DGX GH200 AI Supercomputer
Narrative
256 Grace Hopper superchips. 1 exaflop AI performance. 144TB shared memory. NVLink creates single unified GPU.
Reality
Deployed at Meta, Microsoft, Google. Exascale performance verified. NVLink fabric working but complex. Price estimated $50M+ per system. Long lead times.
Implication
Validated rack-scale thinking over individual GPU optimization. NVLink justified complexity premium. ARM Grace CPU proved viable in AI/HPC. Exascale training costs dropped significantly.