We use variation in nighttime light across the lunar cycle as a novel, plausibly exogenous source of variation in sleep to study how sleep loss affects fatal traffic accidents. Linking 1.8 million fatal crashes (FARS, 1975–2023) to the timing of 606 full moons, along with nearly 3 million sleep responses from the BRFSS and the American Time Use Survey, we show that sleep worsens and fatal crashes rise in the days after the full moon. Daily traffic fatalities increase by 0.65% in the four days after the full moon overall, and by 1.78% in clear-sky states during non-winter months, when clouds are less likely to block moonlight. These effects vanish during total lunar eclipses and are concentrated in daytime hours, when sleep loss can impair driving but moonlight cannot affect visibility. We estimate that the lunar cycle causes 35–38 traffic deaths annually, costing about half a billion dollars per year. By separating the sleep channel from the visibility channel that complicates studies of daylight saving time, our design shows that even small, recurring sleep disruptions impose large social costs.