For example the company servers could be hosted on an island with armed guards instructed to burn everything if anyone approaches and the decryption happens only on those servers: sure they have access by definition, but they really don't.
The guards could decide they’re not getting paid enough and steal the data. Or the government could arrest them. Or the government could MITM the data center. Or any hundreds of different scenarios.
At the end of the day, the only thing preventing somebody from accessing the data is that they just… don’t.
This is very weak security and it is why cryptographers and security professionals call it “effectively plaintext.”
I mean, having to modify server code in order to access data that is "effectively plaintext" is not so different from installing a backdoor inside the client: it's not like the user has any choice of client, so even for apps like whatsapp and signal that run E2EE one is still making a leap of faith.
If we add the fact that everything runs inside an os built by companies who may or may not be constantly spying on their users we could say that by definition there's a lot of stuff in our lives that lives in "effective plaintext".
I also want to repeat the original point that started this whole conversation: the point was how easy it would be for Telegram to access the chats and if the justice system can compel them to do so.
When people say it has the data in plaintext, I take as a "they can access them whenever the want right now without changes", and yes of course the could ultimately access the data (in fact they don't claim to be unable to). What they claim (and I believe it feasible) is that even if a judge seized all the assets and servers under his/her jurisdiction it would be impossible to decrypt any user data.
The point is: even if they could, should they do so when compelled by authority?