It's never been tested in court, but the theory is you with company A run an Affero GPL (modified code) server. It's not accessible to anyone outside company A, except me, with a proxy-ish server in Company B (which may or may not be controlled/related).
Then I'm the only user of your software, and never request the code, even though ten billion users use it via my proxy.
I'm not so sure it's just being cute. Define a bright line difference between this API proxy and something like an IRC server. Fundamentally it's an RPC black box. Client sends a command and gets a response. How that response is generated the client has no way of knowing or differentiating. For the defense to fail, wouldn't you have to prove that end user C was a recipient of software from company A? How could this possibly be true when user C cannot observe even the existence of A?
Like, imagine this hypothetical.
Set up a server. For each unique client that exists, it proxies to one of two backends. One is wrapping an LGPL library. The other is wrapping a cleanroom non-CL implementation.
A given user will always get one or the other. Think of it as being assigned off the high bit of a user hash or something.
Which users have rights under the LGPL? Remember that half of them have never used the (wrapped) LGPL implementation, and NONE of them can tell which group they are in.
I don't actually see the difference as being that distinct from the long accepted fact that a program compiled by a copyleft compiler or running in a copyleft interpreter is not itself inherently copyleft.
Then I'm the only user of your software, and never request the code, even though ten billion users use it via my proxy.