The $176 Billion AI Accounting Trick: How Meta, Microsoft, and Google Boost Earnings

Meta added $2.9 billion to its profits last year without selling a single extra ad. They just decided their AI chips would last longer on paper. Alphabet and Microsoft have been doing the same for years. Across the five biggest tech companies, this one accounting choice is worth an estimated $176 billion in extra reported profits through 2028.

The most telling detail: Amazon bought the exact same Nvidia hardware and did the opposite. They decided the chips age out faster, absorbed a $700 million profit hit, and said openly that AI technology moves too fast for 6-year accounting assumptions.

Let's break down what this is quietly doing to the earnings numbers you read every quarter, and which company's balance sheet cracks first when honest depreciation enters the picture.

How Depreciation Assumptions Boost Big Tech Profits

When a company spends money on expensive equipment like a GPU or AI server, it doesn't book the full cost as an expense immediately. Instead, it spreads that cost over the estimated number of years the equipment stays useful. That spreading is called depreciation. 

Think of it like buying a ₹60,000 fridge for your home. You don’t say, “My household expenses this month is ₹60,000,” because the fridge will not be used just this month. It will probably be used for many years. So, if you expect the fridge to last 10 years, you can think of it as costing you:

₹60,000 ÷ 10 years = ₹6,000 per year

That yearly cost is similar to depreciation.

For a company, expensive equipment works the same way. If it buys a machine for ₹10 lakh and expects to use it for 5 years, it does not show the full ₹10 lakh as one year’s expense. It spreads the cost across 5 years, because the machine helps the business earn money over many years.

Here's the part that matters: Change the number of years, and reported profit changes immediately. No extra customers, no price increases, no cost cuts. Just a different assumption about how long a chip stays useful. Make the life longer, expenses go down, profits go up. Amazon's $700 million hit proves this works in reverse too: they made the life shorter, expenses went up, profit went down.

Nvidia’s Chip Cycle vs Big Tech’s Depreciation Math

Nvidia shipped the H100 GPU in 2022. By 2024, it was already eclipsed. Blackwell arrived and delivered roughly 3 times faster AI training performance at the same scale, verified by industry benchmarks. Rubin is arriving in 2026, and based on Nvidia's public roadmap it will make Blackwell look slow. That's a new major architecture every 18 to 24 months, each generation delivering 2 to 3 times the performance of the last.

However, hyperscalers are booking their Nvidia hardware as though it stays competitively useful until 2030 or 2031. Satya Nadella himself told analysts in 2025 that he didn't want to "get stuck with four or five years of depreciation on one generation" because Nvidia's pace of change had accelerated. He said this while Microsoft's accounting still uses six years.

Amazon and Meta Bought Identical Hardware and Reached Opposite Conclusions

Sources: Amazon Q4 2024 investor release, Meta Q4 2024 earnings release, SEC filings, Deep Quarry

Amazon and Meta made opposite decisions in the same month, about identical hardware. The Big Five collectively saved an estimated $18 billion in annual depreciation expenses by stretching server lives from 3–4 years to 6 years over the past few years, per Introl Research. That saved expense went straight into reported profits.

Why Michael Burry Is Warning Investors About This Risk

Michael Burry flagged this in November 2025, accusing hyperscalers of "understating depreciation by extending useful life of assets," calling it "one of the more common frauds of the modern era." His math: Meta, Amazon, Microsoft, Alphabet, and Oracle will collectively understate depreciation by approximately $176 billion between 2026 and 2028.

Nvidia pushed back. Accounting analysts correctly noted this isn't automatically illegal under US accounting rules, which give management some discretion to estimate useful life based on their own operational judgment. 

But Morgan Stanley and Bank of America filed separate concerns. BofA specifically warned that AI depreciation costs are "only just starting" to hit income statements as billions in recent hardware spending converts from construction into operational assets.

Burry may be overstating the fraud framing. But he's likely right about the direction.

Big Tech Stress Test: What 3-Year AI Depreciation Math Does to EPS

Here is what the reported earnings per share (EPS) of each company would look like if they used a 3-year depreciation schedule, consistent with Nvidia's actual product cycle, instead of their current 5 to 6-year assumptions.

Source: SEC 10-K Filings, SEC 10-Q Filings), Internal Calculation | 

Note: Servers and network equipment represent roughly 60 to 70% of each company's total annual depreciation, based on their asset mix disclosures. This is an estimate, since companies don't separately break out server depreciation in their filings.

What this means: Meta's reported profits are roughly 15% higher than they would be if the company used Amazon's logic. At a stock price of around $600, that means investors are paying for earnings that include several dollars per share of accounting benefit that Amazon decided to stop using in January 2025.

