NVIDIA DGX H200
DGX H200 is NVIDIA's factory-integrated Hopper-generation system: eight H200 SXM GPUs (141 GB HBM3e each) on NVLink 4, dual Xeon hosts, and ConnectX-7 fabric in an 8U chassis with NVIDIA-direct support. It is the prior-generation counterpart to DGX B200, sharing the H100 chassis and board layout but with the higher-capacity, higher-bandwidth H200 GPU swapped in. Buy DGX H200 for a validated, single-vendor-supported node, particularly where an existing H100/H200 DGX or HGX fleet needs a drop-in match; buy an HGX H200 OEM server (Supermicro, Dell, HPE, Lenovo) when you want to control the CPU, storage or networking BOM or already run your own cluster ops.
- At ~10.2 kW max, one DGX H200 still exceeds a standard 5 to 8 kW colo rack's power allocation on its own; plan high-density power delivery even though it draws noticeably less than a DGX B200.
- The H200's 8U chassis is shorter than the B200's 10U but the weight is close (287.6 lb vs 313.9 lb), so rack floor loading and lift equipment planning don't get meaningfully easier by choosing H200 over B200.
- Because H200 shares the H100 chassis and 4+2 PSU layout, an existing H100 DGX SuperPOD rack can usually absorb H200 nodes without a facilities redesign, which is the main reason to pick H200 over waiting for or refitting to Blackwell.
- 141 GB per GPU (vs 80 GB on H100) mainly buys headroom for larger KV caches and batch sizes in inference; if the workload is training-bound and memory capacity isn't the bottleneck, the H200 premium over H100 is often not worth it versus deploying B200 instead if budget allows.
- If your fleet is already HGX-based and OEM-supported, an HGX H200 server from your existing vendor is usually the lower-friction and lower-cost path to the same GPU; reserve DGX H200 for teams that need NVIDIA's own support SLA or SuperPOD reference-architecture compliance.
How much does a DGX H200 cost?
NVIDIA does not publish a list price. Reported street pricing runs roughly $400,000 to $550,000 for a complete 8-GPU system, with one reseller listing it flat at $550,000. Expect a formal quote process and regional/volume variation.
DGX H200 vs DGX B200: which should we buy?
H200 (Hopper) draws less power (~10.2 kW vs ~14.3 kW), fits the same 8U/H100-era rack infrastructure, and costs less. B200 (Blackwell) delivers materially higher training and inference throughput via NVLink 5 and newer tensor cores, at a higher power and price point. Choose H200 for a lower-risk fit into existing H100 infrastructure or budget constraints, B200 for new deployments chasing peak performance per GPU.
Can a DGX H200 use existing DGX H100 rack power and cooling?
Largely yes. The H200 keeps the H100's 8U chassis, 6x 3.3 kW PSU layout (4+2 redundant), and similar airflow profile (1,105 CFM), so an H100-provisioned rack typically does not need a facilities redesign to host H200 nodes.
Is DGX H200 still worth buying with Blackwell available?
Yes for specific cases: it is still shipping as of mid-2026, priced below DGX B200, and is a drop-in for existing H100/H200 racks. It makes less sense for a greenfield deployment with no legacy infrastructure and no power constraint, where B200's higher throughput per dollar of power usually wins.