Content is all ephemeral on some time scale, but you can cache the near-term content to maximize the views and cut back on compute costs. Some model or human made it (the cost), it's trending (the value), so keep it around for a bit.

Everything has a relevancy and penetration decay curve.

The funny thing is, I think this law applied in the classical era (1950's, 1990's, etc.), we just weren't creating at scale to realize it.

Maybe it's just one dominant variable: novelty. I'd be curious to see how we might model this.