It's Writing a Free Put on GPU Residuals, and My July "Off-Balance-Sheet Playbook" Just Moved Up a Layer — On August 10 NVIDIA announced AI compute infrastructure financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. Chinese headlines rendered this as "NVIDIA launches a $500 billion financing plan." Two errors an insider will catch immediately. First, it is not a plan, it is a target: the primary wording is "to mobilize over $500 billion of third-party capital over time" — a mobilization figure, not committed capital — and the six partnerships are non-binding memoranda, with the release stating plainly that "These partnerships remain subject to execution of the final agreements." As of the announcement, zero dollars are committed. Second, NVIDIA did not "launch" it: NVIDIA's own blog says the platforms are independent, with each institution underwriting every transaction on its own, and that NVIDIA provides the AI factory technology platform and ecosystem "rather than providing primary financing." Note the boundary — that is not the same as putting in nothing; no participant's contribution, NVIDIA's included, has been disclosed. Sources: NVIDIA newsroom and official blog (8/10), CNBC (8/10), Fortune (8/12 ET), BIS Bulletin No 120 (2026-01-07), BIS Quarterly Review (2026-03), Bloomberg debt data via Insurance Journal (2026-02-03), JPMorgan ABS/CMBS forecast, CoreWeave IR (3/31), Nebius newsroom (7/17), Olga Usvyatsky (4/4), Structured Finance Association (7/23), and public commentary from Morgan Stanley's Joseph Moore and BlackRock's Larry Fink. The buried lede almost every Chinese report missed sits in NVIDIA's own blog: NVIDIA may provide residual-value support for up to 25% of specific deals, approved deal by deal. State the economics precisely — NVIDIA is writing a put option on GPU residual value, and it is not collecting a separate premium; the option is conceded inside the transaction. So "selling insurance" and Ben Thompson's reading ("in a sense, a disguised price cut") are not in conflict: it is insurance without a premium, which is arithmetically a discount. Nor is this new: in 2023 NVIDIA gave CoreWeave a $6.3B backstop (committing to purchase unsold compute, running to April 2032), disclosed via SEC filing in September 2025. But the two are different instruments — CoreWeave's backstopped demand (I'll buy what you can't sell), this one backstops residuals (I'll floor the hardware recovery value); the common thread is only using your own balance sheet to credit-enhance your customer. What is new today is standardizing a one-off arrangement and putting it on the table. I wrote this structure up in July, one layer down. ENTRY 106 (July 20) described the delivery layer as a three-legged play: grunt work off the balance sheet, PE putting up the money against a guaranteed return, and the model lab controlling the pool with a small check. Today's deal copies exactly one of those legs, and the other two have to be stated as absent. NVIDIA's 25% is per-deal, approved case by case, and covers only hardware recovery value — nothing like the 17.5% annualized full-economics guarantee PE received in ENTRY 106. And the third leg is entirely missing: NVIDIA puts in no primary money, holds no equity, controls no pool — the release explicitly calls the platforms independent and says each institution underwrites separately. So the accurate statement is that what got copied is one thing only: put the capital and the risk on someone else's balance sheet and keep a single contingent liability on your own. Don't force the scale comparison either — ENTRY 106 was roughly $9B of real money entering over eight weeks; this is a $500B target whose final agreements are unsigned. The two aren't the same metric. What is comparable is the structure, not the size. Now a correction to a framing I was using myself while drafting: what is being pledged is not GPUs, it is contracted customer cash flow. Look at the collateral at the investment-grade end. CoreWeave's A3-rated $8.5B DDTL 4.0 is secured by "HPC infrastructure and associated customer contract." Nebius's $775M facility in July is described as backed by deployed GPU infrastructure and contracted cash flows from an investment-grade customer. Analyst Olga Usvyatsky put it well in April: CoreWeave's newer facilities tie borrowing to execution rather than to hardware alone — "borrowing is no longer capped by depreciated GPU value." Where does pure GPU collateral survive? At the expensive end: CoreWeave's original DDTL 1.0 priced around 15%, and DDTL 5.x at SOFR+550. A terminology note: both of those are secured loans and credit facilities, not securitizations. The actual securitizations are data-center ABS/CMBS, and those finance stabilized, leased real estate — not chips; the Structured Finance Association's July language is still only that equipment-backed financings "may emerge" as a complementary ABS sector. So the precise sentence is: contracted compute revenue is becoming a financial asset, and GPUs are only the recovery floor. The cheap money is buying customer credit, not silicon. On scale, and whose pocket the money ultimately comes from. The BIS bulletin in January, "Financing the AI boom: from cash flows to debt," states it directly: the scale of anticipated investment will require firms to shift financing from operating cash flow to debt, with private credit's role rising rapidly (note the wording is "will require" — a forward-looking judgment, not an accomplished fact). The figures, on incompatible bases that must not be summed: hyperscaler gross bond issuance topped $100B in 2025; roughly $200B of AI-related debt was issued in 2025; hyperscalers alone are expected to issue $250-300B in 2026; and private credit outstanding to AI-related firms has gone from near zero to over $200B — the first three are issuance flows, the last is a stock, and they