A lot of people use it to promote their art or to share screencaps/gifs/clips of movies, so I think people would get pretty cranky if the image files were overly compressed (unless it's the magic kind of compression that doesn't cost any fidelity. I don't really understand how compression works).

That's without getting into video content.

It's a social network. Tumblr serves a lot of traffic, and has a lot of data. The product data alone is over 1 trillion rows in relational databases by my estimation (former early eng there), and that's not accounting for replication / high availability.

1000 servers actually sounds surprisingly small. It was a fair bit higher than that a decade ago, and that was on bare metal then, vs (afaik) cloud VMs now. Then again, their traffic is a lot lower now, but still nothing to sneeze at.

For the YC audience: Tumblr has about 60B images using 8 petabytes of storage. 205TB of memory, 46,752 CPU threads, 250 kW of energy. About 100k requests a second, a petabyte a day of bandwidth.

So in short, the answer to my question is that the scale is way way larger than I imagined.