The AI buildout is increasingly financed by the sellers of the equipment
and the buyers of the output, sometimes in the same set of agreements.
A chip vendor funds a model lab. The lab commits spend to a cloud
provider. The cloud provider buys the vendor's equipment. Each leg can be
rational on its own. Together they make the origin and the quality of the
demand harder to read.
The analysis separates three questions: who pays whom, what form the
commitment takes, and how much announced demand is connected to capital
supplied by the recipient.
The map below is that ledger drawn out. Companies sit on the ring in the
order capital tends to move through them: chip vendors first, then the
model labs, then the clouds and the financiers that fund both. Each curve
is one disclosed commitment. Its thickness is the headline dollar value,
and a dashed grey curve is an agreement that has not been signed. Select
a curve, or a company, to open the record behind it.
Explore the ledger
Follow the capital
Drag the diagram sideways to see all of it, or open the table below.
Financing
Cloud and compute
Chips and capacity
Non-binding
Thickness = disclosed dollar value.
Visible relationships
Flow
Headline
Category
Status
Source
Action
Constructed network measure
Circularity score
2026 Q2 disclosure snapshot
12.2%Binding circularity
$102B of $833.4B in
definitive headline commitments is supported by a disclosed return path.
Score definition
Connected commitments, not recycled cash.
For each dollar-denominated arrow, the score looks for a path back to
its payer through no more than two additional relationships. The amount
counted is capped at the weakest leg, so a small investment cannot make
an entire large commercial contract circular.
Binding score
12.2%
All announcements
10.9%
Capacity-only deals
Excluded
Interpretation
The binding score exceeds the all-announcement score because the
$100B NVIDIA letter of intent enlarges the announced denominator without
adding a disclosed dollar path back to NVIDIA. Non-binding value therefore
increases the network total without increasing loop-supported value.
Filing-period comparison
When circularity enters the network
BindingAll announcements
Period
Binding
All announced
Loop-supported value
What changed
2025 Q3
0.0%
0.0%
$0B
No disclosed dollar return path.
2025 Q4
0.0%
0.0%
$0B
Commitments expand, but remain one-way.
2026 Q1
10.5%
9.3%
$76B
Amazon and OpenAI create the first bounded loop.
2026 Q2
12.2%
10.9%
$102B
Amazon and Anthropic add a second bounded loop.
Current two-way structures
The relationships driving the score
Amazon ↔ OpenAI
$38Bsupported in both directions
$50B from Amazon to OpenAI, and $38B
back. Only the smaller direction counts: a return path cannot
support more value than it carries.
Amazon ↔ Anthropic
$13Bsupported in both directions
$13B from Amazon to Anthropic, and $100B
back. Only the smaller direction counts: a return path cannot
support more value than it carries.
These matched amounts add to $51B, while the score above
counts $102B. The two are not in conflict. The
score measures arrows, not pairs: in a two-way structure both directions
are supported, each capped at the same smaller leg, so a pair contributes
twice what it matches.
How this score is calculated
The score starts with each disclosed dollar commitment—for example,
Company A paying Company B—and asks whether other disclosed commitments
carry value back from Company B to Company A. That return can be direct
or pass through one other company.
Only the amount supported by the smallest commitment in the loop counts.
If Company A commits $100B to Company B, but only $20B is disclosed in
the return direction, $20B—not $100B—is treated as loop-supported. When
several return routes exist, the calculation uses the strongest one and
does not count the same starting commitment twice.
Binding score: definitive agreements only.
All-announcement score: definitive agreements plus
letters of intent and memoranda. Capacity agreements stated only in
gigawatts are excluded because they cannot be added to a dollar total.
Each period is a calendar quarter-end snapshot. A relationship enters
the calculation on its public announcement date. This is a constructed
network measure, not an audited filing metric; SEC backlog and remaining
performance obligations usually do not name the customer needed to map
a return relationship.