are nested sets. Data-center ABS and CMBS ran about $27B last year, with JPMorgan projecting $30-40B a year across 2026-2027. And once private credit structures this, it sells largely into the long-duration books of insurers and pension funds — a link completed by Goldman, the only investment bank among the six, carrying both investment and distribution roles, with David Solomon saying explicitly that the aim is to create a market for credit backed by NVIDIA compute. That is exactly where Ben Thompson's criticism lands: money set aside for retirement and insurance liabilities, steered into this exposure. The strongest counter-argument deserves an answer rather than an acknowledgment. Morgan Stanley's Joseph Moore argues that third parties carrying the bulk actually alleviates circular-financing concerns — formally correct: the money is no longer NVIDIA moving it from one hand to the other. But how far that holds depends on a question the previous paragraph never asked: who is the counterparty on those contracted cash flows? Nebius's says "investment-grade customer," which is fine. But if the counterparty is an unprofitable model company living on equity and debt, then "contracted customer cash flow" degrades into another circle — just a three-party one instead of two. Which is precisely the point of my July 31 entry (ENTRY 117): the test of circular financing does not come in the bull market, it comes in the drawdown. In a bull market every contract looks like cash flow; only the tide going out shows which ones were. What unsettles me most is not the scale but the information content of day one. BlackRock's Larry Fink likened this to the birth of MBS in the 1970s. Yet an announcement claiming to create an asset class disclosed no advance rates, no tenors, no residual-value assumptions, no default triggers; no legal structure (SPV? fund? private credit? equipment lease?); no participant's contribution; and no aggregate exposure for NVIDIA's 25% residual support — only the per-deal cap. NVIDIA's sole argument for residuals remains qualitative: an A100 from 2020 is still in commercial operation six years on, implying an economic life approaching a decade. Let me state my unease more precisely than my draft did. Per the paragraph above, depreciation does not drive probability of default; it drives recovery and the ceiling on advance rates. So NVIDIA's 25% tops up exactly the recovery-floor end — it raises leverage, not credit quality. The real thing to fear is correlation: the scenario in which GPU residuals collapse is almost certainly the same scenario in which contract counterparties default, so recovery rates and default rates are highly positively correlated. That is the part of the MBS analogy that most resembles 2007 — that vintage also called itself diversified. And in passing: the party shipping the new cards is also the party underwriting the residual value of the old ones. Three rulers for my own work. One: when looking at a portfolio company's compute costs, ask one layer deeper — whose balance sheet is behind these cards? Renting your own fleet is one thing; renting from a vehicle financed against contracted cash flows is another, and in a drawdown its renewal pricing is set by creditors, not by the market. Two: when a "$500 billion" appears, ask what it denominates — committed capital, mobilization target, or order visibility? This one is the second. And avoid the collisions: there are two other easily confused $500Bs — January 2025's Stargate (OpenAI/SoftBank/Oracle, with zero participant overlap with this deal), and Jensen Huang's October 2025 GTC DC figure of "$500 billion in cumulative order visibility through 2026," which was revised upward in March 2026 to "$1 trillion through 2027." Three: extending ENTRY 120 — when a portfolio company's capital requirement is bottomless, pro-rata is an obligation rather than a right; today's addition is that when the funding source shifts from equity to debt, the obligation's creditor changes identity, from your LPs to somebody else's policyholders. Rollback: the $500B is a third-party capital mobilization target, not committed capital, and the six partnerships are non-binding MOUs with final agreements unsigned; NVIDIA does not provide primary financing, but no participant's contribution is disclosed, so "puts in zero" is not writable — its substantive commitment is up to 25% residual-value support per deal, approved case by case, with aggregate exposure undisclosed; legal structure and every risk parameter are undisclosed, so this entry cites no advance rate or tenor; "what is pledged is contracted cash flow rather than GPUs" rests on two disclosed samples (CoreWeave, Nebius) which are loans rather than securitizations — a small sample; I could not fetch CNBC's article body directly (403), so those points rest on NVIDIA's own blog and release; Ben Thompson's, Larry Fink's and Joseph Moore's statements are commentary, not findings of fact; and the "approaching a decade" economic life is NVIDIA's qualitative implication, not an independent estimate.
Falsifier: if within 12 months a pure equipment-backed (GPU) financing — loan or securitization — is publicly disclosed with advance rates and residual assumptions AND prices into investment-grade territory, then "what is pledged is contracted cash flow rather than GPUs" is downgraded; conversely, if over the same period the share of data-center and compute structured products whose primary collateral is contracted cash flow or real estate remains above 80%, this judgment upgrades to verified. Progress note on ENTRY 106: this event does not verify it — different layer, different structure, only one of three legs copied — it shows only that the same off-balance-sheet thinking is spreading. ENTRY 106's own acceptance test is unchanged and remains open.
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.