Buying H200 NVL in the EU: cards or a server, the NVLink bridge and the included AI Enterprise

How to buy H200 NVL: cards or a server

To buy H200 NVL for a company, choose between two routes: cards for a server whose maker lists the H200 NVL for that exact model, or a server built to order with 2, 4 or 8 cards. Before either, decide whether your models need NVLink bridges and whether a bridge joins two or four cards, because the bridge and the slots it spans are part of the order. Each card comes with a five-year NVIDIA AI Enterprise subscription, which NVIDIA ties to the card’s serial number.

NVIDIA’s H200 page describes the H200 NVL as a PCIe card, “Dual-slot air-cooled”, with 141 GB of HBM3e at 4.8 TB/s and a maximum TDP of “Up to 600W (configurable)”. The same page calls it “ideal for lower-power, air-cooled enterprise rack designs that require flexible configurations”. The server options NVIDIA lists for it are “NVIDIA MGX H200 NVL partner and NVIDIA-Certified Systems with up to 8 GPUs”, so the card goes into a rack server that its maker has qualified for it.

This guide covers the buying decisions. The fitting checks for a server you already own, from the 16-pin power cable to BIOS settings, are in our H200 NVL installation checklist, and this article does not repeat them.

H200 NVL, HGX H200 and DGX H200 compared

A search for H200 returns the H200 NVL card, the H200 SXM module and the systems built on the module. The H200 NVL is the PCIe card, while the H200 SXM is a module that sits on an HGX H200 baseboard, which NVIDIA lists in “partner and NVIDIA-Certified Systems with 4 or 8 GPUs”. DGX H200 is NVIDIA’s own 8U system, and its user guide lists eight H200 GPUs “that provide 1,128 GB total GPU memory”.

NVIDIA H200 product page (specifications marked preliminary by NVIDIA); NVIDIA DGX H100/H200 User Guide, last updated 26 January 2026, whose power rows apply to both systems; NVIDIA AI Enterprise licensing guide, updated 2 September 2026; all read on 10 October 2026.

For a buyer, the main technical difference is the reach of NVLink. In a DGX H200 the user guide gives 900 GB/s of GPU-to-GPU bandwidth for its eight GPUs, while an H200 NVL bridge joins at most four cards. An eight-card H200 NVL server therefore holds at most two NVLink domains of four, which our guide to H200 NVL card counts for large models works through model by model.

We supply the H200 NVL and servers built with it. We do not supply HGX or DGX systems. A model that has to run with tensor parallelism over eight GPUs on NVLink needs an HGX H200 or DGX H200 system rather than an H200 NVL server.

Buying the cards or a server built with them

The two routes differ in who answers for the platform. When you buy cards, the server maker’s support list for your exact model decides whether they go in at all. When you buy a server built to order, the server is chosen for the cards, their bridge and their power from the start.

NVIDIA H200 NVL Product Brief PB-12128-001_v01 (11 April 2025); Lenovo Press LP1944, ThinkSystem NVIDIA H200 141GB GPUs Product Guide (7 November 2025); our AI servers page for the server column.

The card route fits a company that already runs a server model listed for the H200 NVL. Lenovo, for example, lists the card and both bridges under their own part numbers: 4X67A97315 for the “ThinkSystem NVIDIA H200 NVL 141GB PCIe GPU Gen5 Passive GPU”, 4X67A97320 for the 2-way bridge and 4X67A97322 for the 4-way bridge. Its product guide states that the GPU “assumes the server’s base warranty and any warranty upgrades”. For cards bought outside the server maker’s own options, the quote should state which warranty covers them and who handles a claim.

How many cards a listed server holds is set per model. Lenovo Press lists at most two H200 NVL at 600 W in a ThinkSystem SR650a V4, so a plan for four cards at 600 W in one chassis needs a different server model.

The server route fits a new platform for several departments, for example two servers with four cards each for a private AI service shared by several departments. The cards, bridges, power supplies and cooling are then chosen and quoted together.

We supply the H200 NVL as cards with their NVLink bridges, or in AI servers built to order and burn-in tested. Tell us which server the cards would go into, or describe the rack position a new server would stand in.

NVLink bridge for H200 NVL: two-way, four-way or none

NVIDIA’s product brief lists two bridges in its table 4-2, in a column headed “Total NVLink BW”: a “2-slot (spans two cards)” bridge at “900 gigabytes per second” and a “4-slot (spans four cards)” bridge at “1800 gigabytes per second”. NVIDIA’s H200 page states the interconnect as “2- or 4-way NVIDIA NVLink bridge: 900GB/s per GPU”, and the brief’s overview gives “900 GBps bidirectional bandwidth” for up to four connected cards, so a sizing plan works with 900 GB/s per GPU for either bridge. Without a bridge the cards exchange data over PCIe Gen5 at 128 GB/s, the figure on the H200 page.