Capex and cash generation
The ledger above records commitments between companies. This chart asks a
different question about the same buildout: how much of it the buyers can
pay for themselves. If the largest spenders were funding this out of the
cash their own operations throw off, the financing structures above would
matter far less than they do.
Every figure here comes from a filed 10-K cash flow statement, pulled
through the SEC's XBRL company-facts API. The ratio is cash purchases of
property and equipment divided by net cash from operating activities, for
each company's own fiscal year.
That line is deliberately named rather than called capital expenditure.
Filers define capital expenditure differently in their own reporting:
some net proceeds or incentives against it, and some add principal
payments on finance leases to the figure they publish. The accounting tag
varies as well, and Amazon files this line under a different one from the
other four. Naming the cash flow line is what keeps the five comparable,
and the tag behind every observation is shown in the table below.
The path is not the same for all five. Microsoft rises steadily, from 27
cents of purchases per dollar of operating cash flow in FY2022 to 63 cents
in FY2026, and Alphabet from 34 to 56. Amazon and Meta both spent heavily
in 2022, pulled back sharply in 2023, and have climbed since: Amazon from
136 to 62 and back to 94, Meta from 62 to 38 and back to 60. What they
share is the recent direction, not the starting point.
Oracle is the one that crosses over. It sat between 37
and 51 cents through FY2024, then reached 102% in FY2025 and 174% in
FY2026, meaning it spent $1.74 on property and equipment for every dollar
of cash the business generated. The difference has to come from somewhere,
and that somewhere is the capital stack.
These purchases include spending outside AI infrastructure, because the
statements do not provide a complete AI-specific breakout. The ratio
measures spending relative to internally generated cash, not the purpose
of each expenditure. Fiscal years are not aligned either: Microsoft closes
in June and Oracle in May, so the latest column is FY2026 for those two
and FY2025 for the rest.
Conclusions
Everything above is measured: a ratio taken from filed cash flow
statements, a ledger of disclosed commitments, and a score computed from
that ledger. What follows is inference drawn from those measurements, and
it is separated here for the same reason the constructed metric is
labelled as constructed. A reader who accepts the numbers is free to
reject the reading.
The buildout has outgrown the cash it generates. Only
Oracle spends more than its operations produce. For the other four,
purchases of property and equipment now absorb most of operating cash
flow, which leaves less for everything that cash also has to cover:
dividends, buybacks, debt service, acquisitions. The marginal
dollar of this buildout is increasingly an external dollar, which makes
the terms on which it is raised, rather than the size of the
announcements, the variable that decides whether the pace holds.
Circularity is new, small, and concentrated in one firm.
No disclosed dollar return path existed anywhere in this ledger through
the end of 2025. Two quarters later 12.2% of definitive
commitment value sits on one. Both loops run through Amazon: equity into
a model lab, a cloud commitment back. What the score currently measures
is one company's strategy, not a market-wide structure, and it would be
wrong to describe it as the latter.
Most disclosed value still travels one way. 87.8% of definitive commitment value has no disclosed
return path at all. The strong form of the circular-financing claim,
that the buildout is substantially funded by the firms selling into it,
is not supported by what these companies have actually disclosed. It is
not refuted either. A loop no party has announced is invisible to this
method, and the register is 11 sources deep.
The hardware leg of the loop is unpriced. The
3 capacity arrangements carry no dollar value
in any disclosure. The leg the circular thesis leans on hardest, the
equipment purchase that carries money back to the chip vendor, is the
leg the announcements price least often. Every dollar figure on this
page therefore describes a loop with one side missing, and the
measured score should be read as a floor rather than an estimate.
What this implies for the market
Watch the funding mix, not the announcement total.
Headline commitment value is the most reported and least informative
number in this market: it aggregates signed contracts with letters of
intent, and spans terms that are mostly undisclosed. The tractable
version of the question sits in
the financing sections of the same cash flow statements the chart above
is built from. Rising external funding against a flat commitment total
would say more than another record quarter of announcements.
