The Nvidia community this week has truly erupted into open conflict.
On August 13, Michael Burry, the real-life model for ‘The Big Short,’ posted on X directly equating Nvidia’s AI expansion with the 2001 Enron scandal, claiming its degree of danger is several orders of magnitude greater than Enron.Wall Street Insight
Two days earlier, ‘Dr. Doom’ Jim Chanos also delivered a cutting remark, suggesting that the next time these people sit at the same table to explain AI financing, it might be at a 2031 congressional hearing.Wall Street Insight
What the two short-sellers are attacking is the same thing: the $500 billion financing plan that Jensen Huang announced on August 10. On Zhihu, the question ‘Is Nvidia being wronged?’ quickly split into two camps, and the comment section of the relevant video on Bilibili was flooded with things like ‘new subprime loans’ and ‘the bubble will burst eventually,’ with some even joking ‘at least you’ve got the physical card in hand, better than an off-plan property.’Bilibili
Today, let’s lay out the face-up cards, the hidden cards, and the vital points of this deal all at once.
What on earth is this $500 billion play?
On August 10, Nvidia signed a memorandum with six financial giants — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — aiming to mobilize more than $500 billion in third-party capital, dedicated to investment in AI infrastructure. Jensen Huang’s definition of it is this: ‘This is indeed the first time that technology chips have become an investable asset class,’ ‘Compute is revenue.’Zhihu
Translated into plain language: these institutions issue bonds to pension funds, insurance companies, and sovereign wealth funds, pour the raised money into a specially established platform, and then lend it to AI companies and cloud vendors to buy and lease Nvidia GPUs, with the underlying collateral being GPU rental income.36Kr
To sum up in one sentence: GPUs have changed from ‘products sold in a lump sum’ into ‘income-generating assets that can be mortgaged, packaged, and turned into bonds.’
And this is just one move in a series of consecutive plays. In the same period, Nvidia open-sourced the 30B-parameter Nemotron 3.5 Lightning model and released the smart router NeMo Switchyard — according to LangChain’s tests, 93% of agent calls can be handled with just the 30B small model. Selling compute at the hardware layer, locking in the ecosystem at the model + routing layer, and finding money for compute at the capital layer — stacking these three layers is what Jensen Huang calls the ‘AI factory.’Zhihu
What is Burry betting on?
Burry’s original words were actually more restrained than the headlines: ‘Credit structuring is a natural part of the financial system. But in the late stages of a bull market, structuring unnatural credit to sustain momentum — that is what is truly concerning.’ Citing the Bank for International Settlements (BIS) annual report as backing, his core suspicion is one sentence: much of Nvidia’s demand is off-balance-sheet financing self-circulating, and the real end customers are not as numerous as the market believes.
And he isn’t just talking. In July he disclosed that he had added shorts on Nvidia, Micron, and the Philadelphia Semiconductor ETF, and in early August he rolled over the Nvidia short position to June 2027.Weibo
Another ‘Big Short’ prototype, Eisman, threw out a warning from a different angle on the same day: the fatal weakness of the AI boom is that it increasingly depends on the fate of the two companies OpenAI and Anthropic.Weibo
But there is one piece of background worth noting: Burry’s main business now is the paid Substack ‘Cassandra Unchained,’ with subscriptions already exceeding 300,000. The bombardment itself is also his traffic business. The rhetorical element of the ‘Enron analogy’ needs to be discounted.
What are Jensen Huang’s cards?
Jensen Huang doesn’t take up the骂 (insults); he throws out data.
First, old cards still hold value: the A100 released in 2020 is still in active commercial use in training and inference six years later, and customers are still signing multi-year contracts; CoreWeave recently signed an A100 contract at near full price, valid through 2029. The co-founder’s explanation: inference and engineering demand is stretching the lifespan of older GPUs longer and longer.
Second, compute is not dropping, it’s rising: the one-year lease price for the H100 rose from $1.70/hour in October 2025 to $2.35/hour this March, and further rose to $2.70/hour in June, a 59% increase over 8 months; cloud quotes for the B200 have already reached $5.30 to $7.05/hour.Wall Street InsightZhihu
Third, adding another layer of insurance: in specific projects, Nvidia provides up to 25% residual value support, with a total cap of about $125 billion — if the value of the collateral hardware falls below the baseline, Nvidia contributes to make up the difference, positioning it as the ‘first loss-absorbing layer’ of this financing round.36Kr
Also standing in solidarity on the same stage are the heads of Blackstone and BlackRock: BlackRock CEO Larry Fink even compared the current moment to the golden age of his early career, the 1970s, when the mortgage-backed securities (MBS) market was being pioneered. Bank of America analysts see through another layer: with Wall Street financial institutions coming in, Nvidia can instead exit the ‘vendor financing’ model, and the burden no longer rests on its own balance sheet.
