§I
It is one web, not seven balance sheets
The AI build-out is usually described as five rich companies spending their own money, plus two laboratories spending investors’ money. The Q3 2026 filings describe something else: a single web in which the same dollar is simultaneously a hyperscaler’s purchase commitment, a laboratory’s take-or-pay floor, a neocloud’s borrowing base, an SPV’s leased asset, and a retirement saver’s equity and credit exposure. Read the parties one at a time and each looks merely aggressive. Read the edges between them and a different object appears.
The nodes, as filed:
Microsoft Hyperscaler
Microsoft — the strongest balance sheet in the web, and still: $364.6 billion of leases signed but not yet commenced at September 30, 2026, up from $106.2 billion a year earlier, commencing FY2027–FY2034 (10-Q, Note 12). Add $194.1 billion of purchase commitments and $34.6 billion of construction commitments at June 30. Its uncommenced leases alone now equal roughly 54% of its $677 billion commercial backlog.
Alphabet Hyperscaler
Alphabet — Q3 capex $51.4 billion; free cash flow −$9.07 billion, a second consecutive negative quarter. Filed purchase commitments of $924.4 billion, leases not yet commenced of $130.5 billion, and a contingent layer of $98.5 billion — $75.7 billion of backstops on other parties’ data-center and energy obligations, plus $22.8 billion of financial guarantees — that comes due only when a counterparty fails. All-in filed total: ~$1.175 trillion.
Amazon Hyperscaler
Amazon — the quarter’s quiet detonation. Q3 cash capex of $77.26 billion; free cash flow −$33.5 billion in the quarter, −$42.4 billion trailing twelve months. Total contractual commitments of $883.6 billion at September 30 — nearly double the $439.7 billion of nine months earlier — and long-term debt nearly tripled in the same nine months, $65.6 billion to $173.1 billion.
Meta Hyperscaler
Meta — Q3 capex $34.9 billion against $32.8 billion of operating cash flow: free cash flow −$2.12 billion, the first time equipment purchases have outrun the ad business’s cash since 2012. Non-cancelable commitments of $454.52 billion plus $355.73 billion of leases not yet commenced (Q3 10-Q). Its flagship campus is financed by an SPV Meta does not own the debt of — ~$27 billion borrowed by a vehicle 80% owned by Blue Owl funds — while Meta carries residual-value guarantees with a ~$28 billion threshold, and its board has authorized a new share class to raise up to $95 billion more.
Oracle The leveraged edge
Oracle — the web’s leveraged edge. FY2026 capex $55.7 billion against $32.0 billion of operating cash flow; total debt $129.5 billion; backlog of $664 billion, roughly half of it a single customer. S&P downgraded it to BBB− in July 2026, citing leverage above 4x and that concentration. Its credit-default swaps have repriced from ~20 to over 200 basis points while its equity story stayed sunny.
Four of the five are now free-cash-flow negative in the same season. The fifth is guiding ~$190 billion of calendar-2026 capex. This is the funding layer of the web, and it is already under water.
§II
The laboratories owe the landlords
The labs do not own the means of their own production. They rent — on terms that do not flex with demand.
Anthropic’s S-1, filed October 26, 2026, itemizes the floor: Google Cloud ≥$114.5 billion (to July 2033), Amazon ≥$110.0 billion (to April 2036), Microsoft Azure ≥$44.9 billion (to May 2033) — minimum payments, with shortfall true-ups, regardless of usage — plus $161.2 billion of non-cancelable Broadcom equipment leases. Floors and leases sum to $430.6 billion. On the books at June 30, commitments totaled $587 billion, up 5.6x in six months. Anthropic’s own H1 2026 revenue was $16.3 billion.
The landlord lends the tenant the rent, restricted to paying the landlord.
The paper holding this structure together is circular. Amazon holds every outstanding Anthropic convertible note — $6.2 billion of principal. Anthropic carries those notes on its books at $108.6 billion; Amazon carries the same notes at $97.9 billion. One security, two balance sheets, a $10.7 billion disagreement — and the notes’ appreciation is, at once, roughly 90% of Anthropic’s headline $80.7 billion “loss” and tens of billions of Amazon’s paper gains. Broadcom holds up to $42 billion of Anthropic notes whose proceeds may be used only to pay rent on Broadcom equipment, at 9% payment-in-kind rising to 13%. The landlord lends the tenant the rent, restricted to paying the landlord.
