Buy a workstation with 2 to 4 NVIDIA RTX PRO 4000 Blackwell cards in Europe

Who a workstation with RTX PRO 4000 cards suits

You can buy a workstation with two, three or four NVIDIA RTX PRO 4000 Blackwell cards from us, built to order and delivered anywhere in the EU. Each card has 24 GB of GDDR7 with ECC at 672 GB/s and 145 W in a single slot, so four cards give 96 GB of GPU memory at 580 W in four slot widths. Our team sizes the configuration to your workload and budget and replies within one business day.

The RTX PRO 4000 is the card for many small GPUs in one machine. A design office gives four CAD or visualisation users a whole card each from one tower in the server room or under a desk. A development team runs an embedding model, a reranker, speech-to-text and a 14B chat model on four separate cards for a RAG pilot. A training department keeps four isolated GPUs for course participants, one per virtual machine with passthrough. A video team uses the eight encoders and eight decoders of four cards for transcoding beside AI work.

The card is not built for one large model. A 32B model in FP8 needs two cards, a 70B model all four, and every split sends traffic over PCIe. For a single large model, one RTX PRO 6000 Max-Q is the better buy, as our comparison of two RTX PRO 4000 or one RTX PRO 6000 Max-Q shows.

The card: what NVIDIA publishes

NVIDIA RTX PRO 4000 product page and professional desktop GPU overview, read 24 October 2026; CUDA cores, bus width, connector and the datasheet figures from NVIDIA's datasheet of June 2026, as cited in our comparison of the RTX PRO 4000 Blackwell and the RTX 4000 Ada.

NVIDIA publishes two sets of compute figures for this card: the June 2026 datasheet gives 37 TFLOPS of FP32 and 1,178 AI TOPS, the product page 40 and 1,290. We plan on the datasheet. Both give 145 W; Lenovo's ThinkStation PX specification lists the card at 140 W. An adapter from one 8-pin cable to the 16-pin connector comes with the card, as our RTX PRO 4000 comparison records, so four cards need four free 8-pin leads or four native 16-pin cables. NVIDIA describes the cooler only as active.

Single slot is the point of the card. In NVIDIA's desktop GPU overview the RTX PRO 4000 is the only single-slot card; the RTX PRO 4500, 5000 and 6000 Max-Q are listed as dual slot. Our comparison of single-slot GPUs for AI sets it against the single-slot RTX PRO 4500 Server Edition for servers, and the low-profile RTX PRO 4000 SFF at 70 W is a different card again, covered in our 70 W GPU comparison.

Two, three or four cards per workstation

Usable memory follows our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card for activations and the runtime. The driver reports 95.6 GiB on a 96 GB card; for the 24 GB card we scale that by capacity, an assumption, which gives 18.5 GiB.

NVIDIA product page (24 GB, 672 GB/s, 145 W); usable memory and sums by our estimate, ECC off. If ECC reserved the 6.25 per cent that NVIDIA's CUDA Best Practices Guide gives for GDDR memory, usable memory would be about 17.2 GiB per card; NVIDIA publishes no figure for GDDR7.

The four cards remain four pools of 24 GB. Each card addresses only its own memory, so 96 GB of GPU memory is not one 96 GB GPU: a model larger than about 18.5 GiB with its cache is split over cards, and every generated token then crosses PCIe. Two cards suit two users or two services; four suit four users, four services, or one split model of up to about 70 GiB. Our page on four RTX PRO 4000 in one workstation covers the four-card layouts in more detail.

Towers whose makers list four RTX PRO 4000

We checked the makers' own documents on 24 October 2026.

Dell Precision 7960 configurator (Canada), Lenovo PSREF ThinkStation PX specification of 9 October 2026 and HP Z8 Fury G6i specifications, read 24 October 2026.

Two of the Precision 7960's four GPU slots run at PCIe 4.0 x16, which halves the link bandwidth of those two cards to the processor; for one model per card that stays in memory it matters little, for a model split across all four it matters more. Lenovo's specification names "up to four" only for the RTX PRO 6000 Max-Q and the RTX 6000 Ada, so a four-card PX with the RTX PRO 4000 needs a confirmed count before an order.

Single-slot cards also fit boards where the slots sit next to each other, which leaves room for a network card or NVMe adapter beside four GPUs. Four cards side by side draw their air through narrow gaps, so we check the tower's airflow for that layout before quoting. We build four-card workstations to order on Threadripper PRO or Xeon W platforms, as our AI servers page describes, and supply the cards for towers your company already runs. Send us the tower model or the number of users per machine, and our team confirms slots, cables and power before quoting.

What the configuration contains

Our AI servers page (platforms and components); memory and cores by NVIDIA's configuration guide for NVIDIA-Certified Systems (30 September 2026), written for servers and applied here by the same arithmetic.

NVIDIA's guide asks for at least six physical cores per GPU and system memory of at least twice the GPU memory. Four 24 GB cards make 24 cores and 192 GB, filled as eight 24 GB modules on eight channels; two cards need 12 cores and 96 GB, which we round to 128 GB as eight 16 GB modules. Where four people work on the machine at once, through virtual machines or remote sessions, their applications add processor and memory demand of their own, and we size that from the applications you name. AMD lists the Threadripper PRO 9975WX with 32 cores, 350 W, 128 PCIe 5.0 lanes and eight memory channels.

What runs on one card and on the whole build

The counts apply our users-per-card rule: conversations of 8,192 tokens that fit beside the weights in an FP8 KV cache, at 0.625 GiB per conversation for Qwen3-14B, 1.0 GiB for Qwen3-32B and 1.25 GiB for Llama 3.3 70B. gpt-oss-120b is counted at 32K tokens with a 16-bit cache of 1.125 GiB.

