Buying an AI workstation for LLMs in Europe: 2 to 4 NVIDIA RTX PRO GPUs for company teams

AI workstations with two to four GPUs

We sell AI workstations for large language models with two, three or four NVIDIA RTX PRO Blackwell cards, built to order and delivered anywhere in the EU. For most company teams the right start is two RTX PRO 6000 Blackwell Max-Q cards with 96 GB each, in a tower prepared for four; teams with models up to about 32B, or with less room for power and heat, choose the RTX PRO 5000, 4500 or 4000. You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day.

A workstation suits work that one team does next to its machine. An ML team builds and tests LLM services on it. A department runs its first private assistant or a RAG search over its own documents before it decides on a server. A research lab keeps its data in the building and gives several people isolated GPU slices with MIG. A design or engineering office runs a code assistant and an embedding model beside CAD work. When the service runs continuously for several departments or for customers, a rack server is the better buy; our page to buy a GPU server for LLMs covers that range.

This page is the overview of our workstation range. It compares the configurations, shows what each runs by one sizing rule, and links to a buy page for every card at the end.

The configurations we build

All cards below are NVIDIA RTX PRO Blackwell workstation cards on PCIe 5.0 x16 without NVLink, so each card is its own memory pool and a model larger than one card is split over PCIe. Usable memory follows our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card. The driver reports 95.6 GiB on a 96 GB card; for the smaller cards we scale that by capacity, an assumption.

NVIDIA datasheets and product pages read on 24 October 2026 (Max-Q 96 GB, 1,792 GB/s, 300 W; RTX PRO 5000 1,344 GB/s, 300 W; RTX PRO 4500 32 GB, 896 GB/s, 200 W; RTX PRO 4000 24 GB, 672 GB/s, 145 W); usable memory is our estimate by the users-per-card rule.

The RTX PRO 6000 Max-Q is the 300 W edition of the RTX PRO 6000 with a blower cooler; NVIDIA's product page states that it "enables up to four GPUs in a single system". The RTX PRO 5000 is a 300 W dual-slot card with 48 or 72 GB, the RTX PRO 4500 a 200 W dual-slot card with 32 GB, and the RTX PRO 4000 a 145 W single-slot card with 24 GB. We also build three-card configurations on the same rules.

Around the cards we size the platform by NVIDIA's configuration guide for NVIDIA-Certified Systems, updated on 30 September 2026. The guide is written for servers; we apply its rules to workstations by the same arithmetic.

Our AI servers page (platforms and components); NVIDIA's configuration guide for NVIDIA-Certified Systems, read on 24 October 2026.

Which configuration for which team

The largest model you plan to run decides first, then the number of people and the power at the desk.

• One team building LLM services, a 70B model for a department, fine-tuning: two RTX PRO 6000 Max-Q in a tower prepared for four. One 96 GB card holds Llama 3.3 70B in FP8.
• Models of 150 to about 300 GiB, or several 70B copies at once: four RTX PRO 6000 Max-Q, 384 GB. Qwen3-235B-A22B-2507 in FP8, about 236 GB, fits four cards and not two.
• 70B models in 4-bit, QLoRA at 300 W per card, or vGPU on Red Hat KVM: two to four RTX PRO 5000 72 GB, the only RTX PRO workstation card on NVIDIA's vGPU list.
• Models of 14B to 32B, two people per card through MIG: RTX PRO 5000 48 GB.
• RAG, embeddings and coding assistants with models up to about 14B, beside CAD: RTX PRO 4500, two to four cards.
• Small models, compact towers, many cards in few slots: RTX PRO 4000, the single-slot card.

Two other cards suit narrower cases. The RTX PRO 6000 Workstation Edition at 600 W is the fastest single card for long prompts and fine-tuning, and suits one or two per tower. The RTX 6000 Ada, 48 GB at 300 W, extends existing Ada fleets and runs vGPU desktops. Our guide to the best workstation GPU for AI in 2026 compares every card one by one.

