One of Nvidia's actual moats is their system building competency, which AMD lacks. They can sell you a box / a whole server room configuration, since their acquisition of Mellanox together with network equipment.
The cost of Hardware Development is the main cost contribution.
Which I think is not true with regards to both GPU and GPGPU computing. The major cost for GPU is Drivers, and CUDA for GPGPU. i.e It is Software.
Unlike ARM where AWS/GCP/Azure can make their chips and benefits from the Software ecosystem already in place for ARM, there is no such thing on GPU. Drivers and CUDA is the biggest moat around Nvidia's CPU. And unless Developers figure out a way to drive the cost of DL and Drivers down, there is no incentives to switch away from Nvidia's ecosystem.
That is why I am interested to see how Intel tackle this area. And if History will repeat itself again in the Voodoo, Rage 3D and S3 Verge era.
However, I think NVIDIA is still vulnerable—but against AWS/GCP/Azure, not Intel/AMD.
My opinion is that deep learning is moving to the cloud. That's a bigger conversation with a lot of nuances, but if you take that basic assumption, then the development of ASICs like TPU/Inferentia become a big threat to Nvidia.
If the biggest buyers of chips in deep learning are the clouds, and the clouds are increasingly developing their own chips for deep learning, Nvidia is in a tough spot. They'll always have a place among labs that use their own machine, and of course, Nvidia's business is bigger than machine learning, but in general I think the clouds are a real threat.