NVIDIA//H100

- Hopper architecture. Transformer Engine with hardware-level FP8 support.

Hopper architecture. Transformer Engine with hardware-level FP8 support.

2-3× A100 training throughput. ~$25K-$40K per card.

Export-restricted to China — the GPU that made chip policy geopolitical.

Every frontier AI lab's shopping list, perpetually backordered.