Prepared by: Andy Choi
Confidentiality Statement: This document is intended for informational purposes only and is based on publicly available information and the provided source material. However, the analysis and insights contained herein are proprietary and copyrighted. This document should not be distributed, reproduced, or disclosed without prior written consent.
NVIDIA has always been a certain and viable investment for anyone. Strong earnings, main suppliers of chips for Artificial Intelligence, and one of the most owed stock in the market. But I sold all my shares today after doing immense research.
This is an argument that a significant chunk of Nvidia's "demand" is an internal, self-funding loop, also being called a "mirage of growth" rather than true, organic demand from the outside world.
Here are the specific deals that analysts point to as evidence of this loop:
Specific Example 1 (Nvidia <-> OpenAI): Nvidia, the chip supplier, announced it intends to invest up to $100 billion in OpenAI, the AI developer. OpenAI, which is a cash-burning company, then uses this capital to buy Nvidia's next-generation GPUs. Looping profits back into Nvidia, in this case, Nvidia is directly funding its own customer, who then gives the money back to Nvidia in the form of "revenue."
Specific Example 2 (Oracle <-> OpenAI <-> Nvidia): This is a three-way loop.
Oracle (a hyperscaler) and alongside many announced a massive $300 billion cloud infrastructure deal for OpenAI.
To build this infrastructure, Oracle (which already has high debt) had to take on $38 billion in new debt.
Oracle will spend that $38 billion buying the necessary GPUs directly from Nvidia. In this loop, Oracle's future is bet on OpenAI, but its immediate spending is funneled to Nvidia, all while OpenAI's own funding is dependent on partners like Nvidia.
Specific Example 3 (Microsoft <-> CoreWeave <-> Nvidia): This loop involves a "neocloud" (a specialised cloud provider).
Nvidia (the supplier) holds an equity stake in a "neocloud" company called CoreWeave.
CoreWeave raises billions of dollars in debt to buy... as you know it, MORE GPUs from Nvidia.
Microsoft (Nvidia's biggest customer and OpenAI's biggest backer) then signs a $60 billion deal to use CoreWeave's capacity. In this case, all parties are investing in and buying from each other, making it impossible to separate "investment" from "sales."
Now why is this bad?
It's not real demand. It's like a group of people at a table using the same $100 bill to buy things from each other. It looks like $500 in "sales," but only $100 of real money ever existed. It's a fragile, self-funding system that inflates revenue and masks how much organic, external demand (from real, profitable end-users) actually exists.
However, please do keep in mind that, those are specific 3 examples. Nvidia has a lot more customers than just those three. But this hypothesis isn't even the worst one.
Introducing, Hypothesis B - Fraudulent Accounting.
This hypothesis targets the real customers: the "hyperscalers" (Amazon, Google, Microsoft, Meta, and Oracle).
The argument is that these companies cannot actually afford their massive spending sprees, and they are hiding this fact from investors using a specific accounting "trick."
Here are the very specific examples of this "trick":
Specific Example,
The Problem (The Truth): Nvidia has an aggressive, annual upgrade cycle. They launched the "Hopper" (H100) chip, followed by the "Blackwell" chip, and have announced the "Rubin" chip for 2026. Burry's argument is that this makes the true economic life of a cutting-edge AI chip only 2-3 years. After that, it's obsolete for competitive AI training.
The "Trick" (Accounting Fiction): Instead of depreciating (expensing) these chips over the 2-3 years they are actually useful, the hyperscalers have all changed their accounting rules to depreciate them over 5, 6, or even 8 years.
The Company-Specifics:
Google (Alphabet) changed its server useful life from 3 years, to 4, and now to 6 years.
Amazon (AWS) and Microsoft also pushed their server lifespans to 6 years.
Meta (Facebook) uses 5.5 years.
Oracle has been cited as extending its useful life to eight years.
He even singled out the worst offenders. He claims that by 2028, this accounting practice will cause Oracle's earnings to be overstated by 26.9% and Meta's earnings to be overstated by 20.8%.
The Result: This accounting sleight-of-hand artificially inflates their profits by making their expenses look much smaller than they really are. Burry calculates this trick is hiding $176 billion in understated costs.
A bet against Nvidia is a bet that this accounting "house of cards" will collapse. When the market forces the hyperscalers to correct their accounting, their massive profits will vanish. They will be forced to immediately slash their budgets and stop their multi-billion dollar chip-buying spree. When that happens, Nvidia's primary customer base will dry up. In short, they are hiding behind the massive amount of cash burn they do on AI, as well as inflating their profits and revenue.
Is there a bubble? Yes. The analysis concludes that the AI market is a massive speculative bubble. It is propped up by a combination of aggressive accounting (understating depreciation costs) and "circular financing" (creating artificial demand).
Is Michael Burry right or wrong? I think he is directionally correct. His core economic analysis does make sense. He has correctly identified that the market is ignoring a massive, recurring hardware replacement cost that is being hidden by these accounting practices. As in 2005, he is not "crazy", I think he is simply early.
Final Opinion: A tech-sector correction is highly probable. However, the resulting crash will not be a 2008 style systemic banking collapse.
