Isn't a viable horizontal scale solution to have a central allocator, that assigns a new client to a pool instance that then has a direct websocket with the client? So there's no central choke point or "websocket load balancer that all frames flow through"
That's how I've always done it for scalable deployments of my remote browser isolation service.
While it's certainly doable, it's disruptive to horizontally scale websockets servers and depending on how you use it, can result in significant amounts of duplicated data transfer.
From cost and UX perspectives, it's just not worth it. Either use websockets and accept the fact that you'll always be paying for all the capacity you might need or give up and poll a low latency endpoint.