Nvidia’s GPU financing
Neoclouds that cannot borrow on their own name need someone else’s credit. Nvidia guarantees the rent so banks will fund the chips — and SemiAnalysis models an ~18% cut above the floor.
For years the hyperscalers paid for AI factories from cash flow — Google, Amazon, Meta, Microsoft, Oracle. That broke over the last twelve months. Oracle borrowed. Then Meta. Then Google. SemiAnalysis, the semiconductor research shop tracking the buildout, now says financing — not chips, not concrete — is the binding constraint.
Nvidia’s answer is a six-year take-or-pay backstop: a minimum-revenue guarantee on GPU rental hours. Lenders underwrite Nvidia’s AA / Aa2 rating instead of the neocloud’s. The borrower still has to assemble what SemiAnalysis calls the AI Project Trinity — capital, offtake, datacenter. Named deals so far are SharonAI in Australia and Firmus in Batam, Indonesia. Both sit in Asia Pacific. Neither has completed a full six-year cycle.
Josh Brown, on The Compound, described the circularity in plain language. A neocloud announces a data center. Private equity and sovereign money show up. Jensen Huang calls and offers to sell the chips — and to backstop the purchase. “Take that example. Times a thousand,” Brown said. “When the music stops, all these people go for the chairs.”
How Nvidia was becoming the Central Bank of AI. A central bank exists to supply liquidity when others in the banking system are unwilling to step in.
SemiAnalysis, July 2026
Four hundred basis points
The spread collapse is measurable. CoreWeave’s five-year unsecured bonds sat around 10%. The fixed-rate tranche of its $8.5 billion delayed-draw term loan — backstopped by Meta, not Nvidia — cleared at 5.9%. Meta’s own five-year paper was about 5.0%. That is roughly 90 basis points wider than Meta, and about 400 basis points tighter than CoreWeave standing alone.
Five-year funding cost
The 5.9% print is a Meta offtake deal, not an Nvidia backstop — same credit substitution, different guarantor. Numbers from the stock-market wiki’s backstop chain.
As of July 2026 no US or European Nvidia-backstopped financing had been announced. CoreWeave’s flagship loan used an associated customer contract. In America a hyperscaler’s offtake already does what Nvidia’s program does in markets that lack those counterparties. The backstop may be additive to Asia-Pacific unit growth rather than a global rewrite of who buys GPUs.
SemiAnalysis still projects AI debt as a multi-trillion-dollar credit market — more than $7 trillion outstanding by 2029, second only to US residential mortgages at just over $13 trillion. Neoclouds have already used H100s as collateral. The backstop is what lets a lender treat the stack as investment-grade paper instead of project risk. On revenue above the guarantee, SemiAnalysis models an average take of about 18%. Nvidia declined to offer further details. The figure has not appeared in a filing or a call.
Ben Thompson, who does not have a GPU book, asked the same constraint in different words. Everyone talks about compute and electricity. “Maybe the nearest term questions are we going to have enough money?” The stack he names matches the wiki: cash flow, then debt — blown through in about a year — then Google equity, then Nvidia tapping pension funds and insurance floats. “What’s after that?”
A credit market the size of mortgages
SemiAnalysis projection. Not an Nvidia figure.
The foundry said no
TSMC is the other chokepoint supplier — the foundry that actually fabricates the chips Nvidia sells. Asked whether it would finance customers’ customers the way Nvidia does, C.C. Wei said no. TSMC does not do this kind of financial arrangement. The ordinary model, he said, works “smoothly and also successfully.”
That contrast isolates vendor financing as a choice specific to Nvidia’s situation, not an industry-wide response to scarcity. TSMC sells to well-capitalized fabless incumbents. Nvidia sells to newly formed neoclouds. Wei’s “so far” leaves the door open. It does not make vendor financing the default.
At the backstop price itself, SemiAnalysis says a typical neocloud earns a zero or slightly negative project IRR. The equity is a call option on rental rates staying above the floor. If they do not, Nvidia’s revenue share becomes a purchase obligation. AMD already offered similar backstops last year to AWS, Oracle Cloud, and a string of neoclouds — so the structure is not sui generis. What is new is the scale of the name doing it, and the ambition SemiAnalysis ascribes to it: broaden the GPU buyer base beyond the handful of hyperscalers building custom silicon to compete with Nvidia.
Two days later the All-In hosts recapped the same print on an incomplete clip. Chamath Palihapitiya recalled about $24 billion of hyperscaler revenue in the quarter and a matching neocloud book. That is a host’s memory, not a segment table in the 10-Q, and it does not confirm a new backstop. The filed numbers stay $96.2 billion of revenue and $89.0 billion of data center. If a matching neocloud book is real, it is the buyer base SemiAnalysis said the backstop was meant to create. The wiki does not treat the recollection as proof.
Security is now an underwriting filter on the same spine, not a second chain. Dylan Patel’s ClusterMAX 3.0 pass found what he called security horror stories. Frontier labs take bare metal, insist on zero-trust, and generally give the operator read-only access. CISOs sit at the table. Only Together, among the neoclouds he named, runs a paid bug bounty. The September 23 refresh of the same ranking said GPU supply “has gone to zero.” The shop tested 77 managed GPU-cloud providers against a market view of 323. Nebius joined CoreWeave in Platinum — CoreWeave still “sets the technical bar,” Nebius “commands a premium pricing.” Google Cloud sits with Oracle in Gold; Azure moved to Silver; Crusoe dropped to Bronze. Only 19 neoclouds globally made a Medallion grade. Testing now wants Blackwell, not Hopper. The financing line is the Matthew principle: profitable frontier labs lock capacity; smaller ones get fat prepays, worse prices, and a harder time planning years ahead. That is who-can-host and who-can-finance color on GPU-collateral financeability. It does not name a new United States backstop, and it does not flip a step.
Ed Zitron, on The Compound at the end of August, put numbers on concentration. Nvidia’s latest quarter was 16 percent from one customer; the first half of fiscal 2027 was 44 percent from three and 70 percent from five. OpenAI, in his telling, is about 60 percent of all AI data-center demand, and there is not another OpenAI-sized spender except through venture capital. Bear color on the same demand book the backstop is supposed to widen. It does not confirm the program, and it does not falsify it.
A September SemiAnalysis note, paywalled after the Hynix line, put a sovereign customer on the same spine. Korea’s plan is sized at $919 billion, 8.4 gigawatts by 2029 and 18.4 by 2035. SK Group asked for 5 gigawatts of Rubin and is down for 2; Naver is 200 megawatts. That is Nvidia-as-sovereign-customer color, not a new backstop and not a new chain. The same note’s claim that Hynix loses on 2027 HBM and SOCAMM pricing sits behind the Memory Model paywall. Leave it there.
The same shop went back to the second-quarter 10-Q and added up the off-balance-sheet book. Gross guarantees printed $530 billion, from $184 billion — a sum of separate tables, not a single Nvidia-labeled guarantee. Land, power, and shell — SB Energy’s PORTS-Pike, 4.25 gigawatts leased to OpenAI for twenty years — jumped from $3.5 billion to $108.5 billion. A new line: $36 billion of typically-six-year AI-cloud commitments. The filing does not say “take-or-pay.” It does not name SharonAI, Firmus, or GMI. A mid-September recheck still found no named US or European GPU take-or-pay partner. The 18 percent take is still a SemiAnalysis model under an assumed 40 percent above-floor share, not a number Nvidia filed. The supply-and-capacity commitments, $119 billion to $279 billion, were already in the August print. The $500 billion Apollo-to-KKR tap is still subject to definitive agreements. Do not treat the two five-hundred-billion-ish piles as the same money.