"Since its always going to take some amount of time to start a compute, we decided to (mostly) accept that the ongoing cost of keeping compute starts consistently fast across our entire fleet wasn’t something that we could do reliably: so we didn’t. Instead we’re (basically) always starting computes before anyone asks for them, and then when we get a request for a compute, we just take an “empty” compute, give it some configuration and… that’s it? You have a compute in a few hundred milliseconds."
That's not a cold start anymore. You just made a warm instance pool. What a waste of time this was.