GPU GUIDE · UPDATED 11 NOVEMBER 2026

NVIDIA H200 NVL: price, specifications and servers

The NVIDIA H200 NVL is the PCIe version of the H200: 141GB of HBM3e memory with 4.8TB/s of bandwidth in a dual-slot card that fits standard GPU servers, with NVLink bridges to connect two or four cards. This page covers the specifications, how the NVL differs from the H200 SXM, our indicative EU prices for one to eight GPUs and what an H200 NVL server needs.

At a glance

141GB HBM3e

The largest memory of any PCIe GPU we offer, for big models, long contexts and many concurrent users.

4.8TB/s bandwidth

About three times the RTX PRO 6000, which speeds up token generation when a model is memory-bound.

NVLink bridges

Two or four cards can be linked at 900GB/s per GPU, which helps when one model is split across GPUs.

$29,490 at NodeX

Indicative price per card in our configurator on 11 November 2026, confirmed in your written quote.

H200 NVL specs

Figures from NVIDIA’s published specifications, which NVIDIA marks as preliminary. Tensor Core figures are with sparsity, as NVIDIA states.

Specification NVIDIA H200 NVL
Architecture NVIDIA Hopper
GPU memory 141GB HBM3e
Memory bandwidth 4.8TB/s
FP8 Tensor Core 3,341 TFLOPS (with sparsity)
FP16 / BF16 Tensor Core 1,671 TFLOPS (with sparsity)
TF32 Tensor Core 835 TFLOPS (with sparsity)
FP32 60 TFLOPS
FP64 / FP64 Tensor Core 30 / 60 TFLOPS
Decoders 7 NVDEC, 7 JPEG
Multi-Instance GPU Up to 7 instances of 16.5GB
Confidential computing Supported
Maximum power Up to 600W, configurable
Form factor PCIe, dual-slot, air-cooled
Interconnect 2- or 4-way NVLink bridge at 900GB/s per GPU; PCIe Gen5 at 128GB/s
NVIDIA AI Enterprise Included, according to NVIDIA

H200 NVL vs H200 SXM

Both have 141GB of HBM3e at 4.8TB/s. The difference is the form factor and the server around it.

H200 NVL H200 SXM
Form factor PCIe card, dual-slot SXM module on a GPU baseboard
Servers Standard PCIe GPU servers, up to 8 GPUs HGX H200 systems with 4 or 8 GPUs
GPU-to-GPU link NVLink bridges for 2 or 4 GPUs, 900GB/s per GPU NVLink at 900GB/s
Maximum power Up to 600W Up to 700W
FP8 Tensor Core 3,341 TFLOPS (with sparsity) 3,958 TFLOPS (with sparsity)
Multi-Instance GPU Up to 7 × 16.5GB Up to 7 × 18GB
NVIDIA AI Enterprise Included Add-on

The SXM version gives somewhat more compute per GPU, but it only comes in HGX systems built around a GPU baseboard — a different class of server. The NVL fits the same kind of 4U PCIe GPU server as other professional GPUs, which is what we build.

H200 NVL price in the EU

NodeX sells the H200 NVL as part of a complete GPU server, built and tested in Bulgaria. Indicative GPU prices on 11 November 2026:

Configuration Total GPU memory GPUs at indicative prices
1× H200 NVL 141GB $29,490 (about $209 per GB)
2× H200 NVL 282GB $58,980
4× H200 NVL 564GB $117,960
8× H200 NVL 1,128GB $235,920

These totals cover the GPUs only. The platform, processors, memory, storage, networking, NVLink bridges, assembly and warranty are confirmed and priced in your quotation. VAT, delivery and availability are confirmed there too.

What H200 NVL servers are good for

One card: 141GB

70B models in 8-bit with long contexts and many concurrent users.

Two cards: 282GB

70B models in 16-bit with room for context, linked by an NVLink bridge.

Four cards: 564GB

400B-class models in 8-bit, about 400GB of weights, with room for context.

Eight cards: 1,128GB

670B-class models in 8-bit, about 670GB of weights, on a single server.

The H200 NVL also suits FP64 scientific computing, at 30 TFLOPS FP64 and 60 TFLOPS FP64 Tensor Core per card. For the sizing arithmetic, see how much GPU memory a local LLM server needs.

H200 NVL or RTX PRO 6000?

The RTX PRO 6000 Blackwell has 96GB for $11,900 (indicative), about $124 per GB against about $209 for the H200 NVL, and adds FP4 support and graphics capabilities. The H200 NVL wins when memory bandwidth, capacity per GPU, NVLink or FP64 matter. Our RTX PRO 6000 vs H200 NVL comparison goes through the numbers in detail.

What an H200 NVL server needs

  • Airflow. The card is passively cooled and depends on the server’s fans, so the chassis must be designed for passive 600W GPUs.
  • Power. Eight cards can draw about 4.8kW for the GPUs alone, before processors, memory and storage. Plan the power and cooling of the room or colocation before you buy.
  • PCIe 5.0. To use the card’s PCIe Gen5 link, the platform needs PCIe 5.0 slots — for example AMD EPYC 9004-series (SP5) processors in our NX-G8 Dual SP5.
  • NVLink spacing. NVLink bridges need matching slot spacing between the linked cards.

We confirm all of this for your platform in the quotation.

H200 NVL servers from NodeX

NodeX NX-G8 Dual SP5 GPU server for H200 NVL configurations
NX-G8 Dual SP5: up to 8 GPUs with dual AMD EPYC SP5 processors and DDR5 ECC memory.

We build 4U GPU servers in Bulgaria with one to eight H200 NVL GPUs, tested before delivery. We can also arrange colocation in Bulgaria and install the operating system. Tell us your models, number of users and context length, and we will propose a configuration in a written quotation.

Frequently asked questions

What is the difference between the H200 NVL and the H100 NVL?

The H100 NVL has 94GB of memory; the H200 NVL has 141GB of HBM3e at 4.8TB/s. More memory lets larger models and longer contexts run on each card.

Do I need NVLink bridges?

Not for running separate models on each GPU. They help when one model is split across two or four cards, because the GPUs exchange data much faster than over PCIe. Bridges are confirmed and priced in your quotation.

Can the H200 NVL go in any server?

No. It needs a server designed for passive 600W GPUs with sufficient airflow and power. We confirm compatibility for your platform in the quotation.

Sources

Specifications and platform details on this page come from the published sources below. Prices come from our configurator on the date shown at the top of the page.