Methods, systems, and apparatus, including medium-encoded computer program products, for virtualization-based collaborative activity framework with predictive preloading of virtual machines. A collaboration orchestration system can obtain user interaction data describing interactions with a client device. In response to determining, at least in part from the user interaction data, that the client device is predicted to access an activity, and before receiving a request to access the activity, the collaboration orchestration system can: (i) provide an indication to create a virtual machine (VM) instance configured to provide the activity; and (ii) provide an indication to load the VM instance with the activity. The collaboration orchestration system can receive a reference to the VM instance. In response to obtaining, by the collaboration orchestration system, a request to access the activity from the client device, the collaboration orchestration system can provide to the client device, the reference to the VM instance.