In order to mass produce units of Photonic Fabric+Memory Appliance, the multichip modules need to be identical.     

How about CAI having thorough discussions with all the hyperscalers and finally deciding upon a standard configuration - a base model with latest versions of AMD xPUs and Samsung HBM stacks. Manufacture a base model, V1 (version 1) with 16 multichip modules in a box.

A V1 would likely also be suitable for the needs of Fortune 500 companies with applications that have less intensive workloads compared to a hyperscaler. When workloads increase, customers simply buy and connect as many units as required to get the desired performance.

What has not been mentioned so far, is that there will be a CAI server with an operating system that knows the location of all the resources and how to get data to and from all of them within every cluster regardless of size. The server is where the applications execute and it is will be the job of the operating system to perform the input-output operations on behalf of those applications.       

Should the most intensive applications become to much for what V1s can deliver, then develop a V2 by utilizing the latest improvements for AMD xPUs and Samsung HBM.  For me, the commitment appears to be for an ongoing collaboration between AMD and Samsung which means Nvidia might not be in the CAI picture at all.  

If my view proves to be to correct then Nvidia’s main rival AMD, currently at $164.35 is a candidate for a loftier share price down the road. I would never spend $32,870 on 200 AMD shares when I can buy about 5,500 POET shares for the same cash outlay.

Adding more PF+MA units to an existing setup should be trivial. It would involve a shutdown of the server, connect the new cluster(s), reboot the sever and the operating system will recognize all the new stuff and assign addresses to all the available resources and know where they are and how to get to and from them. The base patent for CAI is described as an ONOC. Optical Network On a Chip.