Revenue quality is a counterparty question, not a sector
question. Where a supplier has funded its customer, some part of
the revenue it recognises is a function of the capital it supplied. At the
disclosed level that describes a minority of value and concentrates in a
few named pairs. Claims about circular revenue across the sector are
therefore working at the wrong unit of analysis. The pair is the unit,
which is why this ledger records counterparties rather than totals.
The disclosure regime is what keeps the question open.
Remaining performance obligations are filed without naming the customer,
so the contracted version of this score, built from binding obligations
rather than announcements, cannot be computed by anyone outside the
companies themselves. Until that changes, the alarmed reading of this
market and the reassuring one rest on the same evidence: press releases
written by the participants.
What would change this reading
Three developments would move the conclusions above, and each is worth
watching for. A dollar value attached to any of the capacity arrangements
would price the missing leg and could raise the score sharply. A return
path disclosed between two companies other than Amazon and its partners
would make circularity a market structure rather than one firm's
approach. A filed document contradicting a headline value already in the
ledger would shrink the denominator and cast doubt on the rest of it. The
ledger and the method are published so that whichever arrives first is
visible, including if it is the one that makes this page wrong.
None of the above has to be taken on trust. The table below holds every
annual fact the collector pulled from the SEC, with the filing it came
from and the data label it was reported under, so any figure in the capex
chart can be checked against the document that reported it.
SEC company facts
Inspect the filings
844 normalized facts · 9 public issuers
Period end
Fiscal year
Value
Filed
SEC data label
Source
Annual 10-K observations only. Backlog is shown as a concept candidate, not automatically as management guidance.
Method
Values are annualised only when a term is disclosed. Every numeric record
resolves to a numbered source. Binding status follows the underlying
document: a letter of intent never becomes a contract because it is widely
reported.
That rule is enforced twice, and the two gates behave differently on
purpose. The collector refuses to write a dataset at all if a relationship
names a source that is not declared, so a broken ledger cannot be
published. The site build is deliberately softer: it flags the record on
the page and warns in the terminal, so a half-finished entry can still be
previewed while it is being researched.
The ledger currently holds 13 relationships between
11 companies, backed by 11 sources.
10 carry a disclosed dollar value, totalling
$933.4B of announced headline value, of which
$833.4B (89.3%) sits
under a definitive agreement. The remaining 3 are
stated only as capacity, 22 GW across those
records, and carry no dollar figure at all. That gigawatt total counts
only the capacity-only arrangements; it is not a measure of all capacity
disclosed across the ledger, since a priced deal can mention capacity the
figure does not include.
Limitations
Private companies disclose little. Headline values often span uncertain
terms, and an announcement can describe capacity rather than minimum
spend. Equity ownership changes over time. The ledger captures disclosed
relationships: not side letters, pricing, utilisation, intent, or the
ultimate source of every dollar.
Some headline figures are cumulative rather than a single commitment. The
$13B Amazon investment in Anthropic combines $8B placed before 2026 with
the $5B announced in April, and the ledger dates the whole total to the
announcement. That overstates what was newly committed in 2026, and
because this relationship is one of the two that produce the circularity
score, the score inherits the same timing problem. Splitting the record in
two is the fix, and it waits on a source for the earlier tranche.
3 of the 13 relationships are
stated only in gigawatts of capacity and carry no disclosed dollar value.
They are drawn on the map and listed in the ledger, and they are excluded
from every dollar figure on this page, because capacity cannot be added to
a currency denominator.
Every relationship in the ledger rests on a company publication: a press
release or a blog post written by one of the two counterparties. None of
them rests on a document filed with a regulator. Company announcements
state headline values that filings frequently do not corroborate in the
same form, and closing that gap is the largest open item on this project.
Scaling AI for everyone OpenAI · Company blog · published 2026-02-27 · retrieved 2026-09-02 Backs: Amazon to OpenAI; NVIDIA to OpenAI; SoftBank to OpenAI
EDGAR XBRL company facts and submissions API U.S. Securities and Exchange Commission · Dataset · retrieved 2026-09-02 Every figure in the capex chart, and every row of the filings table.
