Buy a server with 8× NVIDIA H200 NVL and NVLink bridges in Europe

Who the eight-card training node is for

You can buy a server with eight NVIDIA H200 NVL and two 4-way NVLink bridges from us, built to order and delivered anywhere in the EU. It is the training node among our AI servers: eight 141 GB cards in two NVLink domains of four, two AMD EPYC processors, 2.3 TB of DDR5 ECC memory, eight NVMe drives and 400G networking. Our team sizes the configuration to your workload and budget and replies within one business day.

The node suits a company or research institute that trains and fine-tunes its own models, or serves models too large for four cards. Typical jobs are LoRA on 70B models, full fine-tuning up to about 32B, serving DeepSeek-V3.2 or GLM-5.3 in FP8, and a shared platform where up to 56 MIG instances serve development teams next to production. The H200 NVL also computes 30 TFLOPS in FP64 and 60 on its Tensor Cores by NVIDIA's specification, which matters for simulation codes. For inference of 70B-class models alone, four cards or our inference node with RTX PRO 6000 Server Edition cards need less power and rack space.

What the configuration contains

The table shows the reference configuration and what we change per project.

Our AI servers page; NVIDIA AI Enterprise reference architecture (compute node, 2-8-5) for cores, memory per GPU, NVMe and network adapters; NVIDIA-Certified Systems configuration guide (30 September 2026); Dell XE7740 technical guide (Rev. A04) and Cisco C845A M8 spec sheet (Rev. A.11) for power supplies, read on 24 October 2026.

The memory follows NVIDIA's configuration guide for NVIDIA-Certified Systems, at least twice the GPU memory: 2 × 1,128 GB is 2,256 GB, and 24 modules of 96 GB make 2,304 GB with one module on each of the 24 channels of two EPYC 9005 processors. The makers' own rules differ. NVIDIA's reference architecture asks for at least 128 GB per GPU, 1 TB for eight; Dell marks 512 GB and 1 TB as not validated for eight H200 NVL in its XE7740; Cisco recommends at least three times the HBM, 4,096 GB, for its C845A M8. NVIDIA's reference architecture also asks for at least seven cores per GPU plus two per GPU for the operating system, which is 72 for eight cards; two 64-core EPYC 9555, for example, give 128.

Two 4-way bridges and the slot layout

An H200 NVL bridge joins two or four cards, never eight. Dell's technical guide states that "for eight DW H200 GPU cards, two bridges are required", and Cisco's spec sheet requires each bridge to be fully populated. The server therefore has two NVLink domains of four with 900 GB/s per GPU inside each domain, joined over PCIe 5.0. NVIDIA's reference architecture recommends that bridged cards sit under the same processor, and Dell calls bridging across the two processors possible but not recommended. The bridges are fitted with each domain under one processor where the chassis maker's slot layout allows it, and the server is delivered with the bridges in place; Dell, for example, specifies two removal tools per bridge. Which servers take how many cards, with which bridges and power supplies, is listed in our guide to H200 NVL compatible servers.

Variants: memory, processors, network and growth

• Memory. 2.3 TB covers training with LoRA and serving. Full fine-tuning of a 70B model with CPU offload puts about 847 GB of optimiser states into system memory, and prefix caches offloaded to the host grow quickly, so 4 TB is the option for those jobs.
• Processors. MIG adds two cores per instance in NVIDIA's reference architecture; a server split into 56 instances needs over 180 cores by that rule, so we choose larger processors when MIG is part of the plan.
• Network. One 400G adapter suits a single node with fast storage. For a cluster, NVIDIA's reference architecture pairs one 400G SuperNIC with every two GPUs, four per server, plus a BlueField-3 DPU.
• Four cards first. A server can start with four cards on one 4-way bridge, in a chassis with power supplies and slots for eight, and take the second domain later; our guide to adding GPUs later explains what to choose at the start.
• Power mode. The card's power is configurable below 600 W; capped at 450 W, the server needs about 5.0 kW by our estimate, as H200 NVL at 450 W explains.

What it runs

By our users-per-card rule, 0.9 × 140.4 GiB − 3 GiB per card, the eight cards hold 986.9 GiB of weights and KV cache, 493.4 GiB per NVLink domain. The table gives examples; our guide to what an 8× H200 NVL server runs has every layout, the fine-tuning limits and the arithmetic.

Our estimates by the users-per-card rule: concurrent conversations at full context length, FP8 KV cache for Llama 3.3 70B and GLM-5.3, 16-bit for the others; GLM-5.3 and DeepSeek-V3.2 from our model guides.

