If the AI algorithms can only be optimized by pushing them into silicon, then a HW background could be valuable.

However, many software architectures have yet to embrace the pipelining and parallel capabilities of today's CPUs... meaning the need for application-specific ICs could be many year away.