← Noteseconomics
Read hyperscaler capex like a model metric
This week Alphabet, Microsoft and Amazon reported their third quarter, and the number everyone quotes went up again. Across the four biggest spenders, 2026 capital expenditure guidance now adds up to somewhere between $720 and $745 billion (TMT Finance).
I build things on top of that infrastructure, so I care about the number. But I’ve learned to distrust a single headline figure. In my thesis the rule was simple: report each operating point on its own, because a sensitivity from one setting next to a specificity from another makes everything look better than it is. Capex headlines have the same problem.
- Amazon220
- Alphabet195–205
- Microsoft~170
- Meta130–145
Table
| Item | US$ billions |
|---|---|
| Amazon | 220 |
| Alphabet | 195–205 |
| Microsoft | ~170 |
| Meta | 130–145 |
Microsoft as of 29 Oct. Amazon and Alphabet reaffirmed in Oct. Meta as of Aug. Ranges drawn lighter · Source: company guidance via CNBC and TMT Finance
The number moves without the spending moving
- Microsoft cut its calendar-2026 outlook to about $170 billion (CNBC). Part of that “cut” is accounting. It stretched the useful life of its datacentre buildings from 15 to 25 years and some future leases left the capex line. Its CFO said the spending plans “remain unchanged” in economic terms (TMT Finance). The quarter itself came in under target because third-party capacity arrived late, which is a supply problem, not a demand one.
- Amazon held its forecast at about $220 billion. It raised it from $200 billion in the summer, and most of that raise was the price of memory, not more buildings (CNBC).
- Alphabet kept $195–205 billion for 2026 and said spending will “increase significantly” in 2027. The stock fell more than 10% anyway (CNBC).
So the same physical build-out can show up as a cut, a raise or a hold depending on component prices and depreciation schedules. That’s a threshold choice, not a change in the model.
What I actually watch
Cash, because it’s harder to reclassify. Amazon reported negative free cash flow of $42.4 billion (CNBC). By the summer, Alphabet’s debt had gone from about $16 billion to about $100 billion in a year, and its free cash flow had turned negative for the first time since its 2004 IPO (TMT Finance). Those are the figures that tell you how convinced these companies are.
And payback. Andy Jassy says Amazon’s servers and networking gear pay for themselves in under three years, against a five-to-six-year useful life (same source). If that holds, the depreciation fears are overdone. If utilisation drops, it’s the first assumption to break.
Why it matters to someone building applications
My bet, and it is only a bet, is that this much capacity makes raw compute the least scarce thing in most AI projects I’ll work on. What stays scarce is the work around the model: clean data, an evaluation that matches the real task, and an interface people trust. That’s the part I like, and none of it gets cheaper when a datacentre comes online.
Figures as of 30 October 2026, from the linked reports. Company guidance changes every quarter, so check the latest filings before quoting any of this.