Nvidia's $500B AI Compute Financing Plan: What Was Signed

What was signed

NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish independent compute financing platforms, with the stated aim of mobilizing more than $500 billion of third-party capital for AI infrastructure over time. The intent is dedicated pools of capital at scale and at attractive rates for NVIDIA customers, spanning frontier labs, enterprises and AI clouds. [1][3]

The instruments are memoranda of understanding. The release states in its own words that the partnerships remain subject to execution of the final agreements. No platform has been capitalized, no fund has closed and no borrower has been named. [1][3]

Atlas interpretation: A memorandum of understanding among seven institutions of this size is not nothing; it takes months of work and legal review to get six competing capital allocators to sign the same page. It is also not a commitment of capital, and the $500 billion is a target for what the platforms might eventually mobilize rather than a sum anyone has agreed to provide. Both readings should survive contact with the headline. [1]

The underwriting argument, in NVIDIA’s own words

Jensen Huang’s statement is a credit pitch rather than a technology one. NVIDIA compute, he argues, is "broadly adopted, flexible across models and workloads, fungible and transferable across customers and operators, and continuously improved through CUDA software, extending its useful life and improving its economics over time." The release adds that it provides the lowest token cost, the highest revenue and the longest life, with an ecosystem of offtakers built on CUDA. Goldman Sachs describes the opportunity as creating a market for credit backed by NVIDIA compute. [1]

Atlas interpretation: Every clause in that sentence is doing collateral work. Lending against an asset requires a defensible view of what it is worth if the borrower stops paying, which for an accelerator means two things: that somebody else will want it, and that it will still be worth wanting in five years. "Fungible and transferable across customers and operators" asserts the first. "Continuously improved through CUDA, extending its useful life" asserts the second, and it is the more important of the two, because it says software will slow the depreciation of hardware that a successor generation is designed to obsolete. [1]

Atlas interpretation: That is a coherent argument and it is being made by the party with the most to gain from lenders accepting it. The residual value of a GPU fleet in 2031 is not a fact anyone has; it is an assumption, and it is the assumption that determines whether this is infrastructure credit or something with more equity risk in it than the label suggests. [1]

The shape of the circle

Atlas interpretation: NVIDIA is not lending here; third-party capital is. But NVIDIA is convening the lenders, defining the collateral, and selling the thing being bought with the proceeds. That is not the same as vendor financing and it is not entirely unlike it either. The distinction that matters for risk is whether NVIDIA takes any of the loss when a platform’s borrower defaults, and the announcement does not say. [1]

It also sits inside a run of adjacent commitments in the same month. A week later NVIDIA secured land, power and shell capacity at SB Energy’s Ohio campus, providing credit support for the buildout of the initial 4.25 IT-gigawatts, taking an option on the remaining 3.75, and investing $1.5 billion in SB Energy itself. SB Energy builds, owns and operates; OpenAI is the customer under a 20-year lease; capacity is expected to come online in phases beginning in 2028. [4]

Its customers were signing their own multi-year obligations in the same weeks. Anthropic agreed to pay Nscale about $45 billion over six years for 460 megawatts at a West Virginia campus running Vera Rubin chips, capacity expected online late in 2027. That deal was reported rather than announced, and Nscale had not confirmed it. [5]

Atlas interpretation: Read together, the pattern is a supplier appearing on more sides of more transactions: selling the hardware, convening the lenders, underwriting the building it goes in, and holding equity in the developer. Each individual arrangement has a defensible rationale. What none of them establishes is whether the end demand exists at the volume being financed, which is the question any structure built on usage-linked revenue eventually has to answer. [1][4][5]

What would show whether this worked

Atlas interpretation: The checkable milestones are ordinary credit-market ones and none of them has happened yet: final agreements executed rather than contemplated, a first platform capitalized with a stated size, a named borrower, a stated tenor, and eventually a rating or a spread that lets somebody outside the deal price the residual-value assumption. Until a piece of this paper trades, the market has not agreed with the argument; it has only heard it. [1]