OpenAI’s disclosed, priced compute deals sum to ~$710–720 billion — Oracle ~$300 billion, an incremental $250 billion to Microsoft, $138 billion to AWS, $22.4 billion to CoreWeave — against an annualized revenue run rate of roughly $40–50 billion. The annualized cost of the disclosed book is an estimated $117–136 billion a year: 2.3–2.7x current revenue, before salaries. Nvidia, the web’s dominant supplier, has capped a lease guarantee for OpenAI’s Ohio campus at $105 billion, holds a $99 billion equity portfolio concentrated in its own customers, and stands behind the neocloud layer with tens of billions more in backstops. The famous “$100 billion Nvidia investment in OpenAI” never closed; the check that cleared was $30 billion. The supplier is a shareholder in its customers, a guarantor of their landlords, and the recipient of their capex — which is also its revenue.
§III
The same dollar, five costumes
Any total for this web is a construction, and this paper refuses to print a fake one. The overlaps are not errors in the data; they are the data — the mechanism by which the web transmits a failure from any node to every node:
- Anthropic’s $114.5 billion owed to Google sits inside Alphabet’s $560.8 billion cloud backlog. One obligation, two companies’ headline numbers.
- OpenAI’s $300 billion to Oracle is roughly half of Oracle’s backlog and the borrowing base for Oracle’s bonds and ~$72 billion of project debt on its campuses.
- CoreWeave’s $104 billion backlog is the mirror image of Microsoft, OpenAI and Meta purchase commitments; its $35.1 billion of debt — $1.6 billion three years ago, at average rates near 11% — is lent against those same contracts.
- Meta’s Hyperion campus is one asset described three ways: ~$27 billion of SPV debt to Blue Owl’s funds and PIMCO, a ~$12.31 billion initial lease commitment on Meta’s books, ~$28 billion of Meta residual guarantees beneath it.
- Moody’s counts $1.2 trillion of hyperscaler lease commitments, more than $820 billion not yet commenced, and treats them as debt-equivalent. The Bank for International Settlements counts ~$662 billion of “shadow borrowing” in SPVs — larger than the same companies’ combined on-balance-sheet debt. These aggregates are sums of the nodes above, not additions to them.
Michael Burry’s tally — above $3 trillion once guarantees and SPV backstops are counted — and Jim Chanos’s “four-in-one bubble” are arguments, labeled as such here; their inputs, however, are the filings above. OpenAI’s “$1.4 trillion of commitments” is a spoken claim; its own investor plan was ~$600 billion through 2030. This paper uses the priced sums and the filed floors, and flags everything else.
§IV
The waves
The obligations do not share a due date. They arrive in tranches, on a printed schedule, for a decade:
Wave 0 — now
Wave 0 — now. Cash capex: $202 billion in Q3 2026 alone across four companies. Alphabet, Amazon and Meta free-cash-flow negative in the same quarter.
Wave 1 — 2027
Wave 1 — 2027. Morgan Stanley’s credit desk names 2027 the largest free-cash-flow-deficit year; one house forecasts hyperscaler bond issuance jumping to ~$400 billion. Oracle’s OpenAI contract starts billing at ~$60 billion a year. Anthropic’s Broadcom leases come online.
Wave 2 — 2027–28
Wave 2 — 2027–28. Depreciation: the 2025–26 asset placements begin expensing against earnings even if cash spending stops — on five-to-six-year schedules, for hardware generations that turn in eighteen months.
Wave 3 — 2027–2031 and beyond
Wave 3 — 2027–2031 and beyond. The uncommenced leases commence: >$820 billion group-wide. Rent begins when buildings energize, whether or not demand did.
Wave 4 — 2028–2066
Wave 4 — 2028–2066. The maturities. Microsoft’s $25 billion priced on October 29, 2026 runs in ten tranches from 2028 to 2066. Anthropic’s floors run to 2033 and 2036.
