I think this discussion was about switching a pooled model to one where a user's subscription payment is directly split according to that user's listens. That requires unaggregated and unanonymized listen data.

Ah true.

This is solved (well, not solved, but handled) by not having a direct link between usage data and the user. So you store your data as <usage, proxy_id> which references <proxy_id, user_id> which references <user_id, user PII>. This way you can always easily break the connection to PII (if the user requests deletion) while keeping per user data.

But true, it's not anonymized in this case.

Even though they just need a total plays sum to pay out to artists, Spotify for sure stores every play by person. You can request your data from them and see for yourself - it's of course necessary for their recommendation services, and even just providing the listening history back to you is a service many people want badly (it is the whole point of services like last.fm!)