.svg)
On 10 August, NVIDIA said it had signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to set up independent compute financing platforms, with the aim of mobilising more than USD 500 billion of third-party capital for AI infrastructure. The release is written for capital markets, so it reads as an announcement about a new asset class. The mechanic underneath it is ordinary. Somebody other than the operator pays for the hardware, the operator pays for access to it, and those payments repay the funder. That makes it a loan, and loan management software is what decides whether a book of them can be administered at this scale.
Jensen Huang's framing in the release was that in AI, compute is revenue. The financing follows from that claim. If a cluster of chips produces a metered, contracted income stream, it can carry debt the way a toll road or a container ship carries debt.
What lending against a machine looks like
Airlines have worked this way for fifty years. Very few carriers buy their fleets outright. A lessor or a syndicate of lenders puts up the money, keeps a claim on the airframe, and gets repaid out of what the aircraft earns flying routes. Mining equipment, rolling stock and shipping containers are funded on the same logic. What the six managers have agreed to do is apply it to racks of GPUs.
The first real test of that already happened. In August 2023, CoreWeave raised USD 2.3 billion in a debt facility led by Magnetar Capital and Blackstone, secured on NVIDIA H100 chips. Reuters reported the use of the chips as security as unusual at the time, and it was the first facility of that size backed by AI hardware. Nine months later the company signed a second one, USD 7.5 billion led by Blackstone with Magnetar as co-lead, secured on GPUs and customer contracts together, split into investment-grade and speculative-grade tranches at a variable rate averaging around 11%, with repayments starting in January 2026. By early 2026 CoreWeave's total debt had passed USD 21 billion.
The NVIDIA announcement takes that template and industrialises it. Six managers, dedicated pools of capital, and a stated intention to make the structure repeatable across frontier labs, enterprises and AI clouds instead of negotiated deal by deal.
Loan management software wasn't built for collateral that earns
Servicing a mortgage means running a fixed amortisation table against an asset that revalues slowly. A compute-backed facility asks for four things that sit outside that pattern.
The repayment stream is metered. NVIDIA's own description of the platforms refers to long-duration usage-linked revenue, which means the cash available to service the debt depends on how heavily the cluster gets rented in a given period. An amortisation schedule that moves with utilisation has to pull that data in, reconcile it against the operator's billing, and recompute.
The collateral revalues on a hardware cycle. A successor generation arrives every eighteen to twenty-four months and resets the resale value of the fleet securing the loan. Collateral tracking becomes a monthly figure with a depreciation curve attached, rather than a valuation filed once a year.
Contracts form part of the security. CoreWeave's 2024 facility pledged customer offtake agreements alongside the chips, which puts the status of third-party contracts inside the lender's collateral position.
The lender is rarely one party. Six platforms, each raising from their own investors, produces syndicated lending at volume: participation percentages, fee entitlements, notice distribution, amendment votes, and repayment allocation across participants who each need a position that agrees with everyone else's.
Any one of those is a data problem before it's a credit problem.
The market is administering this in spreadsheets
S&P Global launched two products in March 2026, DataXchange and AmendX, both aimed at loan agents. Its stated reason: lending and private credit markets have grown enormously while agents are still managing billion-dollar deals with Excel and email. S&P Global Market Intelligence had observed earlier in the year that private credit portfolios remain predominantly managed in spreadsheets, with rate resets, paydowns, fees and amendment data held in inconsistent formats.
The volume arriving doesn't wait for that to get fixed. JPMorgan projects annual data centre securitisation issuance of USD 30 billion to USD 40 billion across US CMBS and ABS in both 2026 and 2027, up from roughly USD 27 billion in 2025. Outstanding private credit to AI-related companies has gone from close to nothing to more than USD 200 billion, with over USD 40 billion originated in 2025 alone. One forecast puts global AI-related debt issuance near USD 570 billion for 2026, of which USD 236 billion had priced by 31 May, four times the previous year's pace. The NVIDIA platforms sit on top of all of it.
The institutions that end up holding this paper are banks, insurers, pension funds and the credit arms of the managers named in the release. Their question is narrow. Can the loan servicing software they run today produce a participant-level position on a facility whose repayment moves with GPU utilisation, whose collateral revalues quarterly, and whose amendment history lives in an inbox? Where the answer is a reconciliation team and a shared spreadsheet, the cost surfaces later as restatements, disputed fee calculations and audit findings.
Headcount absorbs some of this, and many institutions will hire. The work compounds anyway, because the fund manager's position has to agree with the agent's, the agent's has to agree with the operator's metering, and the auditor needs all three to agree as at a date in the past. Holding the whole facility in one system takes that reconciliation off the table, which is why loan management software has become a live procurement question for anyone planning to write compute-backed credit.
The partnerships are still subject to final agreements, which leaves a short window in which the systems that will administer them get chosen.
