03 · Financing, payment and risk
The round closed.
Did the money arrive?
Some of it demonstrably did, according to the investors' filings. Some remains conditional. Neither fact tells us how much is ultimately earned from independent customers.
Pool topic: Is AI a Bubble? · Question revision 1 · Exact question: Is AI a Bubble?
What this changes
The useful dividing line is no longer “real money or an empty announcement.” It is which money, paid when, under which conditions—and who is exposed if the business disappoints. The evidence documents real funding and substantial commercial ties. It does not prove that every commitment has been paid, that the same dollars came back as cloud revenue, or that the investments will earn adequate returns.
Did it arrive?Follow the Amazon–Anthropic caseWhat does $105bn mean?Inspect the ledger
A closed round can still have more than one payment date
OpenAI's 31 March announcement described a closed round of $122bn in committed capital, at a $852bn post-money valuation. Those words do not say that $122bn arrived in one bank transfer. The earlier February $110bn announcement is an earlier account of that round, not another $110bn to add. March close; February announcement.
Amazon's later filing resolves its own $50bn component: $15bn invested in the first quarter, $13.7bn in the second and $21.3bn after 30 June, reported by the July filing. Our previous report corrected the earlier missing payment evidence. This is investor-reported funding, not an inspection of OpenAI's bank account. Amazon, investment note.
SoftBank's $30bn promise has its own clock
Equal $10bn tranches of the 2026 OpenAI subscription. Status comes from dated SoftBank disclosures—not from the round's headline.
$20bn reported executed. The remaining $10bn is not rendered as cash already paid.
NVIDIA was named for $30bn in February. Its later aggregate portfolio disclosures and management discussion do not, in the passages inspected, settle that particular payment. Unresolved does not mean unpaid. Nor should its older letter of intent for up to $100bn be added automatically to the later round. The documents describe different stages; reports of a replacement are not a definitive cancellation agreement. NVIDIA investment disclosure; earlier letter of intent.
Follow one relationship without inventing a cash loop
Amazon supplies capital to Anthropic. Anthropic commits to buy computing from AWS. The relationship is real—but the documents contain several different financial things, not one interchangeable pile of money. Amazon's June-quarter investment note is more precise than the initial announcement. It reports two separate $5bn subscriptions, one in Series G and one in Series H, and a facility whose size and availability depend on conditions. Amazon, Note 2.
$20bn − $5bn = $15bn. The remainder is a ceiling, not a cash balance.
Amazon's disclosed Anthropic facility, US$ billions. The first and last bars start at zero; the floating middle bar removes part of the original ceiling. All three use the same scale.
Original facility ceiling
Series H funding
Remaining conditional ceiling
At inception, the amount available was zero. Later availability and drawings were not disclosed in the inspected passages.
Reported invested
Amazon → AnthropicTwo equity subscriptions reported in the second quarter: Series G and Series H. These are investor-reported transactions, not an estimate based on the size of Anthropic's round.
Conditional funding
Amazon → AnthropicFunding depends on delivery milestones and other terms. The facility can involve convertible notes or, after a liquidity event, common shares subject to an ownership limit. A maximum is not a drawing.
Future computing purchases
Anthropic → AWSA purchasing commitment announced in April—not ten years of revenue already earned, cash collected or proof that particular investment proceeds paid AWS invoices.
Sources: Amazon's filing and Anthropic's April announcement. These panels label legal/economic directions; their sizes and arrows do not encode conserved money.
Anthropic's May announcement says its $65bn Series H includes $15bn of previously committed hyperscaler capital, including Amazon's $5bn. Adding the old commitments to the whole round would count some capital twice. We do not allocate the remaining included $10bn to a named company without evidence. Anthropic's Series H statement.
The missing bridge is specific: we have not traced these investment proceeds to invoices, collections and subsequent spending, or separated the ultimate customers' retained payments from financing-supported demand. The absence of that trace prevents a strong claim in either direction. It neither proves that the money “just goes in circles” nor proves that every purchase is independently self-financing.
An ownership value is not a cheque. Revenue is not a cash receipt.
Microsoft's June-year-end investment note reports $11.9bn funded against a $13bn OpenAI investment commitment. It also reports $24.1bn of revenue from its commercial relationship with OpenAI during the fiscal year and $6bn of related receivables at 30 June. Those are different measurements: cumulative funding, a year's recognized revenue and an unpaid balance at one date. Subtracting $6bn from $24.1bn would not reconstruct collections without the opening balance and other movements. Microsoft annual filing, printed page 60.
The $135bn value Microsoft placed on its stake in October 2025 was not another $135bn investment payment. The roughly 27% ownership described then and roughly 25% at June year-end also have date and basis qualifications; they are not enough to calculate a precise dilution loss. October recapitalization statement.
Commercial terms changed too. The April 2026 amendment allows OpenAI to serve products across clouds while keeping Microsoft as its primary cloud partner under the stated terms. Microsoft's intellectual-property license becomes nonexclusive through 2032; Microsoft stops paying OpenAI a revenue share, while OpenAI's payments to Microsoft continue at the same percentage through 2030, subject to a cap. The percentage and cap are not disclosed. This loosens one constraint, but it does not establish that the earlier $250bn Azure purchasing commitment vanished. Microsoft's amendment account; OpenAI's matching account.
