Understanding Michael Burry's Bet Against AI: Here's What it Really Means for Investors
Michael Burry, manager of Scion Capital, bets against AI companies like Nvidia, arguing that hyperscalers are overly optimistic about hardware depreciation rates. He claims depreciation is understated by $176 billion, leading to overstated earnings for companies like Oracle and Meta. Burry predicts an AI spending bubble burst as earnings assumptions unravel, affecting investment returns. Critics argue that AI investments focus on long-term gains, not immediate earnings, and question the need for earnings manipulation at the expense of cash flow.
Famed investor Michael Burry's bets against AI-focused companies such as Nvidia (NVDA 1.06%) have attracted significant attention, not least because the reasons he gives to support his views challenge the investment thesis for many major technology companies, notably those in the most exciting growth area of the economy right now.
Burry's case against AI stocks
Burry, the manager of hedge fund Scion Capital, said he believes that hyperscalers are making overly optimistic estimates for the rate at which their AI-related hardware, like servers, GPUs, and network equipment, will depreciate. Basically, he says they assume that the useful lifespan of that hardware will be longer than he thinks is realistic.
Image source: Getty Images.
Instead, Burry thinks these companies will eventually need to raise their depreciation rates (reflecting a shorter useful period for the hardware) to reflect the rapid pace of technological development centered around AI by companies like Nvidia. He projects that across the industry, depreciation has been understated by approximately $176 billion between 2026 and 2028. And because of this, he claims that companies like Oracle (ORCL 5.74%) and Meta Platforms (META +0.87%) could be overstating their earnings by nearly 27% and 21%, respectively.
The corollary to this thesis is that when all of these allegedly incorrect assumptions unravel, the AI spending bubble will burst as the sector adjusts its spending to reflect the corrected level of earnings at these companies. Moreover, those adjustments to earnings will lead to a negative assessment of the returns on investment in AI.
NASDAQ: NVDA
Key Data Points
Why depreciation matters
In accounting terms, depreciation refers to the allocation of the cost of a tangible asset over its useful life. Therefore, a depreciation rate of five years implies that the asset's value will be zero after five years, while a lower depreciation rate, of say 10 years, implying a longer useful life, will leave that asset with a value of zero after a decade.
Because depreciation implies an asset is being used, it is treated as an expense that lowers profitability. As such, a company can bolster its earnings in the near term by lowering its depreciation rate and assuming a longer useful life for its assets -- in this case, network equipment and servers.
Are hyperscalers lowering depreciation rates?
Combing through hyperscalers' 10-K filings with the Securities and Exchange Commission reveals that Burry is correct that the general trend is toward those companies' using lower depreciation rates for their network equipment and servers, with one recent exception: Amazon.com (AMZN +1.51%).
In 2025, Amazon lowered its depreciation rate to five years due to "an increased pace of technology development, particularly in the area of artificial intelligence and machine learning." This somewhat supports Burry's argument, and it may well reflect Amazon's early and aggressive investment in AI-capable hardware, setting a pattern that the other hyperscalers follow.
|
Company |
Last Change Date |
Change in Useful Life Assumption |
|---|---|---|
|
Alphabet |
2023 |
"servers from four years to six years" and "certain network equipment from five years to six years" |
|
Amazon |
2025 |
" from six years to five years" |
|
Meta Platforms |
2025 |
5.5 years (from four to five years) for certain servers and network assets |
|
Microsoft |
2022 |
" server and network equipment from four years to six years." |
|
Oracle |
2025 |
"from five years to six years" for servers |
Data source: SEC filings
What if Burry is right?
If Burry is correct, there are a few possible consequences:
- Earnings growth assumptions will eventually be pared back at hyperscalers to accommodate faster depreciation rates.
- If capital spending is based on an assumption of a useful life of six years, but the reality is three years, then hyperscalers will need to engage in significantly more capital spending, which will be detrimental to their cash flows.
- It could lead to a more pessimistic assessment of the returns available from investing in AI, resulting in a slump in spending.
Three criticisms
First, lowering depreciation boosts earnings, but it also reduces cash flow, as the company must pay taxes on the increased earnings. It's questionable why Alphabet (GOOG +3.33%) (GOOGL +3.53%), Microsoft, and Meta Platforms, in particular, would need to manipulate their earnings to appear better at the expense of cash flow.
Those companies are highly cash-generative despite their significant AI spending, and they can fund investments. Moreover, debt rating agencies and capital markets focus on cash flow and earnings before interest, taxation, depreciation, and amortization (EBITDA), rather than net income.
Data by YCharts.
My second criticism is that AI investment isn't really about near-term earnings. The hyperscalers are doing it to generate long-term earnings and cash flow from recurring customers. For example, Alphabet's Google Cloud generated a $700 million loss in the third quarter of 2022, but in this year's Q3, it reported a $3.6 billion profit with 85% year-over-year growth.
Third, Nvidia's management noted on its earnings call this month that the A100 GPUs "we shipped six years ago are still running at full utilization today." This supports the assumption of a useful life of five to six years discussed above.
Image source: Getty Images.
What it means to investors
The real issue is that, alongside the productive AI investment, there's likely to be a volume of unproductive AI investment that will eventually have to be discontinued. Therefore, the overall level of investment growth will ultimately be scaled back. However, predicting that point, or the underlying growth rate at which AI spending will eventually settle, is extremely difficult, and examining depreciation rates won't provide much help in finding it.
If you want to play it safe, then it's a good idea to focus on cash-flow-based valuations rather than earnings, and avoid the less financially robust options such as Oracle and Amazon.
