We agree that huge capex can be good if the returns on capital being employed are good. How can you or anyone else for that matter tell what the returns on at least $500 Billion (likely significantly more) of AI data center capex will be with any degree of confidence, taking into depreciation schedules.
To me it seems like almost all GOOG fans just take a deep breath & hope that Sundar & Ruth know what they are doing...

I certainly agree that this is a key question, but I don't think it's an existential one for the Alphabet investment case.

Even if the profitability isn't there, I think they are (a) smart enough to stop doing it at some point if/when it becomes clear that it's not a profitable way to go, and (b) they make so much money elsewhere that they can earn their way out of that one time loss in a reasonable time frame. It won't kill them. The share price might be low for a while, but they'll come out of it making money.

My feeling is that there are certainly some data centres being built that won't be needed for their currently intended purposes. Some mix of reasons like technological change (there will be a better way to do that task long before the depreciation schedule runs out), the sponsor not doing that any more (maybe one or two too many firms on the data centre bandwagon, so some may fold their hands), and/or the tasks moving locally to the users that are currently assumed to need data centre services.

It's the "local use" trend that is my pick for the biggest threat to the "infinite data centres" craze: even now, you can run a dang fine model on hardware you can buy within a few months, either RTX Spark or Mac Studio M5 Ultra, and presumably more and better boxes to come soon. I think a *lot* of users will want to spend the money to have the local control, since it's not really all that expensive even now. It's a lot cheaper, in fact, so the only question is whether it's good enough. For a nice little write-up on the feasibility of doing things locally, I strongly recommend reading this little article (even though perhaps a little strong of his conclusions)...
"If This is True, the Hyperscalers are Toast" klementoninvesting.substack.com - If this is true the hyperscalers
...which is reporting on this study
"Intelligence per Watt: Measuring Intelligence Efficiency of Local AI" arXiv: [2511.07885] Intelligence per Watt: Measuring Intelligence Efficiency of Local AI
A possible precis: locally run models are both cheaper and better than you thought and are plainly good enough for most things already, and the gap will keep narrowing.

Jim