Have you considered webtorrent or IPFS for distributing the file? Webtorrent in particular looks very easy to setup, and then the most your server would ever have to handle is the equivalent of a single downloader.

I think it would be better to stream only needed data, using some lazy load mechanism. Still need to figure out what is the best approach

That's definitely the more permanent solution for games like this. I was thinking more along the lines of quickly solving the immediate problem so you don't have to spend more money on a different server. I wouldn't be surprised if you could replace the download with a webtorrent in less than an hour of work and be done with it.

I wonder if you could utilize IPFS as a distributed CDN for p2p file downloading to take some load off the server (assuming the .pk4 is static and stateless)

Excellent job! I wanted to run the native copy for comparison, but it turns out you added support for the demo files yourself! Gah!

What would you guess you could get the bundle size down to if you were to preload only the required assets? I think a lot of us are curious as to how load-time performant webassembly games like this could eventually become.

Swapping the pakfile for a WebTorrent (or maybe a couple of them) wouldn't be crazy, but I guess there would still be some trouble in keeping that from holding things up.