How important are the “normal” cores in an AI workload? Do AI-specific chips like Microsoft’s actually threaten Nvidia’s business?

Nvidia chips contain a huge number of CUDA cores and a separate set of Tensor cores to accelerate instructions relevant to AI workloads.

While I don't expect Nvidia's dominance to change in the next 3-4 years, some of their large customers are researching their own chips. Microsoft and Amazon are known to be working on it, Google is already known to have their own silicon, as is Apple.

What I would like to understand is how would Nvidia's offerings fit into their business model if the cost/performance/space efficiency of those in-house custom chips is better than what Nvidia's current and future models can offer?

I don't think Nvidia's Mellanox business is threatened, because even systems containing those chips will need to be networked to each other, but their current growth comes largely from the sale of datacentre GPUs.

Note: I do own ARM, Nvidia, TSMC, Google, Apple and AMD stocks, which are all relevant to this discussion. I don't currently own Microsoft or Amazon stocks. I also own several other stocks but I'm not sure if they are relevant here.