And this is only the 2025 picture. Burry's 2028 numbers are worse because the build-out is far from over. The Big Five spent $448 billion collectively on AI infrastructure in 2025 alone, per Epoch AI's analysis of SEC filings. They're planning over $600 billion in 2026. All of that new hardware is sitting on balance sheets waiting to start depreciating. The gap in the table above likely doubles by 2027 just from commitments already made.

Meta, Microsoft, or Google: Whose Balance Sheet Cracks First

The answer is Oracle, and it's not close.

1. Oracle: The company made $12.4 billion profit in FY2025, but its future cloud commitments are much bigger. Its signed but undelivered cloud contracts rose to 638 billion. To fulfil them, Oracle is spending heavily on AI data centres.

That spending is already showing up through depreciation. In the first half of FY2026, depreciation was $3.1 billion, or around $6 billion annualised, and it is rising fast. So if Oracle’s AI data-centre assets lose value faster than expected, reported profits could take a major hit. That is why Michael Burry sees Oracle as one of the biggest risks by 2028.

2. Meta: It stands out among Big Tech because it extended the useful life of its assets in 2025. That means it can spread depreciation over more years, which reduces near-term expenses and supports reported profits. But this happened just before Meta planned $115-135 billion of AI capex for 2026.

So the risk is simple: Meta is buying a lot more AI hardware while assuming that hardware will last longer. If that assumption later proves too optimistic, depreciation could rise sharply and hit profits.

3. Microsoft and Alphabet: Both face the same depreciation risk, but their profit cushion is much stronger. Alphabet earned $132 billion in FY2025. Even if depreciation rises by $10 billion, that would be around 7.5% of annual profit. Meaningful, but not balance-sheet breaking.

Microsoft has the cushion too: $281.7B revenue, $128.5B operating income and $101.8B net income in FY2025. With revenue spread across cloud, enterprise software and PCs, higher AI depreciation would hurt margins, but is unlikely to break the earnings story.

Why AI Depreciation Thesis May Be Too Extreme

Burry is right to question AI depreciation. But calling it fraud may be a stretch. The data points to a more nuanced risk: aggressive accounting, not necessarily false accounting.

1. Older GPUs are still earning money

Burry’s argument assumes AI chips lose economic value quickly because frontier training models move fast. But GPUs do not become useless once they stop being cutting-edge.

A chip can move from frontier training to inference, batch processing and analytics. That still generates revenue. For example, CoreWeave’s 2022 H100 contracts reportedly rebooked at 95% of original pricing when they expired. That suggests the market still valued those chips highly even after multiple years of use.

The key test is not whether the chip is the newest. It is whether it still creates economic benefit. If an older GPU is still running paid inference workloads, it still has accounting value.

2. Free cash flow already shows the pressure

The AI capex hit is not hidden. It is already visible in free cash flow. Amazon’s trailing 12-month free cash flow fell from $26 billion to $1.2 billion. Microsoft’s free cash flow is down 22%, while Alphabet’s is down 38%.

That means investors are already seeing the real cash impact of AI infrastructure spending, even if depreciation has not fully hit EPS yet. J.P. Morgan’s stress test also found that applying a 3-year depreciation life to AI hardware would cut EPS and operating margins by roughly 6-8% for most hyperscalers, with Oracle as the bigger exception. That is material, but not collapse-level.

3. Amazon’s change was not a restatement

Amazon shortening server lives in 2025 was treated as a change in accounting estimate, not a correction of past results. That matters. Under GAAP, estimates change when new information becomes available. A restatement would imply the earlier accounting was wrong at the time.

Amazon’s earlier useful-life assumption had been reviewed annually by its auditor. The 2025 change reflects a shift in hardware mix and AI infrastructure economics, not proof that earlier earnings were fraudulent.

What This Means For Investors

If you are an Indian investor holding Meta, Alphabet, Microsoft, or Oracle through INDmoney, the earnings per share numbers you track are computed with these depreciation assumptions already baked in.

Here's the practical issue. The P/E ratio, the most common way to judge whether a stock is cheap or expensive, is calculated by dividing the stock price by earnings per share. If EPS drops because depreciation is revised upward, the P/E ratio rises and the stock suddenly looks more expensive, even if the price hasn't moved

The earnings moved under you but that's not a reason to sell any of these stocks. 

• Alphabet is still one of the most profitable businesses ever built. 
• Meta's core advertising business is fundamentally sound. 

But when you're deciding between a 25x earnings multiple and a 30x multiple, knowing that the "E" includes several dollars of accounting benefit that Amazon chose to eliminate is relevant to the decision you're actually making.

Bottom Line: AI Growth Is Real, But Earnings Quality Matters

The AI profits are real. The growth is real. But some of what shows up as earnings in these quarterly reports is money that has been moved from future years into the present through an accounting choice, not through a better product or a more efficient operation.

Nvidia will ship Rubin this year. The H100s sitting on these balance sheets will be two chip generations old. The accounting still says they're barely halfway through their useful life. One of those two facts will eventually have to give.