NVIDIA H200 Tensor Core GPU

What the H200 changes

The H200 is the same Hopper compute die as the H100 with 141GB of HBM3e in place of 80GB of HBM3, and 4.8TB/s of bandwidth in place of 3.35TB/s. For inference on large language models that is the difference between a 70B model at a real batch size on a single GPU and needing two. For training, the extra memory lets you hold larger activations and cut gradient checkpointing.

Who buys it

• Inference teams serving 70B-class models where memory, not FLOPS, is the limit.
• Fine-tuning shops that want to stay on one node.
• Anyone upgrading an existing HGX H100 chassis: H200 is a drop-in SXM5 module in the same 8-GPU baseboard.

How it is sold

The H200 is not sold as a loose card for practical purposes. It ships inside an HGX 8-GPU baseboard in systems such as the SYS-821GE-TNHR and AS-8125GS-TNHR. Cloud pricing for H200 by the hour is on our price comparison.