Stolen CPU in the cloud refers to a metric that measures cycles a CPU should have been able to run but couldn't due to hypervisor actions, which can be hard to understand and implies malevolent intent from virtual neighbors. It's a relative measure of cycles diverted away from instances by hypervisors enforcing quotas based on ECU purchases. The amount of stolen CPU varies over time, likely due to other instances requesting CPU cycles from the same hardware. This phenomenon is particularly problematic for CPU-intensive applications that frequently approach idle CPU of zero, making it difficult to predict performance without insight into other tenants' CPU usage behavior. To resolve issues caused by stolen CPU, buyers can consider buying more powerful EC2 instances, baselining their application's compute needs, profiling their app on an instance before deployment, or re-deploying their application in another instance.