At Google it's commonplace for a single engineer to be able to run a MapReduce over the entire internet. Or do experiments with training huge deep learning models.

The fact that approval is not required (aka 20% project) to use staggering levels of compute is the advantage I'm talking about.