I don't mean to support or refuse the author's main points or analysis, but you might like to know that the Chrome team is currently working towards shipping the Topics API. I have strong opinions about it but I will try not to editorialize.
My high-level understanding is that they're going to run an ML model over your browsing history (locally on your device) to build a list of "topics" that you care about. Sites you browse can use the Topics API to pull a set of these interests from the browser to show you "relevant" ads. Mozilla has taken a negative position against this standard.
In my experience with modest specifications, you can source servers from tier 1 vendors (Dell, HPE, etc) with 5 years of support for less than a 1 year commitment for the roughly equivalent AWS or GCE instances (without prepayment).
The monthly opex to house and power and cool those servers isn't _negligible_, but if you're doing back of napkin math comparing MRCs to Cloud you can just deduct the costs from their bandwidth charges that have been marked up 10,000%
Aside, your second point is incorrect. The SteamDB folks have a public write up on analysis of the preload system: https://steamdb.info/blog/steam-download-system/