MacroView · Learn · How the AI build-out is financed (the 2026 credit circle)

How the AI build-out is financed (the 2026 credit circle)

Last updated: 2026-09-07

On 10 August 2026 Nvidia announced financing platforms with six financial firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — to mobilise over $500 billion of third-party capital for its own customers: frontier AI labs, enterprises and AI clouds. Nvidia's stated reason is that those customers have been struggling to secure financing on ordinary terms, so the capital is offered at “attractive rates”. NVDA fell about 2% the day the arrangement was reported.

Who is in the loop
RoleWhoWhat they do
The capitalPension funds, insurers, sovereign fundsCommits to the arrangers' private-credit vehicles. Ends up holding the duration risk on chips financed like long-lived infrastructure.
The arrangersApollo · BlackRock · Blackstone · Brookfield · Goldman Sachs · KKRRaise that capital and lend it out. Book origination and management economics up front, whatever the financed assets do afterwards.
The borrowersFrontier AI labs, enterprises and AI clouds — OpenAI, CoreWeave and xAI named as the cohortBorrow to buy compute they could not finance on ordinary terms.
The vendorNvidiaSells the compute — and introduces its own customers to the lenders financing it. That introduction is the leg that closes the circle.
Why it is called a circle

Follow one dollar:

  1. A pension fund commits to a private-credit vehicle run by one of the six arrangers.
  2. The vehicle lends to an AI compute buyer, on terms that buyer could not get from a bank.
  3. The buyer spends the loan on Nvidia compute.
  4. Nvidia books the revenue now. The arranger books its fees now.
  5. The loan is still outstanding, secured against hardware, and the fund is the one holding it.

Nothing here is hidden or improper — it is disclosed vendor-adjacent financing. But the seller of the product is standing next to the lender financing its purchase, which means demand and credit are no longer independent of each other. Sales growth financed this way tells you less about end demand than the same growth paid for out of a customer's cash flow.

Where the risk actually lands
  • A duration mismatch. The arrangers earn up front, while the chips being financed depreciate on a far shorter schedule than the infrastructure debt funding them. Whoever holds the paper at the end holds that gap — and that is the LPs: pensions, insurers and sovereign funds.
  • Collateral that is a depreciating asset. Infrastructure debt is normally secured against something with a twenty-year life. GPUs are not that.
  • No per-firm disclosure. The $500B is an announced aggregate across six memoranda of understanding. No individual commitment was disclosed, so dividing it by six — an obvious first instinct — produces a number nobody published.
The hyperscalers are a separate leg

Microsoft, Alphabet, Meta, Amazon and Oracle buy the same chips, but they are not the cohort these platforms were built for: they fund capex mainly through their own bond issuance. That leg is measurable from SEC filings rather than announcements, and two numbers carry it:

  • Free cash flow — operating cash flow minus capex. Not a filed tag; a subtraction.
  • Net debt issued — long-term debt issued minus repaid.

Negative free cash flow beside rising net issuance is the cash-funded → debt-funded flip, as arithmetic rather than as commentary. It is the single most useful thing to track about the build-out, because it needs no interpretation: either the capex is coming out of the business or it is coming out of the bond market.

The precedent, stated carefully

Vendor financing has form. Lucent and Nortel lent to their own customers through the late 1990s, and their sales looked strongest in the quarters right before the bust — the financing kept revenue growing after end demand had stopped justifying it. That is a precedent worth knowing, not a forecast: the structure rhymes, the outcome is not implied, and the AI cohort's economics are not telecom's.

How to tell whether it is holding

The backdrop every one of these deals is placed into is the price of corporate credit risk. Two ICE BofA option-adjusted spread series cover it: investment grade (BAMLC0A0CM) and high yield (BAMLH0A0HYM2). Widening spreads make the next platform more expensive to fund and the next refinancing harder, which is the channel through which a credit story becomes an equity story.

Note what these are: index-wide spreads. They are not a read on any single borrower's default risk — that would need per-issuer CDS, which is proprietary data MacroView does not carry. The live filings and both spread series are available to agents on the get_ai_financing tool (/v1/ai-financing), which keeps the quarterly scoreboard for the six filers above.

Sources
Nvidia · the announcement itself
nvidianews.nvidia.com · AI compute infrastructure financing platforms

The primary source for the six counterparties, the “over $500 billion” aggregate, the “for NVIDIA customers” scope and the stated rationale. Read it before trusting any summary of it, this one included.

CNBC · contemporaneous coverage
cnbc.com · Nvidia, Wall Street asset managers and the $500B AI push

The market reaction on the day and Huang's own framing of the arrangement.

Desk commentary, not disclosure

One part of the picture above is not in any announcement: the LP → private-credit → borrower plumbing, and the fee-timing-versus-depreciation mismatch. That is how desks described the structure the week it landed. It is reported here as commentary rather than dressed up as a disclosed fact — the same distinction the dataset behind this page keeps.