On Monday August 10, 2026, Nvidia signed a memorandum of understanding with six Wall Street firms — Apollo Global Management, BlackRock, Blackstone, Brookfield Asset Management, Goldman Sachs and KKR — to build an independent compute-financing platform aimed at mobilising more than $500 billion in third-party capital for AI infrastructure. The Decoder and Barchart both had the numbers within hours; BigGo Finance carried the participants and The State of AI carried the critique.

The number is not the story. The single sentence buried inside it is the story.

What was actually announced

Read the disclosure literally. Six firms signed letters of intent, not executed commitments. There are no individual firm allocations, no closing schedule, and no anchor customers named. The $500 billion figure is described in the joint materials as the amount the platforms are "designed to attract over time" — a target, not a commitment. Nvidia's own stock closed down between 1.4 and 2.9 percent on the day of the announcement, taking roughly $60–70 billion off the market cap in the same session that the company was telling Wall Street it had assembled six of the biggest allocators on the planet to underwrite its customers.

Jensen Huang's framing, verbatim across two write-ups: "Nvidia has reached a major milestone. We started by making chips; today, we are helping create a new type of infrastructure — AI factories — that is highly productive and investable." And, separately: "This is really the first time that technology chips have become an investable asset class."

Both quotes describe Nvidia's product. Neither describes the financing arrangement. The financing arrangement is somewhere else in the release.

The sentence everyone skipped was 25 percent

Here it is. Nvidia will provide residual-value support of up to 25 percent per project, reviewed case by case, on the compute the platforms finance. That is a captive-finance guarantee. If, when the loan matures, the GPUs a hyperscaler-adjacent customer bought with borrowed money are worth less than the debt outstanding, Nvidia — the manufacturer that sold the chips into that loan in the first place — makes the shortfall good up to a quarter of the residual value.

Huang describes this as "supplementary in nature" and "far lower than other compute financing arrangements in the market." That framing is doing an enormous amount of work. The comparison class Huang is invoking is Nvidia's own equity investments in customers — CoreWeave, Nebius, Lambda — where Nvidia's exposure can run above 25 percent of the whole deal. A 25 percent residual-value backstop on debt capital raised from Apollo and Blackstone is not the same instrument, and "far lower" is a phrase you use when you are hoping the reader does not stop to reprice the risk.

Three concerns collapse out of the mechanism when you spend a paragraph on it. First, the underwriting is not arm's-length: the vendor of the collateral is guaranteeing the collateral's floor, which is the definition of a captive-finance structure regardless of what the marketing calls it. Second, the residual-value assumption depends on a redeployment market for used GPUs that has, to date, only ever seen a demand upswing — the A100 example Huang keeps citing covers a scarcity period, not a downturn. Third, the $500 billion is soft; the 25 percent is contractual. The number in the headline is the number the platforms would like to raise. The number in the fine print is the number Nvidia is on the hook for.

Fink's mortgage-backed-securities comparison is the tell

The load-bearing quote of the whole announcement did not come from Nvidia. It came from BlackRock. Larry Fink — chief executive of the world's largest asset manager — called the initiative "the next future for financial engineering" and drew an explicit comparison to the emergence of mortgage-backed securities in the 1970s.

That is a sentence he chose to say on the record. It is also, on any fair reading, the single most alarming sentence a BlackRock CEO could have picked to describe an AI-infrastructure financing platform in 2026. Mortgage-backed securities in the 1970s were the trade that let banks originate more housing loans by turning individual mortgages into a fungible, tradable asset class. Mortgage-backed securities in 2008 were the trade that took down Bear Stearns, Lehman Brothers, and — as it happens — cost Larry Fink's shop no small quantity of overnight sleep.

David Solomon, Goldman Sachs CEO, delivered the operational version of the same idea: "We are at a critical juncture in a historic AI investment cycle, and we are very excited about the new opportunity to create a credit market supported by Nvidia compute power." Read Solomon and Fink in the same paragraph and the shape is clear. The six firms are not underwriting Nvidia. They are building the securitisation stack on top of Nvidia, and Nvidia is providing the credit-enhancement layer — the 25 percent residual-value guarantee — that makes the stack investable to the pension funds and insurance balance sheets whose money the platforms exist to attract.

