Technologies are described for protecting compute resources during outage conditions. For example, when an outage condition is detected, currently allocated compute resources can be protected by not releasing them in response to the outage condition. For example, a load pattern representing historical usage of compute resources by a computer service can be obtained. A predicted load pattern of compute resources can be generated based on the obtained load pattern. An outage condition related to the computer service can then be detected based on the predicted load pattern. In response to detecting the outage condition, compute resources can be protected and not released in response to the outage condition.