> Is still Amazon operating those locations, so I assume it's still the USA who's calling the shots. Please do prove me wrong if I am, this is somewhat of an assumption (even if I am fairly confident it works this way in practice).
The whole point of Amazon operating datacenters in multiple regions is that, apart from improving availability of the stuff they host, datacenters outside the US get to be legally independent from US law. (At least at the consulting firm I work for we strongly advise clients to follow GDPR and use the European region only.) My guess would therefore be that European datacenters, for instance, are operated by Amazon EU S.à r.l.
> Domain fronting: didn't Amazon and Google say they were not going to do that anymore, because they didn't want to stand up for the organisations using it at the time? Some countries wanted to block certain services (was it sci-hub? TPB? Tor? I don't remember) and instead of standing up for them, they just banned domain fronting.
You might be right, I'm not sure what the state of domain fronting is, either. In any case, the fact alone that inside an AWS datacenter Signal can operate an almost arbitrary number of servers (as opposed to a single one) without anyone outside the datacenter being able to tell which machine a packet gets routed to or originates from, means that the Signal servers inside the datacenter can potentially handle millions of messages per second which would make correlating them with one another anything but easy. Remember that messages get padded anyway and that there might be additional random delays, depending on which server in a datacenter your message gets routed to. On top of that, judging from the Distributed Systems Developer position Signal advertizes on their website[0] I would assume that Signal's servers need to communicate with one another, too, i.e. your message might go to a European datacenter first and then, from there, go to one on the other side of the world – which would further complicate tracing it. Finally, message notifications on phones usually use Firebase Cloud Messaging (or the Apple equivalent) which adds yet another layer of indirection, delays and mixing[1] and, thus, obfuscation.
All in all, I am not sure, therefore, if I agree with your assessment that
> it should only be a very manageable number of packets, since we're only interested in the routing header and packet size.
as the number and overall scenario don't seem that trivial to me.
> the attack scenario was more about legal interception (which you put in front of the server rather than in front of the datacenter)
Fair enough, if you tap every single server's network cable, it certainly seems more doable. But how do you, as the NSA, carry that out in practice? Knock at the door of that European datacenter and hope employees won't say anything to the press? I'm not saying it's impossible (it's certainly not) but Amazon EU S.à r.l. being a separate legal entity and employees not being bound to US law would certainly be big operational challenges. I'd say it's much more likely that the NSA's European intelligence agencies carry out such a task for them (or provide support) and even then it's not as easy as in the U.S. (no Patriot Act or anything similar).
Finally, don't forget that AWS instances can be deployed within a minute, meaning that "putting something in front of a specific server" becomes a lot less trivial if you don't really know which server it is or new servers get deployed every other day. (For once, an advantage of The Cloud™ haha.) You basically have no other choice than to put your wiretapping device in front of the entire datacenter.
> thus Matrix is king in this regard [legal interception]
I'm not sure it is. There are pros and cons to both centralized and federated networks here: A central big provider might have the legal power and sufficiently deep pockets to successfully fend off a "request" by an intelligence agency. Small ones like in the federated case do not. Sure, it takes longer to subvert a federated network. But once done, the traffic analysis is a lot easier compared to the centralized scenario (as we already agreed).
> Sure, once you know who is running a server
I don't think it's that hard to find out who's running a server (in Europe) when you have its IP address. Almost everyone who operates a server for private purposes ties the server to a domain or at least a bank account. (Remember that the EU forwards financial data to US agencies.)
> But if there is no single central service, you need to install a lot of taps or capture the right internet backbones to get close to the same information.
The NSA already does that! (Compare what's been reported e.g. about NSA's Tailored Access Operations (TAO) team.) My general assumption, therefore, is always that the NSA and its partner agencies are sitting at every major internet backbone.
> And if you're serious about anonymity, if you're hiding from the police or an intelligence agency, then surely you'd host that server somewhere paid for without traces to your real name. Or use some public home server -- they still need to tap that specific home server rather than a centralized server.
I agree with the first sentence but I don't understand the second. Practically all traffic to and from a public server you're running at home gets routed through an internet backbone anyway. (Unless, maybe, your Matrix contacts are all in your neighborhood and are all with the same ISP.)
[0]: https://signal.org/workworkwork/
[1]: https://en.wikipedia.org/wiki/Mix_network