Models and conversations per configuration

The table counts conversations of 8,192 tokens held at the same moment, with an FP8 KV cache: 1.25 GiB per conversation for Llama 3.3 70B, 1.0 GiB for Qwen3-32B and 0.625 GiB for Qwen3-14B. Each cell names the layout it counts: usually the one with the most copies that fits, and for four RTX PRO 4000 one TP=4 copy, since two TP=2 copies of Qwen3-32B would leave room for only 5 conversations each.

Our estimate by the users-per-card rule, memory-bound, not measurements: (usable memory of the copy − weights) ÷ cache per conversation, rounded down. Weights 67.7, 39.8, 32.0 and 15.2 GiB, as in our model guides.

Fewer copies hold more conversations, because each copy stores its own weights: two Max-Q run Llama 3.3 70B in FP8 as one TP=2 copy for 78 conversations, against 24 as two copies. The counts are requests in flight, which is far fewer than the people who use a service; with the example values of our guide to a private ChatGPT server by company size, 500 employees produce about 20 requests in flight at peak. Our comparison of several RTX PRO 5000 or 4500 or fewer RTX PRO 6000 shows the same builds in more detail.

We build every configuration in the table; send us the models and the number of people who will use them, and our team replies with the build that fits and a written quote.

Lanes and slots for two to four cards

Each card takes a PCIe 5.0 x16 link, so four cards need 64 lanes. Both platforms we build on supply that: AMD lists the Threadripper PRO 9975WX with 32 cores, 128 PCIe 5.0 lanes and eight DDR5 channels, and Intel lists its Xeon 698X with 128 PCIe 5.0 lanes and eight memory channels. How the board wires its slots matters as much as the lane count. In a review of 23 June 2026, StorageReview ran two Max-Q cards in an HP Z8 Fury G6i, with every GPU slot at PCIe 5.0 x16, and in a Dell Precision 7875 whose second card ran at PCIe 4.0 x16; serving Llama 3.1 8B in FP8 with vLLM at batch 256, the HP reached 23,004 tokens per second and the Dell 16,833.

Dual-slot cards need double-slot spacing, and Puget Systems advises a free slot between Max-Q cards because the memory on the back of the board needs air. The single-slot RTX PRO 4000 fits four cards into four adjacent slots. A model split over several cards also depends on peer-to-peer access between them, which the IOMMU and ACS settings in the BIOS and kernel decide; our guide to RTX PRO 6000 P2P on Linux lists the checks, which apply to any RTX PRO card.

Built to order, or cards for a tower you run

There are two ways to buy from us.

• Built to order. The workstation is assembled on a Threadripper PRO or Xeon W platform and tested before delivery, and we deliver it with the software installed on request.
• Cards for your tower. If your company standardises on a maker's workstation, we supply the cards for it after checking the power supply, cables and slots. Lenovo's specification for the ThinkStation PX of 9 October 2026 lists up to four RTX PRO 6000 Max-Q and also offers the RTX PRO 5000, 4500 and 4000; HP lists up to four Max-Q for the Z8 Fury G6i, and Dell's configurator for the Precision 7960 offers one to four.

Our pages on two and four RTX PRO 6000 Max-Q list the towers with their power supplies.

Power and room on 230 V

By our estimate, a workstation draws its cards' board power, a 350 W processor of the Threadripper PRO 9975WX or Xeon 698X class, and 15 per cent on top for memory, drives, fans and power-supply losses. A measurement on the delivered system confirms the figure.

Our estimate: NVIDIA board power, plus 350 W for the processor (AMD's default TDP for the Threadripper PRO 9975WX, Intel's base power for the Xeon 698X), plus 15 per cent. A 230 V 16 A circuit delivers 3.68 kW, planned at 80 per cent as 2.94 kW.

