Compared the different silicon players for AI inference (arjunrakesh.substack.com)

I believe that a pretty high level of dependency will always remain but its in the hyperscaler's best interest to de-risk supply and also to use custom ASICS for their core core flows.