AI is all about speed, specifically the speed of parallel computing power enabled by GPUs.
However fast a GPU is, however much computing power a cluster can deliver, there are numerous bottlenecks that hinder AI performance. A single GPU may be fast, but the system only runs as fast as its connections, and there are exponentially more connections as you move up the chain, each with more networking, power, and cooling requirements. Together, those constraints create a lucrative infrastructure market with numerous niches to fill.