Techniques for optimization-based pool protection for a cloud provider network are described. An exemplary method includes receiving historical usage data of virtual machine instances of a capacity pool of a cloud provider network for each account of a plurality of accounts of the cloud provider network, generating a linearly extrapolated usage, based at least in part on an extrapolating parameter, for each account based at least in part on respective usage percentiles of the virtual machine instances from the historical usage data, determining a usage of the virtual machine instances for each account based at least in part on the linearly extrapolated usage for a same extrapolating parameter value, receiving, by the cloud provider network, a request to launch a computing resource for an account, determining a usage limit for the account based at least in part on the usage for that account, and launching the computing resource when a requested usage for the computing resource is less than or equal to the usage limit and not launching the computing resource when the requested usage for the computing resource is greater than the usage limit.