Methods and apparatus for dynamically allocating host resources (e.g., CPUs, GPUs, etc.) to virtual machines (VMs) on host devices in a provider network. The host devices may be provisioned with quantities of each resource type. Customers may request different combinations and quantities of resources for their VMs. Upon receiving a placement request for a VM, a host device is located that can provide a requested combination and quantity of resources for the VM. The host can then be directed to attach at least the requested combination and quantity of host resources to the VM. Future demand for VMs with particular combinations and quantities of resources can be predicted, and logical slots can be predefined in the control plane in anticipation of that demand. If a customer's VM is provided with more resources than requested, the customer may release or sell the extra resources.