• Circuit. Every build in the table fits one ordinary 230 V 16 A circuit; check what else shares it. Four-card builds with more memory and drives can reach about 2 kW, still within the circuit. Our article on Max-Q power on 230 V covers circuits and supplies.
• Cord and supply. IEC C13 cords are rated at 10 A, about 2.3 kW at 230 V; the large supplies of four-card towers use C19 cords rated at 16 A. Workstation supplies reach full rating on high-line input: HP lists 1,700 W at 200 V or more for the Z8 Fury G6i.
• Heat and noise. All of the power ends as heat: 1.8 kW is about 6,100 BTU per hour for as long as jobs run. No maker publishes acoustic figures for a multi-GPU tower under GPU load. Our guide to a GPU server in an office covers rooms shared with people.

Software, MIG and vGPU

On request the workstation arrives with the operating system, the NVIDIA driver, CUDA and a container runtime with the NVIDIA Container Toolkit installed, so the team can start vLLM, SGLang or Ollama on the first day. 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; our article on NVIDIA driver branches and CUDA versions explains the choice. Open-source serving engines on bare metal need no NVIDIA licence.

Sharing the cards between people works through MIG on most of them.

NVIDIA MIG user guide (supported GPUs, last updated 13 October 2026) and Max-Q datasheet; NVIDIA vGPU supported-GPU list, last updated 2 October 2026, both lists read on 24 October 2026; RTX PRO 5000 72 GB MIG size from a board-partner datasheet; Lenovo Press LP2443 (26 May 2026) for the workstation RTX PRO 4500.

MIG on the RTX PRO workstation cards needs Linux, driver 575.51.03 or later, a current vBIOS and the card in compute display mode, which turns its display outputs off; a MIG workstation runs headless or keeps a small extra card for the screen. For the workstation RTX PRO 4500 the sources disagree: NVIDIA's MIG guide lists an "RTX PRO 4500 Blackwell" with two 16 GB instances without naming an edition, and Lenovo's product guide for the RTX PRO 4500 Blackwell Workstation Edition of 26 May 2026 lists MIG as "No support", so we do not plan a build on MIG with that card. Our RTX PRO 6000 MIG profiles list what fits each instance. For virtual machines with vGPU, the RTX PRO 5000 72 GB on Red Hat KVM and the RTX 6000 Ada are the workstation-class options; vGPU hosts at scale belong on a server with the RTX PRO 6000 Server Edition.

Workstation or rack server

A rack server is the better buy when the service runs continuously for several departments or for customers, when nobody should sit next to the machine, or when many virtual machines need vGPU. Our entry server has two RTX PRO 6000 Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives, with redundant power supplies and out-of-band management. The Max-Q belongs in workstations: in a rack server it fits only a chassis whose maker lists it, and for continuous operation in a rack we recommend the Server Edition, as our page on the Max-Q in a rack server explains. AI workstation vs AI server compares the form factors.

Ordering a workstation from us

The steps are the same for one workstation or several for different teams.

1. You describe the workload: models, users, data, the room and the timeline. Sizing is free of charge.
2. We propose a configuration and check that cards, platform, power supply and cooling fit the room and the circuit.
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 workstation is assembled and tested before delivery.
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.

Details for an exact quote

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

• The models you plan to run, with precision and context length
• How many people will use the machine, and whether they share it through one service or through MIG instances
• Other work on the same machine: fine-tuning, embeddings, data preparation, CAD or rendering
• Whether the cards should drive monitors, and the operating system
• The tower model your company standardises on, if any
• The room: circuit rating, cooling and where the machine will stand
• Whether you plan to add cards later, delivery site and invoicing entity

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

Buy pages for each workstation configuration

Each card has its own buy page with the parts list, variants and ordering details, and a technical page on what it runs.

Two comparisons help with the choice between neighbouring builds: two RTX PRO 4000 or one Max-Q and two RTX PRO 5000 72 GB or one Max-Q.

What we supply

We build AI workstations with two to four RTX PRO 6000 Max-Q, RTX PRO 5000, 4500 or 4000 cards to order, and supply every card on its own for towers you already run, on one EU contract and invoice with manufacturer warranty on every component, handled through us. Our engineering partner Vixen.UNO deploys models, RAG and the platform on top as our Private AI/ML service; tell us which models and how many users you plan for, and we size workstation and platform together.

FAQ