The real point of contention is hidden in one detail
Look carefully, and there is a hidden seam in this deal.
Financial executives involved in the deal revealed that many lenders currently require that loans collateralized by GPU leases be fully repaid within 3 to 5 years, premised on the assumption that the value of the underlying chips will be negligible thereafter.36Kr Meanwhile, what Jensen Huang talks about externally is that the chips’ ‘economic useful life is extending toward 10 years.’
In the same deal, the financial side insures on a ‘short life’ basis, while Nvidia tells a ‘long life’ story. This may be the most honest part of the entire $500 billion plan: even the participating institutions don’t fully believe the long-term value-preservation narrative; they are using loan terms to control risk.
Morgan Stanley drew a diagram of this ecosystem’s capital flows long ago: Nvidia’s investments flow to OpenAI and the CoreWeaves, and part of that money comes back as chip orders and rent. So the essence of what the two sides are arguing about can be stated in one sentence: how much will the old cards still be worth in a few years?
Jack Ablin, partner at Cresset, a family office managing $260 billion in assets, put it most harshly: ‘From historical experience, the shelf life of this kind of asset is not much different from a fresh head of lettuce.’ Jensen Huang is betting on land — land doesn’t become obsolete, and rent keeps rising.
To be clear, the Enron analogy does not hold — Enron was accounting fraud that hid debt in off-balance-sheet shells, whereas Nvidia’s $500 billion is transparent third-party capital, independent due diligence, and clearly priced residual value clauses. But the real problem Burry poked at cannot be avoided: the long-term residual value of GPUs is the lifeline of this financing, and no one can prove it right now.
There are two more mirrors in history: at the turn of the century, telecom equipment makers like Lucent and Nortel lent money to customers to buy equipment, and when demand collapsed, the debt brought the companies down with it; in 2008, subprime lending bet on ‘housing prices will always rise.’ This time, the bet is ‘compute will always be insufficient.’
Four signals are more useful than picking a side
Right now, neither the longs nor the shorts can prove themselves right; both sides are betting on tomorrow’s outcome with today’s evidence. Rather than rushing to pick a side, it’s better to keep an eye on these signals:
One is Nvidia’s 5-year CDS spread. It has already surged about 90% this year, and on the day of the August 10 announcement it briefly reached 77.2 basis points, the largest single-day increase in two weeks.Wall Street Insight This is the most direct barometer of credit market sentiment, and if it keeps climbing, it means smart money is seriously pricing in the risk.
Two is the implementation of the terms of the first batch of projects. The $500 billion has not yet raised a single cent, and the 25% residual value backstop has not yet materialized for a single project. Once the first batch actually lands, the terms will make clear how much risk Nvidia is actually carrying.
Three is old-card rent. If A100 and H100 lease prices stay firm or keep rising, it means the flywheel is still turning; the moment a drop appears, that’s the first signal of the ‘oversupply’ narrative.
Four is weathervane companies and earnings reports. CoreWeave’s orders are booked through 2029, Nebius can sell out its 2027 capacity, and adding Nvidia’s new quarterly earnings report later this month — all are answer sheets on the demand side.
One more reference point: in Q2, several domestic private equity firms simultaneously reduced their Nvidia holdings — Gaoyi’s reduction ratio exceeded 70% (only 80,000 shares remaining), Jinglin fully liquidated, and Oriental Harbor reduced by about 15.7% — while simultaneously adding to TSMC and memory, with Gaoyi’s Micron holdings increasing by more than 283%.36Kr Their logic is not necessarily calling an AI peak; rather, it’s more like shifting positions from the most crowded ‘expectations’ to bottleneck segments with verifiable earnings.
What does this have to do with you?
If you’re an investor, this long-short battle determines the valuation direction of Nvidia and the entire AI compute chain. Don’t rush to follow the shorts or the longs; the signals above are more honest than anyone’s mouth.
If you’re a graphics card buyer, the recent wave of GPU price hikes is the most direct manifestation of ‘compute shortage.’ Whether this shortage narrative can be sustained — whether financing can keep rolling, whether rent can keep rising — actually determines whether card prices can hold up going forward. Focus on signals two and three.
Even if you’re just watching for fun, this matter provides excellent talking material: whether GPUs are ‘land’ or ‘lettuce’ will become clear in a few years.
One last sentence: today, no one can prove anyone else wrong. What we can do is write down these signals and check them off one by one over the coming quarters. The above is an organization of information and does not constitute investment advice.