And beneath the waves, the trapdoor: Alphabet’s $48 billion of executed backstops trigger only on a counterparty’s failure. The contingent layer is scheduled, by construction, to come due at the exact moment the firm layers are also failing. In 2008 this property had a name — correlation — and it was the whole disaster.
§V
The demand side is a presumption
Nothing above requires a forecast. What follows is the one judgment in this paper, stated with its evidence and its limits.
The build is sized on an adoption presumption: that the public will take up AI assistants the way it took up search — universally, habitually, at full depth. The evidence available says the presumption is doing the work of a fact. Outside technology circles, the median reaction to these systems is that they are somewhat creepy; the public’s own summary of the technology — “a giant prediction machine” — happens to be technically accurate, and a true sentence cannot be marketed away. Revealed use by the median adult is a search bar: a recipe, a fact, the tab closed. The deep, hours-long engagement is real, but it lives in pockets — the young, the immersed, the companion spenders — and the industry’s retention dashboards are measured on the people who stayed, which is to say on the pockets. The professional readers of human beings — the educators and psychologists whose work is perceiving children — draw the delegation line hardest: they will show the tool a photograph of a bicycle drivetrain and will not hand it a child’s evaluation. Every builder’s plan assumes it becomes the next Google, the default front door. The front door has one chair in it. Five victory scenarios were financed; at most one of them describes the future, and the obligations of the other four are identical in every terminal state.
The bills are fixed, dated, and filed; the revenue is a prayer.
This paper does not quantify future demand; nobody honestly can. The claim is narrower and, being narrower, harder to escape: the servicing of these obligations requires adoption near the heroic end of the range, on schedule — and adoption is the one variable in the system that nobody controls, nobody has measured in the population that matters, and the median member of that population currently declines. I am not clairvoyant. I look at systems. In this one, the bills are fixed, dated, and filed; the revenue is a prayer.
§VI
No counterparty
In every previous bubble, the loss could at least in principle be transferred — sold to a greater fool, laid off on a counterparty, socialized onto a named rescuer. Follow the beneficial ownership of this web and there is no outside party left to transfer it to.
Microsoft’s October 2026 proxy names its only holders above 5%: Vanguard at 7.48% and BlackRock at 7.30% — index managers, which is to say plumbing: the pooled 401(k)s, IRAs and pension savings of households. The same two names top the registers of Alphabet, Amazon and Meta; institutions hold roughly 65–80% of every builder. Federal Reserve data put households’ equity share of financial assets at a record 48.2% in mid-2026, above the dot-com peak. The credit side is the same public wearing a different hat: the >$100 billion of 2025 hyperscaler bonds — and the $250–400 billion forecast for 2026–27 — are bought by bond funds, insurers and pension plans. The New York pension fund currently suing Oracle is suing it as a bondholder, alleging the company concealed its coming debt needs weeks after an $18 billion offering. PIMCO — a manager of retirement money — anchors the Hyperion SPV’s debt. The Bank for International Settlements projects the private-credit layer reaching $300–600 billion by 2030, much of it pension and insurance allocations.
So the loss, when the waves land, does not travel to “Wall Street.” It is already home. The public is the gambler, holding the builders’ equity through its index funds. The public is the casino’s creditor, holding the builders’ bonds through its pension funds. The public is the landlord’s mortgage holder, funding the SPVs that own the campuses the builders rent. Every pocket the loss could be moved into has the same hand in it. It all looks good on paper — until you check whose money it finally is.
Method and evidence. Figures are drawn from Q3 2026 SEC filings (10-Q/10-K/8-K), Anthropic’s S-1 of October 26, 2026, Federal Reserve Z.1 data, and published agency tallies (BIS, Moody’s), compiled November 1, 2026. Each figure in the companion ledger carries its source, its as-of date, and an evidence grade. Company statements (guidance, spoken commitment totals) are labeled as statements, not filings. Skeptics’ constructions (Burry, Chanos) are labeled arguments. Known double-counts are disclosed in §III rather than netted away, because the double-counting is the finding: in this web, every obligation is somebody else’s asset, and the somebody is the same public at every layer.
The author does not claim to be an expert, and does not claim to be right. He claims the inputs can be checked.