Wall Street Millennial 对 $NVDA 的判断变化

It shows the complicated web of financial transactions between Nvidia, OpenAI, and numerous cloud service providers. Everybody knows about the circular financing arrangements. It's done out in the open and is widely reported on by the mainstream media. Yet investors don't seem to care. This has led to a surreal situation where the AI bubble has inflated to a grotesque size even as the red flags pile up.

At the center of this all sits Nvidia. Over the past 2 months, Nvidia has made a series of announcements that are concerning to say the least. On July 1st, they announced a new program where they will give loan guarantees to new data center companies to allow them to finance the purchase of new GPUs.

On July 26th, it was reported that Nvidia is preparing to sign a $250 billion loan guarantee to finance new data centers for Open AI.

As if this wasn't crazy enough, on August 10th, Nvidia announced a $500 billion financing vehicle in partnership with a number of banks and private equity firms. Nvidia is taking on huge liabilities, most of which are off balance sheet.

These circular financing deals tend to be very convoluted, making it difficult to ascertain who will end up holding the bag when the AI investment boom finally ends.

Before we look at Nvidia's circular financing deals, we must first understand the AI data center industry. Broadly speaking, AI data center companies are split into two categories, hyperscalers and neoclouds. These are both colloquial terms lacking precise dictionary definitions. I consider hyperscalers to be cloud service providers that predated the AI boom.

But once the AI boom began, they all hopped onto the bandwagon to build new data centers specifically for AI.

Currently, the hyperscalers account for the majority of Nvidia's revenue, but the hyperscalers are pretty much maxed out in terms of how much they can spend. They are pumping all of their cash flows and then some to purchase GPUs.

To continue growing its revenue, Nvidia is turning to the Neoclouds. The problem is most of the Neoclouds are recently founded. They lack free cash flow and have limited access to capital.

On July 1st, 2026, Nvidia published a press release titled Nvidia unlocks AI compute at scale, inviting partners to power the AI infrastructure buildout.

Under this new program, Nvidia will provide a credit support model to new neoclouds. In return, Nvidia will receive a share of future revenue generated by the Neocloud.

The press release is intentionally vague. The first Neocloud to take advantage of this program was a company called Sharon AI.

Nvidia is giving a vaguely defined credit support agreement to a weird data center company that has crypto connections and more red flags than I can count.

If there was really so much demand for Nvidia's GPUs, why have they resorted to making opaque arrangements with Sharon AI? On July 26th, the Wall Street Journal reported that Nvidia is in talks to guarantee $250 billion for a new data center project which will be leased to Open AI.

The proposed project is a gargantuan 10 gawatt data center to be developed in Ohio by SB Energy. Open AAI does not have an investment grade credit rating and its ability to pay $250 billion is highly uncertain.

lenders would bulk at lending this much money to SoftBank. So, Nvidia will provide a guarantee. In the event that OpenAI defaults on the spending commitment, Nvidia has to make SoftBank whole.

Nvidia has an investment grade credit rating. So, this backs stop will give lenders confidence to finance the project. In other words, Nvidia is taking on a massive contingent liability.

To be clear, the deal has not been signed yet. So, we don't know what the exact terms will be. But realistically, only Nvidia has the capability and motivation to provide the guarantee.

And now Nvidia wants to add another $250 billion on top. On August 10th, Nvidia announced its biggest circular financing program yet. They partnered with a number of large investment companies to mobilize over $500 billion of capital to provide subsidized loans to Neoclouds.

Jensen Huang understood that this would create a lot of criticism and skepticism. He preempted this with a very lengthy Twitter post. First, he addresses concerns of data center overcapacity. He claims that data center customers are broad, including Frontier Labs, AI clouds, enterprises, and nations.

Frontier AI labs are anthropic, open AAI, and Google's Gemini. You could maybe include meta in this category, too. But the second category is absurd on its face. AI clouds refer to neoclouds, the likes of Cororeweave, Nebus, and Sharon AI.

The entire purpose of this $500 billion capital pool is to finance neoclouds. Obviously, the Neoclouds are not their own customers. Jensen Huang is relying on the ignorance of his followers to push a narrative that is absurd on its face.

The reality is that the vast majority of Neocloud revenue comes from the frontier AI labs or from middleman contracts with hyperscalers who resell the compute to the frontier AI labs.

The third category is enterprises. Here, Jensen is not only double counting but triple counting. The vast majority of enterprises purchase AI services from the Frontier Labs.

