You’re charged for all the time an instance is running, regardless of whether or not the GPUs are being utilized.
See Lambda’s docs on billing for more info.