Nvidia wants to mobilize more than $500 billion of third-party capital around hardware it calls an “investable asset class.” But a transferable GPU is not necessarily durable collateral when newer generations can be materially better.
Nvidia Wants Capital Markets to Turn Financing Into Chip Demand
Nvidia has signed memorandums of understanding with Apollo Global Management, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs, and KKR. The goal is to finance GPU purchases and data-center construction for hyperscalers, frontier AI labs, and enterprises.
The ambition is not the same as $500 billion of committed capital. The amount, deployment schedule, loan guarantees, and final terms remain unestablished.
The attraction is clear. Institutional investors provide the money, customers acquire infrastructure, and Nvidia converts access to capital into potential chip demand without necessarily carrying the financing on its own balance sheet.
A Useful GPU Is Not Automatically Good Collateral
Jensen Huang describes Nvidia hardware as productive, revenue-generating, long-lived, fungible, and transferable. Blackstone President Jon Gray compared compute financing with mortgages.
The comparison has limits. A GPU can remain technically useful and CUDA-compatible without retaining enough economic value to protect a lender. A materially better generation could impair its residual value.
Financing can expand access to compute, but it may also make reported demand partially dependent on capital arranged by the supplier. Credit and execution risk could move into banks, insurers, and private funds before the infrastructure has demonstrated its returns.
There is genuine underlying demand. AWS reportedly has much of its capacity committed through the end of 2027 and much of 2028, with customer contracts generally lasting at least five years. Amazon says servers and networking equipment take a little less than three years to break even and have useful lives of at least five to six years.
Microsoft reported commercial remaining performance obligations of $678 billion while Azure grew 43%. These commitments support continued construction. They cannot prove that every financed GPU cluster will earn an adequate return.
Memory Is Becoming the Constraint
Micron reportedly cannot meet even half of customer demand for data-center memory. Buyers are largely insensitive to price, and demand signals extend multiple years. Customers reportedly identify DRAM, rather than power, as their primary constraint.
Micron expects 2027 to be tighter than 2026 because demand is growing faster than supply. Its Strategic Customer Agreements cover three to five years, include binding annual “Take or PAY” commitments, and provide substantial upfront cash. Pricing either floats with the market or operates within floors and ceilings.
Those are not casual expressions of interest. They are multiyear commitments around a component customers believe could remain scarce.
The Binding Constraint May Capture the Value
Goldman Sachs estimates memory at approximately 62% of the bill of materials for Nvidia’s Vera Rubin superchip. Morgan Stanley analyst Howard Kao estimates memory cost inside a Vera Rubin rack will increase 435%, versus 57% for the GPU, contributing to a reported rack cost of $7.8 million.
Amazon increased planned 2026 cash capital expenditure from approximately $200 billion to approximately $220 billion because of higher memory costs. That suggests memory is no longer a peripheral constraint. It is large enough to change a hyperscaler’s spending plans.
The accelerator still matters, but the economics of the complete system are shifting. Multiyear agreements, price-insensitive demand, and memory’s rising share of cost suggest that more value could accrue to whoever supplies the binding constraint.
One estimate places Big Tech capital expenditure at 2.4% of US GDP in 2026 and 3.1% in 2027. At that scale, spending announcements become less informative on their own. Actual AI monetization and credit conditions matter more.
Contracted demand can sustain construction, and transferable GPUs can facilitate financing. Neither establishes residual value or adequate returns. The investment question is now who controls the scarce component, who funds the buildout, and who bears the risk if infrastructure returns disappoint.