The market is asking the wrong question.
The question asked all season is whether AI capex is a bubble, and the search for an answer happens on the demand side: is the cloud growing, are customers paying, are the orders real.
But demand already answered. Microsoft beat its own growth guide for Azure - 43% against a 39%-40% guide. Google Cloud accelerated to 82%. AWS accelerated from 17.5% to 28%. KLA, Teradyne and Lam Research - three independent equipment makers at three different production stages - guided to acceleration in unison.
Demand exists. That is no longer the argument.
The real argument sits in two numbers that appear in the cash flow statement a few lines apart, and that almost nobody places side by side.
The Mechanism: Why Earnings and Cash Flow Tell Two Different Stories
This is the point everything else follows from, and it is not complicated:
Capital expenditure leaves the bank in full, on the day it is spent. It enters the income statement gradually, as depreciation, over the asset's life.
So inside a build cycle, the two statements contradict each other
The cash flow statement absorbed the cost in full, immediately. Free cash flow collapses.
The income statement absorbed only a fraction. Earnings per share look resilient.
And the gap between them is cash that has left the bank and has not yet reached the profit line.
This is not an accounting trick and nobody is hiding anything - it is simply how depreciation works. But the implication is that at this stage of the cycle, the builders' reported earnings are flattered relative to their cash position - and that gap closes by itself, necessarily, in the years ahead.
Hence the central tool of this analysis: the ratio of capex to depreciation.
- Near 1.0 - steady state. The company replaces assets at the rate they wear out.
- Above 1.0 - the company is building. The gap is a future cost already paid in cash and not yet recognized.
- And the return to 1.0 is mechanical. If capex froze in place, annual depreciation must rise by the gap. That is not a forecast - it is the arithmetic of the depreciation schedule.
The Data
Four companies, each from the cash flow statement in its own results release. In millions of dollars.
| Company | Quarter | Purchases of property and equipment | Depreciation and amortization | Ratio | Gap in the quarter |
|---|---|---|---|---|---|
| Amazon | March 2026 | 44,203 | 18,945 | 2.33 | 25,258 |
| Microsoft | June 2026 | 35,802 | 11,022 | 3.25 | 24,780 |
| Meta | June 2026 | 30,116 | 6,356 | 4.74 | 23,760 |
| Alphabet | June 2026 | 44,924 | 7,104 | 6.32 | 37,820 |
The combined gap, one quarter each: about $111.6 billion.
And what matters just as much - the rate at which each side is growing:
| Company | Capex, year-over-year | Depreciation, year-over-year |
|---|---|---|
| Amazon | +77% | +33% |
| Microsoft | +110% | +18% |
| Meta | +82% | +46% |
| Alphabet | +100% | +42% |
Capex is growing at two to six times the rate of depreciation. The gap is not narrowing - it is widening.
A methodological caveat that belongs here, not in a footnote
The depreciation line is not defined identically across the four, so the ratios are not precisely comparable:
Amazon reports "depreciation and amortization of property and equipment and capitalized content costs, operating lease assets, and other" - the broadest definition. Its denominator is inflated, which makes its 2.33 ratio a floor, not a ceiling.
Alphabet reports "depreciation of property and equipment" only - the narrowest definition. Its 6.32 is the most upward-biased of the four and is not directly comparable to the other three.
Microsoft and Meta sit in between.
What the definition does not affect: the direction. Every one of the four is spending multiples of what it depreciates. The exact multiple depends on the definition; the fact does not.
Finding One: The Size of the Bill Waiting in Line
If capex froze at today's level, annual depreciation must rise by the gap. Set that against current operating income, annualized, and the true order of magnitude appears:
| Company | Gap, annualized | Operating income, annualized | Gap as % of profit |
|---|---|---|---|
| Meta | about $95.0 billion | about $75.1 billion | 127% |
| Amazon | about $101.0 billion | about $95.4 billion | 106% |
| Microsoft | about $99.1 billion | about $162.4 billion | 61% |
At Meta and at Amazon, the bill waiting in line is larger than the entire annual operating profit.
