Nvidia should really be able to do both. Their integrated solutions like Orin scale up to a certain point, but the demand for modular compute is still enormous

Nvidia should really be able to do both. Their integrated solutions like Orin scale up to a certain point, but the demand for modular compute is still enormous on the server side (and enthusiast market). Given that CUDA will support both going forward, I don't think there's a technical incentive to kill the market. My guess is that Nvidia will have a line of mobile cores/chipsets for integrators that want them, while also offering PCI-enabled boards for gamers and enthusiasts. Even Apple can't outrun the demand for a PCI-enabled machine, and they don't even support eGPUs. Nvidia's incentive to abandon ATX (or at least modularity) is even slimmer.