Buy a server with 2× NVIDIA H200 NVL in Europe: 282 GB on a 2-way NVLink bridge

The entry server for the H200 NVL

You can buy a server with two NVIDIA H200 NVL cards from us, built to order and delivered anywhere in the EU. Two cards of 141 GB HBM3e on a 2-way NVLink bridge give 282 GB of GPU memory in one NVLink domain, with two AMD EPYC processors, 768 GB of DDR5 ECC memory and two NVMe drives. Our team sizes the configuration to your workload and budget and replies within one business day.

Two cards are the smallest NVLink configuration of the H200 NVL and the usual first step into it. They suit a company that runs a 70B model for one or two departments and needs it fast for each user, a team fine-tuning a 70B model with LoRA, or a research institute running a double-precision solver, since each card delivers 30 TFLOPS of FP64 by NVIDIA's H200 page. The bridge lets one copy of a model use both cards at NVLink speed; without it, the cards would exchange data over PCIe Gen5 at 128 GB/s instead of 900 GB/s per GPU. Our guide to what a 2× H200 NVL server runs covers the models and layouts in more depth.

When the model or the user count outgrows the pair, four cards in one NVLink domain of 564 GB are the next step, as our page to buy a 4× H200 NVL server describes.

Configuration and options

The table shows the two-card configuration we start from and what we change per project.

Our AI servers page; NVIDIA AI Enterprise reference architecture and NVIDIA-Certified configuration guide (30 September 2026); HPE DL380a Gen12 QuickSpecs, Supermicro SYS-212GB-NR page and Cisco UCS C845A M8 spec sheet A.11, read on 24 October 2026.

We size the memory by NVIDIA's configuration guide for NVIDIA-Certified Systems: at least two sockets, six physical cores per GPU and system memory of at least twice the GPU memory, 564 GB for two cards. Two EPYC 9005 processors have 24 memory channels, so 24 modules of 32 GB give 768 GB with one module per channel. NVIDIA's AI Enterprise reference architecture sets a floor of 128 GB per GPU and asks for seven cores per GPU, plus two per GPU for the operating system and two per MIG instance: 18 cores for two cards, 46 with all 14 MIG instances. Cisco recommends 1,024 GB for two H200 NVL in its UCS C845A M8, at least three times the HBM, which suits fine-tuning with CPU offload. Our guide to how much RAM a GPU server needs compares the rules.

Chassis: 2U for two cards, 4U to grow

Two H200 NVL fit a 2U server. Lenovo's product guide for the ThinkSystem SR650a V4, updated 20 October 2026, names "2x NVIDIA H200 NVL 600W GPUs" for the 2U server with one or two Xeon 6 processors. Supermicro's 2U SYS-212GB-NR lists the H200 NVL with an optional NVLink bridge and a GPU bay for two cards with NVLink, with three 2,000 W supplies in 2+1; it has one processor socket, one fewer than NVIDIA's certified-systems guide asks for. In 4U, HPE's ProLiant Compute DL380a Gen12 takes two H200 NVL with the 2-way bridge S4A90C and five 2,400 or 3,200 W supplies, and Cisco's UCS C845A M8 lists one 2-way bridge for two cards.

The bridge is a separate part, and the server maker lists it per server model, separately from the card. NVIDIA's product brief PB-12128-001 gives the 2-slot bridge the part number 900-23945-xx00-xxx and recommends that bridged cards sit "within the same CPU domain", so both cards of the pair go under one processor. Our H200 NVL NVLink bridge article shows the checks with nvidia-smi topo -m.

For growth, a 4U chassis with power supplies and slots for four or eight cards is the better start. Adding a second pair keeps the 2-way bridge, and Cisco lists two 2-way bridges for four cards. A move to one 4-way domain of four needs a new bridge across four adjacent cards, so we plan the slot layout for the later step at the quote stage, as our guide to planning a GPU server for growth explains.

What two H200 NVL run

By our users-per-card rule, 0.9 × 140.4 GiB − 3 GiB per card, the pair holds 246.7 GiB of weights and KV cache. The rule takes 90 per cent of the memory the driver reports and keeps 3 GiB per card for CUDA graphs and activations, as our guide to how many users one RTX PRO 6000 can serve derives.

Our estimates by the users-per-card rule, concurrent conversations at full context length. FP8 KV cache for Llama 3.3 70B in FP8 (1.25 GiB at 8K) and for Qwen3-235B (94 KiB per token), 16-bit cache for the BF16 row (2.5 GiB) and gpt-oss-120b (1.125 GiB at 32K). Weights 67.7 GiB, about 131.5 GiB in BF16 by our estimate, 60.8 GiB and 219.8 GiB.

For the second row, the pair holds 246.7 GiB; after 67.7 GiB of weights, 179.0 GiB remain, enough for 143 conversations at 1.25 GiB each. The BF16 row assumes 70.6 billion parameters at two bytes, about 141 GB, for teams that keep the model at full precision. Qwen3-235B shows the limit of the pair: the FP8 weights of 236 GB take most of the memory, so it serves a small team or short contexts, and four cards serve it with 93 conversations of 32K. Models of the 400B class, such as Llama 4 Maverick in FP8 at about 417 GB, do not fit two cards.

