It feels like every major AI company is using NVIDIA DGX or HGX setups for training large language models.

What makes these systems so dominant? Is it just raw performance, or is the CUDA/software ecosystem the real lock-in?

If you’ve worked with these systems (or alternatives), what’s your experience — are they truly unmatched, or just the most convenient choice?