Smaller networks is one way to go indeed. Have a similar architecture but smaller. Much smaller such that you can have a result in ~1h. Then you can just distribute the process using weights and biases or another similar framework.