A personal post to Michael Burry
The AI bubble: between an accounting illusion and a deep structural reality?
After the narrative about “cyclical investment” seemed to fade, a new argument emerged in the AI arena:
Michael Burry published a new claim, stating - “The AI bubble is an accounting illusion – GPU depreciation does not reflect its true lifespan.”
According to him, in the reports of major cloud providers (AWS, Azure, GCP), the GPU depreciation period is 6 years, while studies indicate a true lifespan of only 2-3 years.
The implication?
Accounting profit is inflated, and profitability appears too high.
But is he right?
Let's dive a bit deeper into the technical details - it is recommended to read carefully and delve deeper
Does the GPU really “die” after 3 years?
The claim is mainly based on Meta's 2024 report regarding Llama 3 training: The training lasted 54 days on 16,384 H100 units. - 466 activity interruptions were recorded, of which 419 were unplanned.
An estimated annual failure rate of about 9%, meaning a quarter of GPUs could fail within three years
But wait — not every interruption is a failure, and the data was skewed upwards.
Today, thanks to automation and early quality checks by Nvidia, the annual failure rate is estimated at 6% or even less.
Training ≠ Inference. GPUs used for model training wear out faster due to high load and heat. In contrast, GPUs used for inference (ongoing model operation) work under low and stable load —
The failure rate there is less than 3%, sometimes even 2%.
Therefore, a depreciation period of 5-6 years is reasonable – especially as the market gradually shifts to the inference phase, which is the main source of revenue.
- Many companies still use A100s from 5 years ago, and they are still business-relevant.
But wait, there's another very fundamental question here
What about rapid generation replacement?
The answer is very interesting...
True, ostensibly, Nvidia's pace of advancement should lead to early retirement of older generations.
But in practice – demand is so strong that there is no “early scrapping”:
CoreWeave reported that renewed contracts for H100s were sold at almost the same price, and all A100s were snapped up.
Note what CoreWeave CEO, Michael Intrator, said - ‘The A100 is out of stock across the entire industry’.
Listen to what he says - “Our A100s are out of stock. L40s are out of stock. H100s are out of stock. H200s are out of stock. When a contract expires, it is immediately renewed.”
(Even 5-year-old A100s are out of stock - the demand for AI computing is growing exponentially) and that's not all...
He also said the following - “Our TPUs, 7-8 years old, are 100% utilized, which shows the level of demand.”
It's crazy - the demand is so great - that old hardware simply doesn't become obsolete...
So if not depreciation – where is the bubble?
The bubble could still form on the applications side.
AI today has an inverse structure to that of the dot-com bubble: Dot-com era - excess infrastructure, excess investment in applications, unused fiber optics
The AI era?
Exactly the opposite - shortage of applications, shortage of infrastructure, GPUs, electricity, and data centers full until 2026
In the dot-com bubble, infrastructure companies collapsed first
This time – the big risk is precisely in application companies
The risk: the pace of AI Agents' progress is slower than expected, while companies continue to invest huge sums in computing infrastructure to “bridge” the shortage.
This creates a situation where real demand does not keep up with infrastructure growth – a core for a real bubble.
Take CRM stock as an example, a leading company investing heavily in developing AI agents and other applications
But the investors?
Simply not interested..
They are afraid, perhaps rightly so..
Another point to consider
Who holds the debt? - Dot-com bubble – debt with infrastructure companies; profit with applications. - AI bubble – both debt and value with the applications themselves.
Nevertheless, if players like OpenAI and Anthropic maintain a growth rate of 3-9x per year over time, and demand for computing continues beyond 2026 – it is possible that the “bubble” simply reflects a structural shift to a computation-based civilization.
GPUs may wear out quickly, but the bubble (if it exists) is not accounting-based – but psychological and economic.
The race for computing power is essentially borrowing time from the future.
If future growth justifies the investment – we will not remember it as a bubble, but as a revolution.