Understanding Michael Burry's Nvidia Short: A Deep Dive Using Microsoft's Numbers

When a $4 trillion company issues a formal rebuttal to a short seller, you know something serious is happening. That's exactly what happened when Nvidia responded to Michael Burry's bearish thesis with a memo.

Most investors are confused about what Burry is actually arguing. Headlines scream "Burry shorts Nvidia!" but miss the nuance of his position. This isn't a story about fraud, accounting manipulation, or even Nvidia's business model.

It's far more subtle, and potentially more dangerous for the entire AI infrastructure ecosystem.

To cut through the noise, we're going to explain Burry's thesis using Microsoft's actual financial data from our platform. Not estimates. Not assumptions. Real numbers from their SEC filings that show exactly what he's warning about, and why it matters for every investor holding big tech stocks.

What Burry Is NOT Saying

Let's start by clearing up the biggest misconception.

Michael Burry is not claiming Nvidia is cooking its books. As he explicitly stated: "I made no such claim because no such claim is plausible. Nvidia is a fabless company that only designs chips. It has very little capital expenditure and very little depreciation."

Nvidia designs chips but doesn't manufacture them - Taiwan Semiconductor does. This means Nvidia's balance sheet is light on long-lived assets. They don't have massive depreciation expenses because they don't own the factories.

So if Burry isn't targeting Nvidia's accounting, what's he actually saying?

The thesis is about Nvidia's customers: Microsoft, Google, Amazon, and Meta. These "hyperscalers", i.e., companies that own and operate the world's largest data centers, are spending hundreds of billions of dollars buying Nvidia's chips. And Burry argues they're systematically misstating the economic reality of their GPU purchases.

The Core Thesis: Economic vs. Physical Lifespan

Here's Burry's argument in plain English:

When hyperscalers buy GPU servers, they capitalize them as long-term assets and depreciate them over their "useful life." Microsoft extended server depreciation from 4 years to 6 years in 2022. Meta recently extended to 5.5 years.

But Burry argues there's a critical mismatch between physical lifespan and economic lifespan.

The technology is advancing so rapidly that older chips become economically obsolete long before they physically break. Consider the progression:

• A100 chips (2020): The workhorse of the early AI boom
• H100 chips (2022): 2x more power efficient per watt, 3-6x more performance
• Blackwell chips (2024-2025): Up to 30x faster inference and 25x better energy efficiency for LLM workloads

This isn't incremental improvement - this is generational acceleration.

Imagine running a data center with A100 chips from 2020. They work perfectly fine, but while A100s consume less power per chip (400W vs 700W for H100), they deliver roughly half the performance per watt. For the same computational output, you need significantly more A100s, resulting in much higher aggregate power consumption.

In a world where data centers consume megawatts of power, your operating costs are exploding while competitors with newer chips run the same workloads for a fraction of the power cost.

As Burry writes: "An older chip will be a residual value unit that is an energy hog and extremely costly, not to mention difficult to run outside a same-vintage data center."

The chips don't die. They just become uneconomical to operate.

The Accounting Impact

Let's understand this using Microsoft as an example. Microsoft purchased 485,000 Nvidia chips in 2024. At approximately $35,000 per chip, this represents roughly $17 billion in GPU purchases in 2024 alone.

Here's the math:

If depreciated over 6 years (current policy): Annual depreciation = $17B ÷ 6 = ~$2.8B/year

If economic life is actually 3 years (Burry's argument): Annual depreciation = $17B ÷ 3 = ~$5.7B/year.

The difference: ~$2.9B/year in overstated earnings.

Microsoft's FY2024 net income was $88.1B. A $2.9B overstatement represents 3.3% of reported profits.

That might not sound like much, but this is just from one year of GPU purchases. If similar spending occurred in 2022, 2023, and 2025, the cumulative overstatement could be $10-12B annually, or roughly 11-14% of reported earnings.

What This Means for the Stock Price

Currently, Microsoft trades at approximately $492 per share with a P/E ratio of 34 and earnings per share of $14.11. If earnings were adjusted down by 11-14% to reflect realistic GPU depreciation, the adjusted EPS would fall to $12.13-$12.56. Assuming the P/E ratio remains at 34, the stock price would drop to $412-427 - a decline of $65-80 per share, or roughly 13-16%. However, if investors also lose confidence in AI infrastructure returns, the P/E multiple could compress further, potentially amplifying losses beyond the accounting adjustment alone.

The Valuation Impact

If Microsoft's GPUs truly become obsolete in three years instead of six, the company faces more than just an accounting adjustment - it faces a structural cash flow problem.

Current DCF Value

Stockoscope's DCF model values Microsoft at $384.93 per share, with our DCF model estimating MSFT priced ~22% above its intrinsic-value estimate at the current price of $491.92. This valuation assumes capital expenditure will average approximately 15.7% of revenue going forward, based on historical levels ranging from 11.2% in FY2020-2021 to 18.1% in FY2024 and 22.9% in FY2025.

However, Burry's thesis suggests otherwise: if GPUs must be replaced every three years instead of six, Microsoft will face ongoing replacement capex of roughly $17 billion annually (replacing $17B purchases every 3 years vs. every 6 years), keeping total capex elevated at 20% of revenue or higher indefinitely.

