Methods, systems, and computer-readable media for scaling for virtualized graphics processing are disclosed. A first virtual GPU is attached to a virtual compute instance of a provider network. The provider network comprises a plurality of computing devices configured to implement a plurality of virtual compute instances with multi-tenancy. The first virtual GPU is replaced by a second virtual GPU based at least in part on a change in GPU requirements for the virtual compute instance. The first and second virtual GPUs are implemented using physical GPU resources that are accessible to the virtual compute instance over a network. Processing for the virtual compute instance is migrated from the first virtual GPU to the second virtual GPU. An application is executed using the second virtual GPU on the virtual compute instance.