The heuristic that when total cost is a sum of competing terms with power-law shapes, the minimum tends to sit where the terms are roughly equal — surprisingly load-bearing for back-of-envelope reasoning about training vs. inference.
Tensions
It is a rule of thumb, not a theorem, and "equal" can mean 50/50 or 33/33/33 depending on how many cost terms you split into. The looseness is fine because the underlying inputs (traffic forecasts, model lifetimes) carry order-of-magnitude error bars themselves.