Follow both the money and the accounting and you'll see the part of the capex debate most coverage skips. Spending on servers is the headline; how that spending is depreciated is the quiet number that decides what profit looks like afterward.
When a hyperscaler buys a data center's worth of accelerators, it doesn't expense the whole bill at once. It spreads the cost across the equipment's estimated useful life. A four-year life means roughly a quarter of the cost hits the income statement each year; stretch the assumption and each year's hit shrinks, even though the cash already left.
Alphabet makes this visible. Its filings disclose that it "completed an assessment of the useful lives of our servers" and adjusted them — language preserved in the sec.gov filing, surfaced via EdgarBeast. That's the lever in plain sight: the same physical buildout can look more or less expensive on paper depending on this one estimate.
This is why the capex argument is genuinely hard to settle from the outside. The cash-flow statement shows the spend; the income statement shows a depreciated slice of it; and the useful-life assumption sits between them. The FY2025 report continues to lean on "servers and networking equipment" as infrastructure, per the sec.gov filing — but the disclosures don't isolate an AI-only capital number, so any "AI capex" figure you see is an estimate layered on top of bundled hardware totals.
The honest version of the capex question, then, isn't just "how much are they spending?" It's "over how many years are they choosing to recognize it, and does the revenue arrive before the depreciation does?" The filings give you the assumption; they leave the judgment to you.