. Public bond curves get expensive and thin past 10–15 years;
a fully amortizing 24-year private placement wrapped around a specific
building with a specific tenant is, mechanically, a *better-matched*
instrument than a bullet bond — it is how toll roads, pipelines, and
power plants have been financed for decades. Project finance did not
become respectable by accident.

**Ratings are a hard budget constraint.** Oracle put the buildout on its
own balance sheet — the "honest" route — and got downgraded to one notch
above junk, with CDS at 18-year wides. The downgrade is not cosmetic: it
shrinks the eligible buyer base for every future bond. Keeping the debt
in rated SPVs preserves the parent's rating capacity, which *is* the
capacity to keep building.

**And investor mental models break.** These companies traded for a
decade as "net cash, high buyback, capital-light." Piling $300 billion
of construction debt onto that picture doesn't just move ratios — it
evicts the shareholder base that owns the stock for that picture, before
a new base has priced the new reality. CNBC called the buyback cut a
breach of an "unspoken contract." The SPV architecture is, in part, a
managed transition between those two shareholder bases.

This is how the financial system has always metabolized a capital
super-cycle — railroads invented modern project finance, real estate
invented the REIT, mortgages invented securitization. AI compute is
inventing its instrument set in real time: SPV net-lease bonds, GPU
funds, compute-backed credit. The evolution is legitimate. What deserves
scrutiny is not that the structures exist, but that the *headline
numbers most investors read do not add them up* — and that the risk,
carefully carved out of every balance sheet, has not disappeared. It has
moved: into rents, guarantees, and residual values, payable only in the
world where AI demand stalls.

How big is that risk, and what would it take to trigger it? GPU lending
rates falling from 15% to under 6% in three years say the market believes
compute has become a durable asset class — NVIDIA now finances it like
real estate. Whether that belief survives its first recession is the
subject of our next piece.

## What this does to our buybacks piece

Our [previous analysis](https://evidinvest.com/blog/hyperscaler-buybacks-to-bonds-2026-08-22)
stands, but its scale was conservative and one conclusion needs
sharpening. The ~$210B of bonds is the smallest of three layers: above
it sit the on-book-but-not-called-debt items (Microsoft's $66.6B of
finance leases, Alphabet's $43.8B of guarantee notional carried at
$815M), and above those the **$1.09 trillion** of signed leases that
accounting does not yet see — plus $39.8B of SPV bonds and $41B of RVG
exposure at Meta alone. And Microsoft is not the unlevered exception —
it is the *differently levered* exception: no bonds, but the group's
single largest pile of future lease commitments and a doubling finance-
lease book. Funding "from operating cash" today, having promised
$329 billion of tomorrow's.

The capital-return era did not just end. It has been replaced by the
most elaborate corporate financing architecture since the GSE era —
rated, documented, legal, disclosed in footnotes, and largely invisible
in every headline debt number you will read this earnings season.

*Research, not investment advice. Not a price call.*

---

Cite the canonical URL above (https://evidinvest.com/blog/hyperscaler-shadow-debt-spv-map-2026-08-22), not this Markdown file. Research, not investment advice.

## Sitemap

All posts in Markdown: https://evidinvest.com/blog/index.md · feeds: https://evidinvest.com/blog/rss.xml and https://evidinvest.com/blog/feed.json ·
the [index for language models](https://evidinvest.com/llms.txt) · the full [sitemap](https://evidinvest.com/sitemap.md).