Nvidia found $500 billion of AI capex that does not touch its balance sheet

Nvidia said this week that it is forming financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, aiming to mobilize more than $500 billion for AI infrastructure, with the money coming largely from third-party investors rather than Nvidia's own balance sheet ^{[1] [2]}. The consequence is not the headline number; it is that the party most exposed to a glut of GPUs would no longer be the party that makes them.

The six counterparties are the tell. These are not strategic chip buyers or hyperscaler treasuries; they are the intermediaries through which institutional money, including retirement money, reaches private assets. Fortune put the implication in its headline: "Nvidia wants your pension fund in the AI trade" ^[7]. That is a framing, not a disclosure, and the account as reported names no pension plan, no fund vehicle, no closing date and no fee or leverage terms ^[8]. Operators reading this as a done deal should note the verb: the platforms aim to mobilize the capital ^[1].

The pitch is a reclassification. Jensen Huang described the arrangement as treating AI compute like "productive infrastructure" ^[3], which is a request to underwrite accelerators the way an allocator underwrites a toll road or a substation: long-lived, contracted, and financeable at infrastructure cost of capital. Morgan Stanley and Bank of America have both called Nvidia's chips unusually financeable ^[5]. Ben Thompson of Stratechery was less convinced, calling it "a completely new nerve-racking thing to bring safety-seeking assets to bear" on a pile of GPUs ^[6]. Both positions can be right. A chip is financeable when there is a deep resale market and a stable depreciation curve; the question an infrastructure committee has to answer is whether an accelerator four years from now looks like a turbine or like a 2019 server.

Which brings up the sweetener. Nvidia is offering residual-value support of up to 25% on some deals ^[4]. Read plainly, that means Nvidia is not transferring all of the downside it is moving off its balance sheet; it is keeping a slice of it in contingent form, and using that slice to lower the return third parties need. If that 25% applied to the entire $500 billion, the notional ceiling would be about $125 billion, though the source says the support attaches only to some deals, so the real figure is lower and unquantified ^[9]. Nobody outside the deal rooms currently knows the actual number, and that gap is the thing worth tracking.

Structurally, this is vendor support recast as asset management. It converts a capex cycle that would otherwise be constrained by customers' balance sheets into a funding problem solved by allocators, and it puts Nvidia in the position of both selling the asset and partially guaranteeing its terminal value. When demand holds, that is efficient. When it does not, the residual-value support and the resale market get tested at the same time, in the same direction, and the holders are institutions whose mandates describe them as safety-seeking ^[6].