NVDA is being positioned as the center of a new GPU-finance market: a $500B MOU with banks and asset managers, plus direct backstops to frontier labs, could turn Jensen Huang into the financier of AI buildout as much as the supplier of it. The bullish case rests on scarce GPU supply, rising rent rates, long-lived hardware use, and a 2027-2029 capacity crunch that keeps assets valuable and refinanceable. The bear case is that demand wavers, hyperscalers’ returns on capex weaken, or NVDA’s faster release cycle crushes resale values and blows up the debt-backed structures. A second thread says Elon’s AI buildout, SpaceX’s dense data center, and Vera Rubin scale-up all reinforce that NVDA remains the bottleneck and beneficiary.
Larry Fink compared the $500B deal to early-2000s mortgage-backed debt securities, but the pitch is a formal MOU: non-binding, not a guarantee that $500B will flow.
GPU rental prices are up 40% on the year while there still are not enough GPUs to back deals MSFT, GOOGL, ANTHROPIC, and OPENAI are signing with NVDA.
NVDA stock dropped 4% on the news, but the move is framed as too early to judge because the structure is still just an intention/alignment among participants.
Jensen Huang personally called the listed banks and none said no; the assembled group is APO, BLK, BX, BBU, GS, and KKR, with Apollo at $1T AUM, BX above $1.3T, Brookfield over $1T, and $3.4T combined.
The banks’ concern is that NVDA could ship a GPU that is 'a thousand times better' and crush margins, so the package includes up to 25% NVDA backstop — capped at $125B — described as plunge protection/depreciation insurance.
The financing is pitched as long-term debt and asset-backed structures that let even non-hyperscaler renters get GPUs and pay interest, with lenders able to claw back GPUs if needed.
The assets are argued to be transferable and valuable because a GPU can be unplugged, moved, or the data-center control transferred; in the worst case NVDA can claw them back, fold them into its own Neo Cloud, and redeploy them.
NVDA is already backstopping frontier labs by fronting GPU purchases for revenue splits; SSI is cited with a claimed 10% revenue split tied to a breakthrough launch, and CRWV is expected to see similar treatment after reporting a 464% YoY revenue increase.
David Solomon said $500B sounds large, but there are $9T in US money market funds and more than $100T in US equities; Larry Fink called the setup a very attractive, long-dated return opportunity.
The bull comparison is aircraft finance, not mortgage finance: GPUs are standardized, liquid, transferable assets like a 2005 plane versus a 2025 plane, while 2008 MBS blew up on oversupply and permanently rising-price assumptions.
Supply is still constrained by GPU substrate layers, packaging, lithography-machine counts, and data-center power, while memory capacity rises about 20% a year against 45% annual demand growth, implying constraints until at least 2028-2029.
GOOGL backlog is said to have doubled this year and hit $460B at the halfway mark, with more upside expected; CRWV also signed an A100 contract running to 2029, showing older NVDA hardware still clearing.
The A100, launched in 2020, was initially expected to last 3.5 years, but CRWV now thinks it is good through 2029 because inference and other workloads extend useful life beyond bleeding-edge training.
The next bottleneck is shifting from chips to energy and power by the end of the decade, with sold-out supply already visible for at least 2027 and likely 2028, and visibility for the next 18-24 months from lithography, packaging, and power buildout.
The downside list is clear: if AI demand plateaus or quarterly revenue growth falls from the current 100%-500% YoY range, GPU demand can drop and the $500B debt-backed structure can default; if NVDA’s release cadence floods the market, depreciation accelerates; if hyperscaler ROIC weakens, capex cuts follow.