[mjevans]

Back of the napkin algorithm:

Borrow what DHCP does for leases, but with payment quotas.

First layer is a burst quota, up to that maximum in any spike. This only regenerates when the (monthly?) average quota has enough slack at current that an additional burst can be harvested without the remaining average being under the initial set value. (The math would probably be different, remaining period quota less burst average over time greater than period quota average, but that's the sales description of the concept.)

The second layer is the monthly (or some other period) average for the maximum expenditure allowed.

A billing endpoint would maintain a fractional bucket of spending (divided up as makes sense) but in the case of a single quota consumer would receive an estimate for the period (ideally the burst) and allow up to half of that to be used. At that point it will 'renew' the quota (lease), and flush billing, including sending alerts. If there isn't any further quota the remaining released balance would be consumed and then requests fail.