The NSA figured this out in the 90s -- you filter based on metadata and then retrieve the corresponding data if necessary. Aside from the fact that this (in the NSA's view) allows them to sidestep the 4th amendment, it's usually much more effective than sifting through billions of records every day. That same system lives on today with XKeyScore (or whatever they've replaced it with in the past 7 years).

Hard drives are pretty cheap, particularly for a government. Store it all now, target your analysis narrowly later at your leisure.

Do you know how much data that would have to be? Scaling that seems improbable.

The NSA has built a data center in Utah specifically for this problem[1], so it's hardly beyond the realm of plausibility.

[1]: https://en.wikipedia.org/wiki/Utah_Data_Center

Year-old data isn't worth very much.

There's no reason to think they'd have to wait a year. High value targets, like SpaceX had they not banned the use of Zoom, would obviously receive priority treatment by the Chinese intelligence community. My point here is that the analysis doesn't need to be done in real time, they could store the data and review it a few hours later, or whenever they wanted.

(For that matter, there is certainly a lot of data that would be useful a year later. Some data could be valuable even many years later. Taking SpaceX as an example, it should be obvious that old data could be valuable.)