Alphabet just delivered the biggest quarter in its history and the stock fell 7% the next morning. Nothing in the revenue line explained it. What explained it sat two statements further down: for the first time since its 2004 IPO, Google’s parent generated negative free cash flow.
That sell-off — which dragged the S&P 500 down 1.20% to 7,409 and the Nasdaq down 2.15% to 25,138 on 23 July — is being written up as a story about AI spending getting too big. It isn’t. The size of the spending is public, guided, and repeated on every earnings call. The number nobody is arguing about, and the one that quietly decides how much of that spending shows up as profit, is AI capex depreciation.
The number everyone is arguing about is the wrong number
The big five hyperscalers spent roughly $412 billion on capital projects in 2025. Guidance for 2026 now sums to somewhere between $700 billion and $760 billion: Amazon around $200 billion, Alphabet $195–205 billion, Meta $125–145 billion, Microsoft $110–120 billion. Roughly three-quarters of it is AI infrastructure. We have already walked through who actually gets paid out of that bill, and the second-order version of it that shows up in your electricity account.
Here is the part that gets skipped. Capex is not an expense. It is a cash outflow that becomes an expense slowly, over an assumed life. That assumption — how many years a server earns its keep — is where reported profit is manufactured.
What “useful life” actually means for your portfolio
Buy a $30,000 GPU rack. On a four-year schedule, $7,500 a year hits the income statement. Stretch the assumption to six years and it is $5,000 a year. Same machine, same cash out the door, $2,500 a year more reported operating profit. Now multiply that by hundreds of billions of dollars of hardware.
Between 2022 and 2024, Microsoft, Google, Meta and Oracle each extended the assumed useful life of server and networking equipment from roughly four years to five or six. The disclosed first-year benefit across the group was north of $10 billion in additional operating income. Nobody had to sell one extra ad or one extra seat of software to earn it.
One company went the other way. In February 2025, Amazon shortened the useful life of a subset of its servers and networking gear from six years back to five, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” When one of the five largest buyers of AI hardware on earth tells you the kit ages faster than the books assume, that is not a footnote. That is a disagreement about physics, settled in accounting.
$46 billion of 2026 profit is an assumption, not a result
Run the two numbers side by side and the gap is hard to unsee. Against roughly $760 billion of 2026 capex, the AI-related depreciation and amortisation actually recognised in 2026 is about $211 billion. More than half a trillion dollars of spending sits on the balance sheet waiting its turn at the income statement.
Estimates built on the hyperscalers’ own 10-K disclosures put the depreciation suppressed by longer useful-life assumptions — measured against a four-year counterfactual — at roughly $46 billion in 2026, $75 billion in 2027 and $107 billion in 2028. That is well over $200 billion of reported pre-tax income across three years that exists because of a scheduling choice made in a conference room in 2023.
To be precise about the accusation, because it matters: this is legal, disclosed, audited and entirely defensible. It is not fraud and it is not even unusual. But it is a judgement, and the hardware cadence is arguing with it. Hopper gave way to Blackwell gave way to Rubin faster than a six-year schedule contemplates. A 2022 GPU still switches on. The question is whether it earns its rent when the rack beside it does the same work for a fraction of the power.
The financing tell: 9% to 32%
The second thing that changed while everyone stared at the capex headline is where the money comes from. Incremental annual debt has gone from 9% of capex in FY24 to 32% on a trailing-twelve-month basis by mid-2026. Alphabet priced an $84.75 billion equity raise in June. Oracle has flagged roughly $40 billion of combined debt and equity for FY27. Gross issuance topped $100 billion in 2025, and forecasts of the multi-year pipeline run toward $1.5 trillion. The Bank for International Settlements has separately flagged the growth of off-balance-sheet joint ventures and special-purpose vehicles funded with private credit.
Translated: the phase in which AI was funded out of the best cash flows in corporate history is over. It is now funded partly out of borrowed money and partly out of a depreciation schedule. Both are claims on future earnings. Neither shows up in the EPS number that leads the headline.
The contrarian read: this is not a bubble call
First, nothing here says the build is wasted. Data centres, transmission, cooling, fabs and networks are real assets doing real work, and the demand for compute has not yet blinked. The honest position is uncomfortable but simple: the capex is probably productive, and the reported profit is probably flattered. Both can be true at once, and most of the commentary refuses to hold both.
Second, the market has already half-worked this out. The tell is behavioural, not analytical. Investors sold a record quarter. When a stock drops 7% on excellent EPS and an ugly cash-flow statement, the market has quietly moved its trust from the income statement to the cash statement. That is a healthy repricing — and it is why “own the hyperscalers because their earnings keep beating” is now the crowded, lowest-quality-of-earnings version of the AI trade.
The cleaner expression is the part of the chain that gets paid in cash regardless of what the buyer assumes about useful life: the chipmakers, the memory oligopoly, the power, cooling and grid names, the fabs. Their revenue is the buyers’ capex. It does not need a schedule to be real. That pattern — the money moving before the headline catches up — is the same one we saw when oil crossed $100 and the bond market shrugged.
What to actually do this week
- Read the cash-flow statement first. Microsoft and Meta report on 29 July, Apple and Amazon on 30 July, straight into an FOMC decision on 28–29. Free cash flow and capex guidance will move those stocks more than EPS will.
- Find the useful-life footnote. It is in every 10-K under Property and Equipment. It takes two minutes and it tells you more about earnings quality than any analyst note you will read this quarter.
- Check what you already own. If you hold a total-market or S&P 500 index fund, roughly 40% of it sits in the top ten names, and most of them are on both sides of this trade. You did not choose the AI trade. You inherited it. That is not a reason to sell — it is a reason to know your position size.
- Do not confuse a re-rating with a collapse. A market that starts discounting reported profit is a market repricing risk, not one falling apart. The people who get hurt are the ones who sized their portfolio as if 2024’s earnings quality still applied.
The bottom line
Record earnings in the AI era are partly an engineering achievement and partly a decision about how many years a graphics card lasts. The spending is visible, loud and endlessly debated. The assumption that converts it into profit is a single sentence in a footnote — and it is currently worth about $46 billion a year to the companies that dominate your index fund.
Wealth is built by people who read the second sentence. This quarter, the second sentence is in the depreciation policy.
Want to track your own path to financial independence? The Wealtharian Wealth Tracker lets you monitor your net worth, FU money progress, and investment milestones in one place. Try it free →