If I were you, I'd look into smart caching, traditional (high and scaling) bandwidth limits, and instead applying tier limits on updates/edits/data pulls. More user friendly, more generous for smaller projects and slightly less open for exploits. In the current model (with no heuristics), you could shut down one of your customers in 5 min just using curl.
Of course you need some sort of traffic restriction for things not to spiral out of control. Putting this restriction on actual MB's transferred is an acceptable safeguard. Cache bandwidth is cheap using the right strategies, so your bounds could be pretty high while keeping it profitable.