Ie a library thats optimized to drive a bunch of hosts with ARM cores, strapped to nvidia gpu's, connected by mellanox interconnects.
Sell it a a unit to the public clouds for a massive markup, or maybe offer on their own cloud for an unbeatably low cost.
ARM is effectively embarrassingly parallel, low-power compute, and GPU's are parallel high-power compute.
If NVidia got into datacenter design (with ARM+GPU), dropped K8S+S3+Postgres onto it, charged $1999.95/mo for access, what could they deliver to customers?
What would a vertically integrated datacenter look like where you are the designer + manufacturer for everything from the concrete up?
A lot like AWS, Oracle Cloud, and soon to be Google Cloud and Azure?
people are losing their shit but this isn't even that novel lol, quite a few other players already own their full stack.
CUDA provides software lock-in to Nvidia GPUs. This would provide software lock in to the combined Nvidia Arm / Nvidia GPU platform.