What does NVIDIA's $105bn guarantee actually mean?
In August, NVIDIA disclosed residual-value guarantees associated with the initial 4.25 gigawatts of critical computing load at SB Energy's PORTS-Pike project, with an OpenAI affiliate as tenant. A guarantee transfers risk under a contract. It is not a $105bn investment cheque, current expenditure or expected loss. The disclosure expects phased effectiveness, with the first phases in 2028. The subsequent filed forms reveal why a single payment arrow would be misleading. NVIDIA's August disclosure.
There are contractual gates between the headline and a covered-loss payment
Stated aggregate cap for the initially contemplated guarantee arrangement.
Not a forecast of what NVIDIA will lose.
- Delivery and commencementThe covered-loss provisions depend on lease commencement and ready-for-service conditions. The lease separately abates each phase's rent until delivery.
- A qualifying defaultMonetary or insolvency events, notices and cure/verification requirements matter. Permitted rent relief is not automatically a payment default.
- Remedies and recoveryThe guarantor can assume, relet, sell, reject or defer, under specified terms. Interim payments, other remedies and recoveries affect the result.
The lease form makes an economically important distinction. Delivery or insufficient-power conditions can allow specified rent relief; a weak market is not a general escape clause. Its post-commencement force-majeure definition excludes financial inability and capital-market volatility or economic downturn. That does not settle how an executed contract would be enforced, but it narrows the claim that the customer can simply walk away when demand disappoints. Lease sections 18 and 34.
There is a favorable commercial interpretation: support for infrastructure can bring forward usable supply, secure a market for NVIDIA-based computing and allow equipment or capacity to be redeployed. The form also specifies long-term NVIDIA reference architecture. The adverse interpretation is that a supplier has taken on exposure to a customer's infrastructure economics. Recovery rights are useful, but we have not established what reletting or collateral would fetch in a shared downturn. Guaranty remedies and architecture terms; the companies' commercial case.
Two $1.5bn contracts are not proof of a $3bn payment
SB Energy's September filing exhibits contain an NVIDIA prepaid-forward agreement and a separate IPO-linked share subscription, each for $1.5bn and both dated 17 August. The prepaid forward calls for payment within three business days; the subscription closes after a qualifying IPO. The IPO condition requires at least $1.5bn of new company-issued shares—sales by existing shareholders do not count toward that threshold. A liquidation preference in the forward is also junior to creditors and specified securities; its formula is not guaranteed recovery. Filed exhibit index; prepaid forward; IPO subscription.
The main prospectus exceeded the available retrieval limit, so its complete funding chronology was not inspected. The exhibits establish contractual obligations and conditions, not independently verified receipts. We therefore keep the contracts separate and do not report $3bn as cash paid. The public filing and a proposed IPO are not themselves a completed listing.
What this says—and does not say—about an AI bubble
The strongest favorable reading
Documented subscriptions are stronger than promotional promises. Financing and cloud partnerships can fund real capacity, reduce uncertainty and support useful products. Contractual conditions and recovery rights can limit exposure. The amended cloud terms also show that these arrangements need not stay fixed.
The strongest adverse reading
Funding and procurement can depend on the same small set of firms. Delivery gates, concentrated customers, borrowing needs and guarantees can transmit a disappointment. A supplier's current revenue need not imply adequate returns for the buyer—or for whoever ultimately carries the risk.
These are interpretations to test, not verdicts established by the diagram. The FTC's earlier compulsory-information study documents investment and cloud-purchasing links, but it covers an earlier period. Commissioners separately disputed speculative harm inferences; the vote to issue the report should not be presented as unanimous agreement with every conclusion. The September BIS speech likewise discusses both AI's benefits and financial-stability risks, not an inevitable collapse. FTC study; Ferguson statement; Holyoak statement; BIS speech.
The next discriminating evidence is retained customer cash, actual facility drawings, delivered and paid-for capacity, and returns through replacement and financing costs. Strong results on those measures could justify expansion. Weak results combined with binding obligations could expose local excess or financing fragility. Neither an aggregate fundraising headline nor a circle of arrows supplies those answers.
Inspect the arrangements and the limits
Source-linked arrangement ledger and chart data (JSON) · Arrangement summary (CSV). Missing paid amounts are null or empty, not zero. The files distinguish commitments, subscriptions, debt, purchasing and guarantees; they do not offer an additive “AI money” total.
Evidence is mainly issuer reporting and public contract forms, not an audit of private books. Named legal entities are used only where the document supplies them. In particular, an anonymous Alphabet funding commitment is not silently assigned to Anthropic, and an aggregate NVIDIA investment number is not assigned to OpenAI. The native-language check here is one Japanese SoftBank release, not broad non-US coverage. No security valuation, loss probability, complete borrower underwriting or whole-economy demand census is claimed.
Previous: spending, leases and usable capacity · All dated research · The answer so far
This is research on an open question, not a prediction of a crash or a personal investment recommendation. Source publication, transaction dates and the date of our inspection are different.