If that's the case, shouldn't HTTP/2 have solved that issue? HTTP/2 is multiplexed, so if you make something like 10 HTTP calls, only a single TCP connection would be established. Additionally, according to section 9.1 of the HTTP/2 spec[0], connections are persistent so unless the server closed the connection, an already established connection will be available from when you first loaded the page.
For one, the most major http client library (Axios) doesn't have support for that kind of pipelining.
But also, you don't necessarily know what to pipeline until you get the first request through. Most web apps will make a REST request, look at the contents of the response, then make another REST request. Maybe you could keep the socket open through that, or you could even use websockets to bypass HTTP altogether and use jsonrpc, but still, you don't know what you need to fetch in the second round until you finish the first one, so either way you're looking at a waterfall of requests, one after another. Pipelining them would scarcely help.
What would help is having one API request that basically says, "here's who I am and here's the page the user wants. Give me everything for it." Then you're moving the controller back to the server, and leaving only the view on the client.
Axios doesn't control what protocol is used, the browser and the server will negotitate a suitable protocol and version, which the browser will transparently use for XHR (which Axios use under the hood). Don't confuse multiplexing with pipelining which was introduced in HTTP/1.1. Multiplexing is a much better solution than pipelining to counter head-of-line blocking, since large or slow responses will delay subsequent requests while multiplexing allows parallel response/request communication.
I do agree that waterfall-style requests are more prominent with a REST structure, and that something like GraphQL could solve that - but GraphQL seems largely incompatible with how many develop web apps today; smaller components that request their own data. I'm also unsure how compatible a dynamic query language like GraphQL is with denormalized databases like Cassandra or ScyllaDB where you can't model your data before you've established what queries your site/app will perform. I've yet to see a codebase where the REST waterfall problem is a huge problem though.
Note that I too dislike how slow websites are these days, and much prefer to both build and use an entirely server-side rendered one. None of the SPAs that I know of has been a success in terms of performance; Facebook, Google Mail, Reddit's redesign, Youtube, Twitch - most of them are dreadfully slow and sluggish.
[0]: https://httpwg.org/specs/rfc7540.html#rfc.section.9.1