The big issue here is how to train something like that. Deepmind, as an AI, needs a set of parameters it can tweak, and a heuristic to judge whether it is doing "better" or worse than another AI. What would this heuristic be?
Furthermore, what data would we train it on? There are a lot of rooms, and chances are most rooms when players start out are super inefficient. And many higher GCL players have very varied setups.
Not to say it is impossible, just in a game like Screeps with a poorly-defined end/goal state, training an AI is hard.