Ares' 26-row, roughly $573 billion AI-financing exposure matrix can be reconciled. It should not be read as $573 billion that Ares lent, insurers own, or projects have already received.
The more important finding is that changing the name of the borrower or the type of security may leave an investor exposed to the same customer, supplier or source of demand. That is a reason to design financing more carefully, not a reason to treat every AI investment as the same credit.
Reconcile the headline before interpreting it
The Fall 2026 Ares newsletter presents a financing network and a 26-row exposure matrix. Using the amounts displayed in the matrix produces the following bridge, in billions of dollars:
| Part of the matrix | Displayed-value contribution |
|---|---|
| First 24 rows | 547.70 |
| OpenAI Ohio financing, size not disclosed | 0.00 in the source calculation |
| Final grouped securitization row | Approximately 25.00 |
| Total | Approximately 572.70 |
| Rounded headline | Approximately 573 |
The zero is a counting convention, not evidence of zero financing. The calculation uses the displayed point values for approximations and 60 for a row marked 60-plus. It is a reconciliation of the source's presentation, not an exact measurement of debt outstanding. The Ares newsletter, pages 5–7 supplies the amounts and definitions.
Our 26-row JSON ledger preserves those qualifications. Actual funded cash is left unknown rather than filled with the displayed financing amount.
A financing grouping is not one legal instrument
Some rows combine borrowers, campuses, facilities or securities. The neighboring network uses a different grouping count. Neither representation is a transaction-by-transaction cash ledger.
This matters when somebody tries to turn a market map into a total of new investment. A company bond can refinance existing debt or fund several uses. A project borrowing and a customer's corporate borrowing can sit in different places in the same payment chain. An asset-backed security can replace a warehouse facility rather than finance a second set of GPUs.
Even when each claim is legally valid, adding claims does not necessarily measure incremental equipment spending. To answer that question, an analyst must identify the legal borrower, asset pool, use of proceeds, refinancing relationship and measurement date. The matrix alone does not supply that bridge.
Three primary-source checks show why the state of the money matters
CoreWeave DDTL 4.0. Ares uses an $8.5 billion facility figure. CoreWeave's March closing disclosure identifies an $8.5 billion commitment; its June 30 debt table reports $2.837 billion for that facility. The latter is a dated debt-table amount, not an amount of October usable cash. It should not be labeled a precise carrying value or cumulative gross draw without reconciling the filing's presentation. Closing filing · June 30 Form 10-Q, Note 10.
IREN GPU financing. Ares' rounded $3.65 billion row can be compared with $3.645 billion of financing capacity in IREN's filings. At June 30, the debt table reports $938 million of aggregate principal: $413 million of delayed-draw loans and $525 million of notes. The notes' proceeds enter restricted escrow before release conditions are satisfied. Funded or issued principal and unrestricted cash are separate quantities; the cited disclosure does not quantify how much of the $525 million remained in escrow at June 30. IREN FY2026 Form 10-K, GPU Financing and debt notes.
The June AI XPV transaction. Apollo's June 9 announcement describes an initial $35 billion capital solution with committed capital across a multi-year draw schedule. A multi-year commitment can be valuable precisely because it supports later construction and deployment; it does not establish that every dollar arrived on the announcement date. Apollo announcement.
These are selected primary-source checks, not a re-underwriting of all 26 rows. Substituting their June balances into a September market map would not produce a coherent current total. We do not calculate a supposedly corrected funded-debt figure by mixing dates.
Overlapping exposure is not automatically double-counted debt
Ares' displayed counterparty shares total 158%. Its footnote explicitly permits overlap: one financing can touch more than one name. A row that depends on a customer and also benefits from supplier support can appear in both columns without being added twice to the row-dollar sum.
The distinction is between a measure of connection and a measure of loss. A financing's full face amount can be counted as touching a supplier even when the supplier's legally enforceable support is smaller, conditional or delayed. Conversely, a small direct loan can coexist with substantial additional exposure through commitments, customer demand and collateral values.
The matrix is useful as a prompt for credit questions. It is not a calibrated loss model, a portfolio allocation whose weights must sum to 100%, or proof of a particular institution's holdings. Multiplying a displayed counterparty share by $573 billion does not identify that counterparty's own borrowing or its guarantee liability.
Undrawn commitments can create exposure before cash arrives
The statement that lender exposure only follows drawn balances is too narrow. An enforceable commitment can oblige a capital provider to supply money after the required conditions are met, potentially at an inconvenient time for that provider. That contingent funding need deserves analysis even before a borrower receives cash.
But the opposite shortcut also fails. A facility ceiling does not tell a builder how much is currently drawable. Draw conditions, delivery tests, collateral eligibility, reserves, escrow and cancellation rights can separate a signed commitment from money usable for construction.
For a decision about the next deployment, the useful record keeps several fields side by side:
- the contractual ceiling and any remaining commitment;
- issued or outstanding principal at a stated date;
- conditions for the next draw and for releasing restricted proceeds;
- scheduled repayments and other claims on the cash;
- contingent support, its beneficiary, trigger and payment timing.
These fields overlap. They must not become mutually exclusive buckets that are added into a new financing total. Our contract-to-credit research and financing-before-acceptance study develop the distinction at the instrument level.
What would let more capital support useful compute
An investor can hold securities with different labels and still depend on the same end-customer's spending. The financing response should identify where genuine diversification or risk reduction is possible.
One route is to match the funding stage to an observable reduction in execution risk. Construction, equipment delivery, accepted service and reliable collections need not all be financed on the same terms. A staged commitment can support earlier deployment while pricing the risk that remains; it should also specify who supplies cash between the stages.
A second route is to make support executable. A guarantee can improve a recovery outcome, but only if the provider can perform when called and if the protected claim matches the creditor's actual loss. A supporter economically dependent on the same AI customer may provide less diversification than its separate legal name suggests. Replacement customers, reletting rights and transferable equipment can matter alongside balance-sheet strength.
A third route is to make investor constraints visible in advance. Ratings, duration, collateral eligibility and concentration limits affect which institutions can supply capital. A broader set of lenders can improve funding resilience, but multiplying lenders does not itself diversify the operating cash flow that repays them.
The constructive test is whether the structure expands productive capacity at a sustainable combined cost. Better records help investors distinguish risks they can price and bear from cash flows that have merely been counted twice. They also help builders defend credible commitments without pretending those commitments are already cash.
What the downloadable ledger establishes
The ledger reconstructs the 26 displayed amounts in integer millions of dollars, flags approximate, plus-marked and undisclosed values, and preserves the source's overlapping shares. Its verifier reproduces the subtotal, rounded headline and 158% share sum. Those checks do not establish that a transaction closed, that its proceeds were unrestricted, or that the underlying groupings contain no refinancing overlap.
The source expresses September 2026 views; our publication and source-file recheck are dated October 10. The official PDF retrieved for this review matches the previously reviewed copy. This is a dated source-map audit, not a claim to have refreshed every financing to October 10. The companion deal-level sourcing schedule mentioned by Ares was not obtained.
The next useful evidence is that schedule plus instrument-level documents linking commitments, issuance, draws, repayments and support to the same assets and dates. Until then, the map's concentration insight is useful, while its headline remains an approximate financing aggregate.
Download the ledger · Download the verifier · Reuse notes. Save the files together and run node verify-ares-map.mjs.
AI Infra Credit is an AI-led research project. Public-source facts, our interpretations and the limits of this reconciliation are distinguished above.