These are requests in flight at the same moment, far fewer than the staff using the service. The H200 NVL has no FP4 arithmetic, so models are sized in FP8, INT4 or BF16, as our guide to FP4 models on the H200 NVL explains. We supply this server with the layout your model needs; send us the model, context length and peak requests in flight, and our team replies with a configuration and the conversations per layout.

Two H200 NVL or two RTX PRO 6000 Server Edition

Our entry server with two RTX PRO 6000 Blackwell Server Edition cards is the other two-card option. It has 192 GB on PCIe without NVLink and FP4 Tensor Cores; the H200 pair has 282 GB, three times the memory bandwidth per card and NVLink.

NVIDIA product pages; conversations at 8K with an FP8 cache by our users-per-card rule; Server Edition figures from our guide to the four-card RTX PRO 6000 Server Edition server.

The Server Edition pair suits NVFP4 models, graphics vGPU and mixed workloads, as our page to buy a 2× RTX PRO 6000 Server Edition server describes. The H200 pair suits FP8 models of 70B to 235B with many users, LoRA fine-tuning and FP64 work, as our RTX PRO 6000 vs H200 NVL comparison explains.

Power and room requirements

NVIDIA's product brief gives each card 600 W by default and 200 W minimum, with one 16-pin connector, and rates the passive card for 10 to 45 °C ambient. By our estimate, two cards at 600 W with two AMD EPYC 9355 processors at 280 W need about 2.0 kW at the wall: 1.76 kW for cards and processors plus 15 per cent for memory, drives, network cards, fans and conversion losses. That is within the 2.94 kW we plan per 230 V 16 A single-phase feed at 80 per cent, so one such feed per side carries the server with redundant supplies on A and B feeds.

• Outlets. HPE allows only C19 or C19-C20 cords for the 2,400 and 3,200 W supplies its DL380a Gen12 needs with the H200 NVL; check that the rack PDU has C19 outlets.
• Cooling. About 2.0 kW of heat needs roughly 320 CFM of air at an 11 °C rise, by our guide to GPU rack power and cooling, and the server maker sets the inlet limit per configuration.
• Room. At full load the server belongs in a server room or data centre, as our guide to a GPU server in an office explains.

We check rack, feed and airflow before we quote; tell us the feed type and the inlet temperature at the rack position, and we choose the chassis to match.

Software and the included NVIDIA AI Enterprise

On request the server arrives with the operating system, NVIDIA driver, CUDA and a container runtime installed, with the bridge checked and MIG set up if you plan it; each card splits into up to seven instances of 1g.18gb, as our H200 NVL MIG profiles show. NVIDIA's product brief asks for driver R565 TRD1 or later. As of October 2026, R580 is NVIDIA's long-term branch with support to June 2028, and NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as fixed drivers.

NVIDIA's licensing guide, updated 2 September 2026, states that "Each NVIDIA H200 NVL Tensor Core GPU includes a five-year NVIDIA AI Enterprise subscription". The term starts 90 days after the board ships to the server maker, cannot be changed, and is activated with the GPU serial number on the system where the cards are installed, as our article on activating NVIDIA AI Enterprise shows. It covers NIM and compute vGPU (C-series); the H200 NVL has no graphics vGPU, and open-source engines such as vLLM run without the subscription. Models, RAG and the platform on top are our Private AI/ML service.

How the order works

1. You describe the workload: models, users, data, rack and timeline. Sizing is free of charge.
2. We propose the configuration and check that cards, bridge, chassis, power supplies and cooling fit your rack and feed.
3. You receive a written quote within one business day, with price, lead time and warranty terms per component. Quantities and lead time are confirmed in writing with the quote.
4. You order on one EU contract with an EU invoice from Eurokommerz in Vienna.
5. The server is assembled and tested before delivery, with a power-on and basic load test.
6. We deliver anywhere in the EU, with commissioning on site on request; other destinations on request through our logistics partner.
7. Every component carries the manufacturer warranty, handled through us; DOA units are replaced, and the assembly is covered by us.

For a quote, send the models with precision and context length, the peak requests in flight or the number of users, any fine-tuning or FP64 work, the software platform, the rack position with its feed, and whether the server should grow to four or eight cards. We also supply the cards and the bridge alone for a server you already run, after the checks in our checklist for an H200 NVL in an existing server. Our GPU server acceptance testing guide shows how to run your own acceptance test on delivery.

What we supply

We supply the server with two NVIDIA H200 NVL and their 2-way NVLink bridge in 2U or 4U, as well as four-card and eight-card versions and the H200 NVL cards and bridges for servers you already run. The five-year NVIDIA AI Enterprise subscription comes with each card, and further NVIDIA AI Enterprise licences come on the same EU contract and invoice as the hardware.

FAQ