NVIDIA H200 NVL GPU Server

H200 NVL capacity is being installed across our data centers now. Submit the reservation form with your preferred configuration and region; our engineers confirm the lead time for that build, hold the hardware for you, and send a quote. Reservations are free and non-binding until you approve the quote.

Both carry 141 GB of HBM3e at 4.8 TB/s on Hopper silicon. H200 SXM runs at 700 W with 1,979 TFLOPS FP16 Tensor and links 8 GPUs through NVLink and NVSwitch at 900 GB/s on an HGX baseboard; H200 NVL is a dual-slot PCIe Gen 5 card at up to 600 W with 1,671 TFLOPS FP16 Tensor and a 2-way or 4-way NVLink bridge at 900 GB/s per GPU. Choose H200 SXM for 8-GPU training and the highest per-GPU throughput; choose H200 NVL for inference, RAG and HPC in air-cooled PCIe servers or racks that cannot take HGX systems.

H200 NVL raises memory from 94 GB HBM3 to 141 GB HBM3e and bandwidth from 3.9 TB/s to 4.8 TB/s, on the same Hopper architecture and PCIe form factor. The extra 47 GB per card lets a single GPU hold larger models and longer contexts, and the 4-way NVLink bridge pools 564 GB across four cards. TDP moves from 350 to 400 W on H100 NVL to up to 600 W configurable on H200 NVL, still air-cooled.

Pricing is quoted per configuration because card count, host CPU, memory, storage and uplink choices vary widely. Volume discounts apply to multi-GPU and multi-server orders, and every GPU server qualifies for committed-term discounts on 6 and 12-month billing. Crypto, card, PayPal and bank wire are accepted.

NVLink bridges connect two or four H200 NVL cards in the same server at 900 GB/s per GPU, so tensor-parallel inference across a bridged group does not cross the PCIe bus. A 4-way group pools 564 GB of HBM3e, enough for 400B-class models in FP8 or 70B-class models with very long contexts. An 8-card node runs as two 4-way bridged groups; if you need all-to-all NVSwitch bandwidth across 8 GPUs, choose H200 SXM.

Yes. H200 NVL is Hopper, the same architecture as H100 and H200 SXM, so CUDA 12 builds, PyTorch, JAX, TensorFlow, TensorRT-LLM and vLLM run unchanged. NVIDIA lists a 5-year NVIDIA AI Enterprise subscription with H200 NVL, which adds NIM microservices and enterprise support for the software stack. We install Ubuntu, Debian, Rocky/AlmaLinux, Windows Server or your own ISO, with the NVIDIA driver, CUDA toolkit and container runtime preinstalled on request.

H200 NVL is a dual-slot PCIe card with a configurable TDP of up to 600 W and air cooling, so it fits standard enterprise servers rather than HGX baseboards. Our GPU racks are provisioned for high-density PCIe servers; an 8x H200 NVL node is deployed with the power feeds and airflow it needs. You do not need to plan facilities, only the workload.

Yes. Each H200 NVL can be partitioned into up to 7 MIG instances of 16.5 GB, each with its own memory and compute, for many concurrent models on one card. Hopper Confidential Computing keeps data and model weights protected in GPU memory, useful for private inference services and regulated workloads.