MTBF is not exactly the right value for predicting the number of failures of a certain amount of units. You'll need to figure the actual probable failure rate to figure failures of parts per million.
MTBF will allow you to predict the probability of a part working based on a certain time. R(T)=exp(-T/MTBF) T is the time you're interested in and R(T) is the probability the part will work over that course of time.
What is your physics of failure, is exponential the correct distribution? I agree with the Kyle, you will need to calculate reliability at a given time to get ppm.