The annualized figures are the quarter multiplied by four - a rough approximation that assumes the quarter is representative. This is an order-of-magnitude presentation, not a forecast.
Which is why revenue growth is not a nice-to-have - it is a structural requirement. Revenue has to grow faster than a depreciation schedule that is already set, otherwise the margin erodes even if nothing in the business breaks.
Finding Two: The Metric That Called Who Would Lose Their Margin
Here the analysis turns from description into a working tool. The metric: revenue growth minus depreciation growth.
| Company | Revenue | Depreciation | Spread |
|---|---|---|---|
| Microsoft | +18.0% | +18.3% | -0.3 points |
| Amazon | +16.6% | +32.8% | -16.2 points |
| Alphabet | +24.2% | +42.1% | -17.9 points |
| Meta | +28.0% | +46.4% | -18.4 points |
And now the check that turns this from a model into a verified observation:
The most negative spread in the table belongs to Meta - minus 18.4 points. And Meta's operating margin this quarter actually fell from 43% to 31% - twelve points.
The only balanced spread belongs to Microsoft - minus 0.3 points. And Microsoft's operating margin held.
The metric does not merely describe after the fact. It pointed correctly at both the company whose margin would collapse and the one that would hold - from a ratio computable off the cash flow statement before you look at the margin at all.
Finding Three, and the Important One: Not All Revenue Was Built Equal
The table above uses total revenue. But this capex was not built for total revenue - it was built for the cloud. Swap the numerator and the picture inverts completely:
| Company | Cloud growth | Depreciation | Spread |
|---|---|---|---|
| Alphabet (Google Cloud) | +82% | +42.1% | +39.9 points |
| Microsoft (Azure) | +43% | +18.3% | +24.7 points |
| Amazon (AWS) | +28% | +32.8% | -4.8 points |
| Meta | no cloud line | +46.4% | cannot be computed |
Two conclusions, both sharp.
A. AWS is the only one of the three clouds growing slower than its own depreciation
And this is not a single-quarter artifact. A second check, on a twelve-month basis - the more reliable base, less exposed to seasonality:
Amazon, twelve months ended March 31, 2026
· Purchases of property and equipment: $151.0 billion · Depreciation and amortization: $70.4 billion · Ratio: 2.14. Gap: $80.6 billion a year.
And depreciation growth over that period: 27.2%. Against AWS growth of 28%.
Those eight tenths of a percentage point are the entire margin Amazon has between the new revenue it generates and the new cost it has already committed to.
B. Meta has no numerator at all
It has no revenue line that shows the return on the investment. Its return is embedded in the price of an ad, which rose 12% in the quarter - while its depreciation rose 46%.
That does not necessarily make it a bad investment. It makes it an investment that cannot be measured from the outside. And when it cannot be measured, the market prices what is visible: the cost.
What This Does to the Bubble Question
This is not a demand bubble. Demand is verified across three clouds, three equipment makers, and one company beating its own guide.
Nor is it merely "investment in the future". It has an accounting timetable that is already set, and that timetable is the size of annual operating income at two of the four.
The right question, and this is the whole analysis:
The question, stated measurably
Is cloud revenue compounding faster than the depreciation schedule already locked in?
At Alphabet and Microsoft - yes, comfortably. At Amazon - currently no, within a margin of error under one percentage point. At Meta - unknowable from the outside.
And that is a measurable question, not a matter of belief.
Three Metrics to Track - and Three Things That Would Refute This
What I will follow every quarter:
- The capex-to-depreciation ratio. As long as it is above 2, the bill is still waiting. Convergence toward 1 is the end of the build - and the start of full absorption into earnings.
- The spread between cloud growth and depreciation growth. This is the metric that pointed correctly at Meta and at Microsoft.
- Free cash flow against net income. When they diverge, the gap is the investment. When they converge again - the build is over.
