If your read the Levitt study, you will see that he took a different approach to the problem than just counting up cases. He worked out the likelihood of causing a crash instead with the only assumption being a binomial distribution.
As you point out, most crashes are not caused by drunk drivers, that's because the overwhelming majority of drivers on the road are sober. Even people who drive drunk are typically sober most of the time. So drunk drivers are very rare, but they cause an outsized share of the fatal accidents.
In calculating their harm to society, he excluded fatalities when the drunk driver killed themselves or a passenger in their own car, noting that these people have presumably accepted the risk of death. He counted up the remaining fatalities, only cases where the drunk killed someone in the other vehicle.
From this, and standard insurance rates for human life, he worked out that it costs society about ten grand every time, and that's in the neighborhood of the average car.