These are requests in flight at the same moment, which a company's gateway logs show at the busy hour. We supply this server in every variant above; send us your models, training jobs and peak requests, and our team returns a configuration with the figures per layout.

NVIDIA AI Enterprise for five years with every 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." Eight cards therefore come with eight subscriptions, and no separate licence is needed for NVIDIA's AI software on this server during the term. Three details from the same guide matter for planning.

• Start date. The term starts with "the ship date of the GPU board to the OEM partner, plus 90 days", and "cannot be modified, as it is tied to the specific card". It does not wait for activation.
• Activation. The subscription is activated with the card's serial number; our guide to activating NVIDIA AI Enterprise shows the steps.
• Where it runs. The software "can be installed on the NVIDIA Certified Systems in which the GPUs are installed, but not on a different server or instance". After five years the subscription has to be renewed to stay active.

The subscription covers NIM microservices, NVIDIA support and compute vGPU, which NVIDIA licenses only through AI Enterprise. Open-source engines such as vLLM and SGLang run on bare metal without it. We state the subscriptions on the quote with the cards, so the term is visible before you order; our NVIDIA AI Enterprise page explains the licence types.

Power and room requirements

By our estimate, eight cards at 600 W with two AMD EPYC 9555 processors at 360 W need about 6.3 kW at the wall, and about 5.0 kW with the cards capped at 450 W; a measurement on the delivered system confirms the figure. We plan a 230 V 32 A single-phase feed at 80 per cent of its rating, 5.89 kW, so the full-power server needs a 400 V three-phase position (8.87 kW planned at 16 A). With redundant power supplies on an A and a B feed, each side carries the whole server.

• Power supplies and outlets. Dell's XE7740 uses six supplies for the GPU zone in 3+3 and rates its 3,200 W supply at 2,900 W between 200 and 220 V; HPE requires eight 2,400 or 3,200 W supplies for eight double-wide GPUs in the DL380a Gen12, with C19 or C19-C20 cords only. Check that the rack PDU has enough C19 outlets.
• Cooling. About 6.3 kW of heat needs roughly 1,000 CFM of air at an 11 °C rise, by the 160 CFM per kW in our guide to GPU rack power and cooling. Dell sets the inlet limit for eight H200 NVL in the XE7740 at 30 °C, against 35 °C for other cards.
• Weight and room. Cisco gives 55 kg for its 4RU C845A M8, so plan two people or a lift for racking. A server of this class belongs in a data centre or a server room with three-phase power.

Software delivered with the server

On request the server arrives with the operating system, NVIDIA driver, CUDA and a container runtime installed, so your team can start training or a serving engine on the first day.

Our AI servers page; driver minimum from NVIDIA's product brief PB-12128-001_v01 for the H200 NVL.

As of October 2026, NVIDIA AI Enterprise 8.2 ships the R595 data-centre driver 595.91.07, and R580 is NVIDIA's long-term branch with support to June 2028; NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as fixed drivers. Our article on NVIDIA driver branches and CUDA versions explains the choice. Models, RAG and the platform on top are a separate service, Private AI/ML.

How ordering works with us

The order runs in the same steps for one server or several.

1. You describe the workload: models, training jobs, users, data, rack and timeline. Sizing is free of charge.
2. We propose the configuration and check that cards, bridges, 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; the AI Enterprise subscriptions included with the cards are listed on the quote.
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.

We also build into your own chassis or with parts you already own, after a compatibility check of the platform, power and cooling. For your own acceptance test on delivery, use the checks in our GPU server acceptance testing guide.

What to send us for a quote

A short description is enough to start; these points let us quote the exact configuration in the first reply.

• Models you plan to train or serve, with precision, context length and method (LoRA, QLoRA, full fine-tuning)
• Peak requests in flight for serving, or the number of users and teams
• Whether one node is enough or the server joins a cluster, and the storage it reads data from
• Software platform: bare metal, Kubernetes or a hypervisor, with the version
• Rack position: free height units, feed type and current, A/B feeds, PDU outlets and inlet temperature
• Delivery site, invoicing entity, timeline, and any tender requirements, which our GPU server tender specification helps to word

Describe these points in the form below, and our team replies with a configuration and a written quote within one business day.

What we supply

We supply the server with eight NVIDIA H200 NVL and their NVLink bridges in 4U or 5U, as well as two-card and four-card versions and the H200 NVL cards and bridges for servers you already run. Each card comes with its five-year NVIDIA AI Enterprise subscription on the same EU contract and invoice as the hardware.

FAQ