And what would refute what is written here:
- A longer asset life than assumed. If these assets are depreciated over 10-15
years rather than 5-6, the bill spreads thinner and the annual impact shrinks
accordingly. **Depreciation policy is disclosed in the full financial statements
- that is where to check, not in the results release.**
- A sharp acceleration in cloud revenue. If AWS returns to growth above 35%, the spread against depreciation reopens and the problem resolves itself.
- A capex slowdown. If the pace slows, the gap stops widening - but what has already been spent gets absorbed either way.
The debate in one line
The bulls see demand verified across three clouds, two clouds growing comfortably above their depreciation schedules, and a capex cycle that creates a decade-long operating asset. The bears see a $111 billion quarterly gap between cash out and cost recognized, a future bill the size of annual operating profit at two of the four, and one cloud already growing slower than its own depreciation. Both sides are reading the same filings.
My Angle
A personal opinion of Ilan Abramov - not advice, not a recommendation
What catches me in this work is that the question everyone is asking has already been answered, and the question that decides is barely being asked.
"Is there demand for AI" was a good question a year ago. Today it has been answered by three clouds, three equipment makers, and one company that beat its own guide. Anyone still asking it is looking backward.
The question that decides is one of pace, not of existence. The investment has been made, the accounting bill is already set, and the depreciation schedule does not care how we feel about it. The only thing that can outrun it is revenue.
And what surprised me in the numbers: I expected the gap to be large. I did not expect it to be the size of the entire annual operating profit at two of the four. Nor did I expect this simple metric - revenue minus depreciation - to point with that precision at the company whose margin collapsed. When a simple model explains an outcome that has already happened, it is worth respecting what it says about outcomes that have not.
And the caution I hold on myself: the depreciation definitions differ across these companies, and comparing the ratios is problematic - I put that in the body of the analysis rather than a footnote, because it is a material caveat. Also: Amazon's figures here are from the March quarter. It reports the June quarter tonight - and within hours we will know whether that 28% accelerated.
And what I will watch first: the spread between AWS and Amazon's depreciation. Right now it is eight tenths of a percentage point. That is an unusually thin gap between new revenue and new cost at a company of this size - and in cloud, eight tenths of a point is one quarter.
Summary
We pulled two numbers from the four giants' cash flow statements that sit a few lines apart and that almost nobody places side by side: purchases of property and equipment, and depreciation.
The combined gap in a single quarter: about $111.6 billion - cash paid out that has not yet been charged to the income statement. The ratios: Amazon 2.33, Microsoft 3.25, Meta 4.74, Alphabet 6.32 - where a steady state sits near 1.
And if capex froze, depreciation must rise by the gap - about 127% of annualized operating income at Meta, about 106% at Amazon, about 61% at Microsoft.
And the metric derived from it - revenue minus depreciation - pointed correctly: the most negative was Meta, whose margin fell 12 points; the balanced one was Microsoft, whose margin held. Swap in cloud revenue: Alphabet +39.9 points, Microsoft +24.7, and AWS -4.8 - the only cloud growing slower than its own depreciation.
The question for the investor is not whether demand exists - it does, and that is no longer the argument. The question is whether revenue compounds faster than a depreciation schedule already locked in. Amazon reports tonight, and that is the first test.
Sources: the cash flow statements in the official results releases as filed with the U.S. Securities and Exchange Commission - Amazon.com (Form 8-K, April 29, 2026), Microsoft (8-K, July 29, 2026), Meta Platforms (8-K, July 29, 2026) and Alphabet (8-K, second quarter 2026), including purchases of property and equipment, depreciation and amortization, cash from operations and operating income; cloud growth rates as reported by the companies. Microsoft does not disclose Azure revenue in dollars, so only its growth rate is used. Annualized figures are the quarter multiplied by four and are an order-of-magnitude approximation. Amazon's figures are from the March 2026 quarter - the most recent reported at the time of writing. Data accurate as of the time of writing. Charts are shown in real time via TradingView. Nothing herein constitutes a forecast, recommendation or advice - see the full disclaimer at the bottom of the page.