I worked at Demon Internet in the late 1990s, and our “homepages” service ran on thttpd. My colleagues had hacked an earlier version to support mass virtual hosting - there were tens of thousands of homepages web sites at a time when that was a lot.

On the front end we had a few cacheing reverse proxies running Squid on FreeBSD, like an early CDN (we had nodes in London and New York). The FreeBSD kernel was hacked to support configuring IP addresses by the CIDR block instead of individually. When Homepages was set up, the HTTP Host: header was not yet well supported, so it needed an IP address per web site. The Squid boxes had 96k addresses each (a /16 and two /18s) and the network was configured to spread the load across them.

Squid was modified to translate the server IP address into a Host header as necessary, so the back end did not need many IP addresses. There was one box running thttpd, one box running an FTP server for uploads, and a NetApp NFS storage server. There were hacks in thttpd to shard the web site names across multiple directories to avoid performance problems with very large directories.

We made some effort to get the modifications upstream, tho thttpd had diverged a lot by the time we sent our code to Jef. But, IIRC thttpd’s vhost setup has some resemblance to the way Demon did things. (This is also why Apache httpd mod_vhost_alias is so weird.) And these days, firewalls can be configured to do the CIDR IP address trick, so there is no need for kernel hacking.

(edited to add) Here's an old blog post on the topic, with the links I left out of this comment https://dotat.at/@/2012-09-25-large-scale-ip-based-virtual-h...