Nvidia Secures $500 Billion Pool with Apollo, BlackRock, Blackstone, Brookfield, Goldman, and KKR
The manufacturer signed MOUs with six Wall Street firms to mobilize third-party capital that finances GPUs instead of drawing from clients' balance sheets. Jensen Huang says Nvidia can provide a $125 billion backstop.
Jensen Huang arrived at the series of announcements between Monday and Tuesday with a phrase that summarizes the shift in thesis: AI chips are an investable asset. This is not just marketing. During the same window, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create financing platforms capable of mobilizing over $500 billion in third-party capital over the next few years, aimed at purchasing GPUs and building data centers. In a press conference on Tuesday, the CEO added the missing detail: Nvidia has the option to backstop up to $125 billion, or 25% of potential deals, if a specific vehicle does not close.
The structure matters more than the number. So far, those looking to purchase large-scale computing of Nvidia's latest generations had three pathways: hyperscalers financed with their own cash, amortizing over the years; neoclouds took on expensive debt or venture capital; and smaller companies rented third-party capacity. The current proposal is to create vehicles that buy GPUs and data centers with long-term institutional money, leasing them to AI operators without those operators needing to touch their own balance sheets.
Who's Involved and What Each Side Gains
The six chosen partners collectively manage over $15 trillion in assets. Each is expected to establish its own pool. For the managers, the appeal lies in guaranteed infrastructure returns from GPU leasing, with timelines compatible with infrastructure funds, insurance companies, and pension funds. For Nvidia, it alleviates the financial demand bottleneck that started to limit who could purchase its chips. For the end customer, it shifts the decision from capex to opex, widening the denominator of who can join the race.
Huang told CNBC that he only approached these six firms and that none declined. "We started by building chips; today, we are helping to create a new class of productive and investable infrastructure: AI factories," declared the CEO. "In AI, computing is revenue. Nvidia's computing is uniquely suitable for this role." In the Tuesday press conference where he confirmed the backstop, Huang emphasized: "We are not financing ourselves. Independent capital providers are assuming real credit risk."
The Blank Check Comes With a Stamp
Skepticism has a destination. Credit analysts have publicly noted that much of Nvidia's revenue in recent quarters depends on buyers who do not yet generate positive cash flow, which brings the design closer to indirect vendor financing. In this format, the residual credit risk remains somewhere in the chain, and a pullback in the price of inference per token—which has been significantly declining year over year according to research by Andreessen Horowitz on major open and closed models—tightens the economics of rentals.
On the positive side, the design mitigates part of the concentration risk: Meta, Microsoft, and Google are expected to spend over $340 billion in capex in 2026, primarily related to AI, which acts as a lifeline supporting Nvidia even if the second tier of demand slows down. What the new platform finances is precisely this second tier—frontier labs outside of hyperscalers and regional clouds that currently compete for GPUs under worse conditions.
How the Movement is Interpreted in Other Markets
In Europe, the immediate effect is on operators like the French Orange and the German SAP, which announced in recent months the construction of sovereign clouds for generative AI and face the same prolonged capex equation. A European pool with conditions comparable to those of these six firms becomes the next natural regulatory demand in Brussels, especially after the transparency phase of the AI Act came into effect on August 2.
In India, the Reliance Group and Yotta announced frontier data centers in Jamnagar and Noida with hundreds of thousands of Nvidia GPUs. Access to a financing pool with costs comparable to the U.S. alters the timeline of these constructions and the price of inference offered to Indian software exporters, who currently compete on margins against American providers.
In Brazil, the reading is more operational than strategic. Banks like Itaú and Bradesco have been announcing their own data centers with capacity dedicated to AI workloads and may, in the medium term, become clients of such structures rather than operators, a path that reduces exposure to currency in the capex of imported chips.