The bridge is needed when one model is split across cards with tensor parallelism, and for training that exchanges gradients between them. When each card serves its own copy of a model, the cards do not exchange data and no bridge is needed.

The bridge decision belongs before the order because it fixes the slot layout. NVIDIA’s product brief states that “bridged NVIDIA H200 NVL cards should be within the same CPU domain”, and lists “Bridge GPUs under different CPUs” under “Allowed but Not Recommended”. A server whose risers place two cards under each processor suits 2-way bridges. A 4-way bridge within one CPU domain needs four adjacent cards under one processor. Changing from 2-way to 4-way later can mean a different riser layout or a different server, so state the bridge type in the request.

For an eight-card plan the choice is between one server with two sets of four and two servers with one set each. NVIDIA’s recommended topology has the same number of GPUs under each CPU socket and identical bridge spans in every CPU domain, which in a two-socket server with eight cards gives one 4-way set per processor. Our comparison of one 8-GPU server or two 4-GPU servers covers failover, maintenance and rack space for that choice.

The five-year NVIDIA AI Enterprise subscription included with each card

NVIDIA’s licensing guide, updated on 2 September 2026, states: “Each NVIDIA H200 NVL Tensor Core GPU includes a five-year NVIDIA AI Enterprise subscription.” It adds “Software activation required”, and the subscription “can be activated with the serial number of selected GPUs”. NVIDIA’s H200 page lists AI Enterprise as an add-on for the H200 SXM and as included for the H200 NVL. For DGX H200, the licensing guide states that AI Enterprise “is included in the NVIDIA DGX software bundle for NVIDIA DGX systems using the NVIDIA Hopper architecture”.

The term is tied to each card rather than to the server. NVIDIA counts the start from the ship date “of the GPU board to the OEM partner”, plus 90 days for integration and delivery, and states that the start date “cannot be modified, as it is tied to the specific card”. A server with eight cards therefore carries eight subscriptions, and cards added in a later year end on later dates. After the stated term, NVIDIA’s guide says, a subscription “must be renewed to remain active”.

Record the serial numbers when the cards arrive, before the server is racked. The steps on NVIDIA’s portals are in our guide to activating NVIDIA AI Enterprise licences.

Power and cooling for H200 NVL servers

NVIDIA’s product brief gives the total board power for the 600 W cable strapping as “600 W maximum (default)”, with a “350 W power compliance limit” and a “200 W minimum”. Planning starts from the default. Four cards at 600 W draw 2,400 W and eight draw 4,800 W before processors, memory, drives and fans, so the power feed of the rack position can set the card count before the rack space does.

The card is passive, and the server’s fans move the air through it. That is why the server maker’s list names the fan modules, air ducts and slots for each model, and why an eight-card server is a different chassis from a two-card one.

The fitting details for an existing server, the cable part numbers, thermal rules and power planning tools of each maker, are in the installation checklist linked above. For a new server, we check the rack, power and airflow before we quote.

Decisions before the order

These six decisions, taken in this order, cover what a quote for H200 NVL cards or servers needs.

1. Models and precision: list the models the cards will serve or fine-tune and their weight precision, since these set the number of cards per model.
2. Cards and servers: decide the total number of cards and whether they go into one server or several, for example 8 cards as one server or as two servers with 4 each.
3. NVLink bridge: none, 2-way or 4-way for each set of cards, and the CPU domain each set sits in.
4. Server: an existing server, by full model number on the maker’s list, or a server built to order for the cards.
5. Power and cooling: the feed per rack position in amps and phases, the outlets and the inlet temperature.
6. Licences and partitioning: the five-year AI Enterprise subscription per card, any further licences for the other GPUs in the platform, and whether the cards will be split with MIG into up to seven instances of 16.5 GB, as our H200 NVL MIG profiles guide explains.

We check the rack, power and airflow before we quote. Send us your answers to these six points through the form below, even if some are still open.

What we supply

We supply the NVIDIA H200 NVL with its 2-way and 4-way NVLink bridges and its five-year NVIDIA AI Enterprise subscription, as cards or in AI servers built to order with 2, 4 or 8 cards, assembled and burn-in tested, with manufacturer warranty on every component. For a server you already run, we check the platform, power and cooling first and say in advance if the card is not supported in it. Further NVIDIA AI Enterprise and vGPU licences come on the same EU contract and invoice, together with the L40S, L4 and RTX PRO cards from our professional GPU range where the platform mixes them. Operating system, drivers, CUDA and a container runtime are installed on request, and if you want the deployment done for you, our engineering partner Vixen.UNO handles it under the same contract.

FAQ