Enterprises represent the majority of revenue for both OpenAI and Anthropic. Fewer than 2% of enterprises purchase compute directly from Neoclouds. This means they run openweight models on Neocloud servers.

Google, OpenAI, and Anthropic all use Neoclouds to varying extents. So if you pay for Chat GPT, there's a chance that the computers being run on a Neocloud. Thus, many enterprises are indirectly customers of Neoclouds without even knowing it. But that of course creates a double counting problem.

A negligible percentage of enterprises purchase Neocloud services directly. So, why is Jensen Huang so intent on pushing a narrative that is so obviously false? It's because he's trying to obiscate the real situation.

The uncomfortable truth is that the vast majority of Neocloud capacity is purchased by just two companies, OpenAI and Anthropic.

Microsoft, Google, and Amazon purchase compute from Neoclouds. But for the most part, they're just reselling that compute to OpenAI and Anthropic. Both of which are losing huge amounts of money and are being propped up by circular financing deals.

Next, Jensen directly addresses the question of circular financing. He says this is not circular financing. The $500 billion being lent to the Neoclouds will come from the likes of Apollo, Black Rockck, Blackstone, Goldman Sachs, and other financial institutions.

It's not Nvidia's money being lent. But here's the problem. Black Rockck and any other investor is free to lend money to any Neocloud they want to. They don't need a partnership with Nvidia to do this.

Many of the new Neoclouds are shady startups like Share and AI with highly questionable creditworthiness. And there are legitimate concerns about overcapacity and the residual values of GPUs.

This $500 billion lending fund to circular financing by stealth. Nvidia may provide a residual value support mechanism for up to 25% of an opportunity.

This is substantially lower than other compute financing arrangements. Nvidia can provide support because Nvidia compute is unique. It is funible, universally adopted, software upgradable, and redeployable across a large ecosystem of customers.

Our role is to help unlock a very large pool of independent capital while maintaining disciplined risk exposure. Hang's framing is extremely misleading. In reality, Nvidia will be taking on massive contingent liabilities.

Let's look at a hypothetical example. Let's suppose a new Neocloud startup is founded. Black Rockck gives them a $1 billion loan that matures in 5 years. The Neocloud uses this money to buy Nvidia GPUs.

For the sake of simplicity, we'll just assume there is no interest rate. The Neocloud sets up a data center and generates $150 million of Ebida per year by leasing the data center to OpenAI or Anthropic.

At the end of 5 years, the Neocloud has cumulatively generated $750 million of Ebida. But this is not enough to pay back the loan. So, the Neocloud defaults.

Black Rockck can repossess the GPUs. But how much are these 5-year-old GPUs worth? Maybe by this point they're worn out.

Or maybe by this point there's a compute glut and rental prices have declined to below the operating costs, thereby rendering the GPU worthless. This could happen either because AI demand slows or Nvidia releases new generations of GPUs that are far more efficient.

There are a lot of potential scenarios where these GPUs might be worth significantly less than $250 million in which case Black Rockck will take a loss.

This is why lenders are hesitant to lend money to Neoclouds.

That's why Nvidia created the residual value support mechanism. In the event that the GPUs are worth nothing, Nvidia would have to pay $250 million to make Black Rockck hole.

Jensen Hang tries to frame it as it's only 25% so it's not a big deal. But this misses the point. In no scenario will the NeoCloud go bankrupt immediately. They'll generate some Ebidot and they'll be able to pay

The argument that Jensen is trying to make is that his GPUs are so useful that even after 5 years, they'll still be worth at least 25% of their original sale price. Jensen Huang might believe this, but Black Rockck, Apollo, Goldman Sachs, and the others are not convinced. That's why they're demanding Nvidia to provide them with this residual value support mechanism.

What Nvidia is doing is very devious. If they wanted to, they could lend money to the Neoclouds directly. But if Nvidia were to lend the money directly, this would show up on Nvidia's balance sheet, and Nvidia would have to estimate provisions for credit losses. So instead, they're laundering the loans through other financial institutions, while Nvidia bears some, if not most of the risk. This is a contingent liability and thus does not show up on Nvidia's balance sheet. If in a few years time the Neocloud start defaulting, Nvidia will be on the hook to recognize enormous losses which will seemingly pop out of nowhere.

Over the past 12 months, Nvidia generated $ 160 billion of operating profit. Even without circular financing, they would be one of the most profitable companies in the history of the world. With a market cap of $5 trillion, Nvidia is the most valuable company in the world.

But for Hong, it's never enough. He wants even more money. So, he's now having Nvidia take on enormous risks while doing everything he can to obiscate this from his shareholders.