Procurement profile
Where this model makes sense
Strong option for LLM inference/training, retrieval, analytics and large-context clusters.
- large-context LLM workloads;
- memory-bound inference and training;
- HGX, MGX and OEM server nodes for H200.
Hopper / High-memory GPU
NVIDIA H200 for LLM and AI infrastructure
H200 targets memory-bound workloads where H100 is constrained by memory capacity and bandwidth. Quotes start with server, GPU count and network architecture checks.
Specs
Procurement specifications
The table helps qualify the request quickly. For a commercial offer we verify datasheet, OEM SKU, availability and warranty region.
| Architecture | Hopper |
|---|---|
| Memory | 141 GB HBM3e |
| Memory bandwidth | 4.8 TB/s |
| Form factor | SXM / PCIe NVL by SKU |
| MIG | up to 7 MIG instances |
| Workloads | memory-bound LLM / inference / training |
RFQ
What to verify before quoting
- H200 SXM or H200 NVL;
- server QVL and firmware;
- cooling, power and rack readiness;
- networking and storage profile for the workload.
Critical limitation
NVIDIA H200
H200 should not be sourced without site checks: server cooling, industrial power and compatible nodes are required.
Specification sources
The table helps qualify the request quickly. For a commercial offer we verify datasheet, OEM SKU, availability and warranty region.
Internal links
Related NVIDIA pages
These links keep model-level pages inside the current catalog architecture without overloading the global menu.
RFQ
Request price and availability
Attach SKU, BOM or model list. For server systems, include site, power, cooling, network and compatible server details.