The Burry-Adjusted Fair Value

We recalculated our DCF model assuming capex remains at 20% of revenue instead of declining to 15.7%. The additional $15-20 billion in annual capital expenditure reduces free cash flow each year, which lowers the total enterprise value by $547 billion and the per-share intrinsic value by $73.44.

Thus, the Burry-adjusted fair value becomes $311.49 per share.

This represents a 19% reduction from our baseline fair value, with our DCF model estimating MSFT priced ~37% above its intrinsic-value estimate at the current price of $491.92.

Sensitivity Analysis

Table 1. Microsoft’s fair value under different capex scenarios.

Sustained Capex (% of Revenue)	Intrinsic Value	Upside / (Downside) vs $492
15.7% (Base Case)	$384.93	(22%)
17.5%	~$354	(28%)
20.0% (Burry Base Case)	$311.49	(37%)
22.5%	~$281	(43%)
25.0%	~$246	(50%)

The key insight: you don't need to accept Burry's full thesis to see meaningful overvaluation. Even a modest increase, where capex remains at 17.5% of revenue rather than declining to the baseline 15.7%, implies a 28% overvaluation.

If capex stays elevated at 20% (Burry's base case), the overvaluation reaches 37%. In more aggressive scenarios where AI infrastructure spending climbs to 22.5-25% of revenue, Microsoft could be overvalued by 43-50%.

The margin of safety is thin. Unless capital spending moderates significantly from current levels, Microsoft appears richly priced even under conservative assumptions.

The Nvidia Connection: How This Destroys Demand

Now we come full circle to why Burry is short Nvidia, not Microsoft.

If hyperscalers recognize that GPUs become obsolete in three years instead of six, the financial pressure intensifies. Sustained capex at 20%+ of revenue permanently reduces free cash flow by 15-20%. When CFOs present the reality ($60-80 billion annually for infrastructure with a 36-month shelf life), boards will demand better returns and more disciplined spending. As capital allocation tightens and upgrade cycles extend, demand for Nvidia chips could collapse, potentially destabilizing the entire AI infrastructure market.

The evidence suggests this is already starting:

• Amazon's admission: Shortened GPU depreciation from 6 years to 5 years in January 2025, costing them $700 million in additional depreciation.
• Microsoft's FCF stall: Our data shows FY2025 operating cash flow grew 15%, but free cash flow declined 3.3%. Capex exploded to $64.6B - 22.9% of revenue.

Table 2. In FY 2025, Microsoft's operating cash flow grew 15%, but free cash flow declined 3.3%.

Fiscal Year	Operating Cash Flow	Capex	Free Cash Flow	Capex Growth
FY2021	$76.7B	$20.6B	$56.1B	-
FY2022	$89.0B	$23.9B	$65.1B	16%
FY2023	$87.6B	$28.1B	$59.5B	18%
FY2024	$118.5B	$44.5B	$74.1B	58%
FY2025	$136.2B	$64.6B	$71.6B	45%

• Nadella's warning: "I didn't want to go get stuck with four or five years of depreciation on one generation." He's essentially admitting chip generations turn over faster than accounting assumes.

Why Short Nvidia Instead of Microsoft?

Microsoft has a diversified business. Even if Azure AI disappoints, it still has Office, Windows, LinkedIn, and gaming. The stock might be overvalued, but the company isn't going away.

Nvidia is a pure play on AI infrastructure demand. If hyperscalers slow purchasing even modestly, Nvidia's revenue collapses. The company is priced for perfection, assuming indefinite exponential growth.

But if our analysis is correct, and Microsoft's fair value is 36% lower once you account for sustained high capex, then the appetite for continued GPU purchases at current prices evaporates. CFOs can't justify it. Boards won't approve it.

That's the trap Burry sees: Nvidia's revenue depends on customers making economically irrational decisions. Once the music stops, the stock has nowhere to hide.

The $500 Billion Question

Michael Burry isn't betting against artificial intelligence. He's not claiming Nvidia makes bad products. He's not even saying Microsoft is a bad company.

He's asking a simpler, more fundamental question: Can Microsoft and its peers sustain $60-80 billion per year in capital expenditures indefinitely, when the infrastructure they're building may need to be replaced every 3 years instead of 6?

The market is betting "yes" - that AI will generate returns justifying this spending, that chips will maintain economic value for their full depreciation period, and that demand will remain robust.

Burry is betting "no" - that the accounting assumptions don't match reality, that CFOs will eventually rein in spending when the math doesn't work, and that Nvidia's demand will cliff when that happens.

Time will tell who's right. But as Burry has shown before, most memorably in 2008, sometimes the consensus is wrong, and the uncomfortable truth is hidden in the accounting footnotes and cash flow statements.

The data is available. The assumptions are quantifiable. The math is clear. The only question is whether the market is pricing in reality or hope.

---

This article was originally published on Medium on 5 December 2025. All data, prices, and results reflect information available at the time of publication. We have retained the original publication date on this platform for consistency.

Disclaimer: This analysis is for educational purposes only and should not be considered investment advice. The author may hold positions in stocks mentioned. Always do your own research and consult with a financial advisor before making investment decisions.