I'm going to go against the grain here. If this were my project I would try to spin up render machines dynamically in response to load. (Similarly to how gitlab CI runners can be spun up automatically, for instance.) There is no need to replace a working technology stack to make it faster. 1 minute is a reasonable wait time, and if your wait time goes over 5 minutes, then spin up some more workers.

Plus we have to be real and say, is the only reason the queue so long, because you let a bunch of nerds on HN add to the queue for free? I'm guessing the answer is, "probably." No need to engineer a fix for a problem that will typically not exist.