Nvidia spent the last three years selling shovels in a gold rush. This week it did something different: it started arranging the loans that let people buy the shovels. On August 10 the company announced agreements with six of the largest asset managers on Wall Street to build financing platforms aimed at mobilizing more than $500 billion for AI infrastructure. Within a day, the word “circular” was attached to it on every trading desk in the country.
What Nvidia Actually Announced
Read the announcement carefully and it is narrower than the headline number suggests. These are memorandums of understanding, not funded commitments. The $500 billion is a target for third-party capital to be mobilized over time, not a pool sitting in an account waiting to be drawn.
Here is the structure in plain terms.
| Element | What it means |
|---|---|
| Who provides the money | Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR, each through its own platform |
| Nvidia’s role | Connector and standard-setter, not a lender and not a guarantor |
| The collateral | GPU systems, treated as income-producing equipment rather than depreciating electronics |
| The safety valve | Borrowers use Nvidia-specified architectures so another operator can assume the hardware after a default |
| Status | Preliminary. Rates, borrowers and timelines still being worked out |
The standardization piece is the genuinely clever part and it is easy to skim past. A lender’s core problem with GPU-backed debt is that if the borrower goes under, the lender is holding a warehouse full of hardware it has no idea how to operate or sell. By insisting borrowers build to a common Nvidia reference architecture, the hardware becomes fungible. Any competent operator can pick it up and keep the revenue flowing. That is the difference between collateral and scrap.
Why “Circular Financing” Keeps Coming Up
The criticism is straightforward: a chip vendor helping arrange the capital that customers use to buy its chips is booking revenue that its own ecosystem financed. Nvidia’s defense is equally straightforward, and it is not wrong on the facts. The company is not the lender. Six independent institutions with their own fiduciary duties and their own risk committees decide what to fund.
The problem is that the strict definition is not really what skeptics are worried about. They are worried about the shape of the thing. Nvidia has already been an anchor investor in vehicles that buy Nvidia hardware and lease it out. Apollo committed roughly $7 billion across two back-to-back deals in early 2026, structured as sale-leasebacks through an entity called Valor Compute Infrastructure, which leased capacity to xAI. In that arrangement Nvidia sat as an anchor limited partner in a fund purchasing Nvidia’s own product. Whatever you call that, the money leaves through one door and comes back through another.
The historical comparison analysts keep reaching for is telecom vendor financing in the late 1990s. Lucent and Nortel lent aggressively to carriers so those carriers could buy switching equipment. It worked beautifully while demand grew and became catastrophic the moment it stopped, because the vendors had converted a sales problem into a credit problem and then discovered they owned both.
The Number Everyone Should Be Watching Is the Residual Value
Every one of these deals rests on an assumption about what a GPU is worth in three years. Early GPU-backed securitizations, starting with a $500 million Lambda Labs offering arranged through Macquarie in mid-2024, were built on the premise that the hardware would retain roughly half its value after three years.
That assumption is doing enormous work. If it holds, the loans are well collateralized and the whole edifice is boring and sensible. If GPUs depreciate faster than expected, whether because Nvidia’s own product cadence makes each generation obsolete quickly or because Chinese manufacturing pushes accelerator prices down, then the collateral erodes faster than the debt amortizes. Lenders end up undersecured on hundreds of billions of dollars of paper, and they find out at the same time.
This is where the price war in AI services becomes relevant rather than just interesting. Compute economics on the demand side have been moving fast and mostly downward, and we have covered how OpenAI and DeepSeek have been racing each other toward the floor on inference pricing. Cheaper inference is wonderful for anyone building products. It is less wonderful for the models that assume each GPU generates a predictable revenue stream for the life of a loan.
What Wall Street Actually Said
The reaction from analysts was notably measured, and the measured version is more damning than an outright pan would have been. Mizuho’s read was that expanding available capital does not answer the question underneath all of this, which is how much genuine end-user demand and economic return exists to service the debt. Morgan Stanley made a related point: whether the AI ecosystem should be taking on this much leverage is its own investment debate, separate from the question of where the leverage comes from.
Translated out of analyst language, both firms are saying the same thing. Solving the funding constraint does not create demand. It just makes it possible to build much more capacity much faster, which is exactly what you want if the demand is real and exactly what you do not want if it is not.
There is a China dimension too. A financing structure built around Nvidia-specified architectures assumes those architectures remain the standard everyone builds to. Export restrictions and a domestic Chinese accelerator industry that keeps improving both chip away at that assumption in ways nobody has priced precisely.
The Efficiency Question Nobody in This Deal Wants to Discuss
All of this capital is being raised on the premise that frontier AI requires enormous, centralized, expensive compute. That has been true. It is becoming less uniformly true.
Capable models keep getting smaller and cheaper to run. Meta open-sourced a 30-billion-parameter agent that runs on a single consumer gaming GPU, which would have sounded absurd two years ago. Meanwhile the enterprise picture is messier than either camp admits, since token prices have collapsed while corporate AI bills have kept climbing, because usage grew faster than unit costs fell.
That combination is the genuine uncertainty at the center of a $500 billion bet. Total demand for compute may well keep rising for years. But the specific assumption that it must be served by the most expensive hardware, financed over multi-year terms, at data center scale, is an assumption rather than a law.
What to Watch From Here
- Whether the MOUs become funded deals. Memorandums are cheap. The first priced transaction, with a real rate and a real borrower, will tell you what lenders actually think the collateral is worth.
- Residual value assumptions in new paper. If the 50 percent three-year figure starts drifting down in deal documents, that is the market repricing the risk in public.
- Who the borrowers are. Established hyperscalers with real revenue are one credit story. Newer AI-native companies burning cash toward an uncertain business model are a very different one.
- Nvidia’s own participation. The company insists it is not the lender. If it keeps showing up as an anchor investor in the vehicles buying its hardware, that distinction gets harder to defend.
None of this means the plan is doomed. Standardizing GPU collateral is a real innovation, and unlocking institutional capital for infrastructure that would otherwise be funded from a handful of corporate balance sheets is a reasonable thing to want. But the history of vendor-adjacent financing in a booming technology sector is not a comforting one, and the people raising the objection are not cranks. They are the same analysts who will have to explain to clients what happened if the residual value assumption turns out to be wrong.