Fink's comparison is either a genuinely candid statement of what the mechanism is, or the most successful attempt at plausible-deniability irony in recent finance PR. Either reading is bad.

Burry versus the decade

The subtext of the entire announcement is a running argument with Michael Burry, who has spent the last quarter publicly warning that hyperscaler GPU depreciation schedules are wrong. Burry's specific claim, on the record: the companies Nvidia finances or invests in "turn around and spend the money on Nvidia hardware," creating a circle that inflates both parties' revenue without underlying end-demand. That is the argument Nvidia's residual-value guarantee has to defend against, and Huang's counter is the useful-life claim: the A100 first shipped in 2020, remains "in commercial use six years later," and has an economic lifespan "stretching toward a decade."

The tell that the argument is live is in the hyperscalers' own books. Amazon shortened its assumed GPU useful life from six years to five in a recent filing. That is a one-year cut on a fleet the size of AWS's, and it is the sound of a CFO conceding a percentage point to Burry without saying his name. If the useful-life assumption is wrong by even 12 months on a $500 billion financing stack, the residual-value math the 25 percent guarantee is meant to backstop moves in the wrong direction for everyone. Nvidia's guarantee absorbs the first slice. The remaining slices are the pension funds and insurance balance sheets Solomon and Fink are pitching.

Nvidia's stock reaction — down on a day that should have been up, if the market believed the announcement was clean — is the second tell. The tape is quietly telling you the mechanism is not free money, no matter what BlackRock's press desk called it.

What to watch

1. Whether the first executed financing lands with the residual-value clause disclosed. The $500 billion is a letter of intent. The 25 percent is contractual — but only if it appears in the executed loan documents, and only at whatever terms and haircut the individual project underwriters actually agree on. If the first closing lands at 25 percent and the SEC filing carries the language, the structure is real. If the first closing lands quietly at a lower figure or is buried in supplemental collateral, Nvidia has walked back its own headline mechanism before the ink was dry.
2. Whether Amazon's five-year useful-life number gets matched by Meta, Microsoft or Google. Amazon moved first. If a second hyperscaler shortens its GPU useful-life assumption on its next quarterly filing, the industry-wide depreciation schedule is being repriced in public, and the residual-value guarantee moves from a marketing footnote to a load-bearing part of Nvidia's earnings quality. Watch for Meta's next 10-Q.
3. Whether the fund-raise gets pension money or just insurance money. Insurance balance sheets already carry long-dated illiquid credit; they can absorb structured GPU debt without the political headline risk. Public-sector pension funds cannot. The composition of the eventual $500 billion — how much comes from CalPERS-class LPs versus how much comes from Apollo's captive insurance side — is the honest measure of how far the "asset class" framing has actually travelled beyond a BlackRock press release.
4. Whether Michael Burry adds to his short. Burry is the on-the-record short-seller whose thesis this announcement is explicitly designed to counter. If his next 13F filing adds to the Nvidia short after Aug 10 rather than covering it, the residual-value guarantee did not close the trade — it opened a new one, and the market has a named counterparty betting against the mechanism at whatever gross size that filing discloses.

The August 10 announcement is what happens when the CEO of the world's largest asset manager decides to compare a vendor-guaranteed compute-financing platform to the emergence of mortgage-backed securities, on stage, in front of five other Wall Street CEOs who did not correct him. Nvidia will backstop 25 percent of the residual value of its own chips. The other 75 percent is what Apollo, Blackstone, KKR, Brookfield, BlackRock and Goldman Sachs are now in the business of selling to somebody else. The A100 has been in commercial use for six years. Whether that history holds for the H100, the B200, and whatever ships in 2028 is a question the residual-value clause exists to make somebody other than Nvidia's shareholders pay for getting wrong.

The Charging Bull statue on Wall Street was installed in 1989, one year after Michael Milken's junk-bond desk peaked at Drexel Burnham Lambert. Financial engineering has a way of leaving statues.