Actually I have $0 hardware/automation costs: in always free Oracle Cloud I have small flask server receiving webhooks and triggering swarm of Kaggle notebooks (max 10 concurrently running, each 4 CPU/30GB RAM) doing training and submitting.
That may explain my strange way of looking at Numerai. Additionally, I do not believe that size or method of model can give you superior performance (I would need some evidence to change my mind) because input data are extremely noisy. Also signal (if present) is ever changing. Simply put: it may be just a random walk down Wall Street.