You can easily build a localized, access-controlled compute grid using Apple Silicon hardware. Haven't tested but as far as I can see, the key to making it performant would be keeping full models resident on individual nodes rather than attempting to distribute the tensor processing across the network.