Our estimates by the users-per-card rule with 18.5 GiB usable per card, ECC off; weights Qwen3-14B FP8 15.2 GiB, Qwen3-32B NVFP4 19.3 GiB and FP8 32.0 GiB, Llama 3.3 70B NVFP4 39.8 GiB and FP8 67.7 GiB, gpt-oss-120b 60.8 GiB, as in our model guides. Memory arithmetic, not measured throughput.

One card holds a 14B model in FP8 with room for five conversations, which suits one person or a small team; gpt-oss-20b, at 12.8 GiB in OpenAI's MXFP4 release, leaves 5.7 GiB on a card by our rule. A 32B model in 4-bit does not fit one card by our rule, because NVIDIA's NVFP4 checkpoint is 19.3 GiB. With the example values of our guide to a private ChatGPT server by company size, about 20 requests are in flight at peak for 500 staff; four copies of Qwen3-14B cover that by memory, and two TP=2 pairs leave a margin.

Speed per conversation follows bandwidth. By our arithmetic, one card generating Qwen3-14B in FP8, which reads about 16.3 GB per token, is held to about 41 tokens per second for one user; Qwen3-32B in FP8 as a TP=2 pair reaches a ceiling of about 39 before the cost of exchanging results over PCIe. These are ceilings from the specifications, not measurements, and we found no published measurement of tensor-parallel scaling on two or four RTX PRO 4000 cards. Our article on one model on several GPUs over PCIe or NVLink explains the cost of a split, and our guide to RTX PRO 6000 P2P on Linux lists the BIOS and kernel checks for peer-to-peer access, which apply to any RTX PRO cards.

Graphics, video and CAD

Puget Systems measured the RTX PRO 4000 Blackwell in its December 2025 roundup at 83 per cent faster than the RTX 4000 Ada in V-Ray GPU and 40 per cent faster in Blender, and its engineering roundup found 12 to 21 per cent for the midrange Blackwell cards in SOLIDWORKS, as our RTX PRO 4000 comparison records. Those are single-card results. Each card keeps its four DisplayPort 2.1b outputs and its two encoders and two decoders, so a four-card tower has sixteen display outputs and eight of each video engine.

No MIG, no vGPU: one card per user

NVIDIA's MIG user guide, updated on 13 October 2026, does not list the RTX PRO 4000, and NVIDIA's list of GPUs supported by vGPU software, updated on 2 October 2026, does not include it. A card therefore cannot be split into isolated instances or shared between virtual machines with vGPU profiles. It can be passed whole to one virtual machine, so a four-card workstation gives four virtual machines a GPU each; plan a passthrough test with your hypervisor before relying on it.

That makes the card count the user count for anything that needs its own GPU. If more people need isolated GPU memory than a tower has cards, an RTX PRO 6000 Max-Q in four 24 GB MIG instances, or a rack server with vGPU on the RTX PRO 4500 Server Edition, does the job; our guide to MIG and vGPU: VMs per GPU counts the seats.

Power on 230 V and the room

By our estimate every build here stays far inside one 230 V circuit: the cards, a 350 W processor and 15 per cent for memory, drives, fans and power-supply losses.

Our estimate from NVIDIA's 145 W board power and the 350 W TDP of the Threadripper PRO 9975WX, plus 15 per cent, at full load; a 230 V 16 A circuit planned at 80 per cent carries 2.94 kW.

The processor takes a larger share here than in our other multi-card builds: in the four-card tower it draws more than half as much as all four cards. A tower for two users with lighter applications can take a processor with a lower TDP, which brings the figure down. A four-card tower turns about 1.1 kW into heat, roughly 3,700 BTU per hour, which an office with ventilation handles more easily than the 1.8 to 2 kW of a four-card 300 W build. Our guide to a GPU server in an office covers noise and heat in shared rooms.

Four RTX PRO 4000 or fewer, larger cards

NVIDIA product pages and datasheets; usable memory by our users-per-card rule, as in our buy pages for these cards.

One RTX PRO 6000 Max-Q has the same 96 GB as four RTX PRO 4000 in one pool, at about half the board power, and splits into four 24 GB MIG instances; for language models it holds more conversations of every model in our table. Four RTX PRO 4000 stay ahead where four users each need a whole GPU with display outputs and graphics, where the tower has single slots free and no room for a dual-slot card, and where four separate small services should not share one card. Our pages to buy an RTX PRO 4500 workstation and to buy an AI workstation for LLMs cover the neighbouring builds.

Software on delivery

On request the workstation arrives with a Linux release or your company's image, the NVIDIA driver branch you will run, CUDA and a container runtime with the NVIDIA Container Toolkit. NVIDIA's CUDA GPU list places the RTX PRO 4000 Blackwell under compute capability 12.0, like the other RTX PRO Blackwell cards. As of October 2026 the R580 branch has long-term support to June 2028, and NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as fixed versions; our guide to NVIDIA driver branches and CUDA versions explains how to pin the branch.

How ordering works with us in the EU

1. You describe the workload: users, applications, models, the room and the timeline. Sizing is free of charge.
2. We propose the configuration and check that cards, tower, 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. We assemble the workstation and test it 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 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.

Our GPU server acceptance testing guide lists checks for your own acceptance test on delivery.

What to send us for a quote

• How many people or services use the machine, and whether each needs a whole GPU
• The applications: CAD, rendering, video, and the models with precision and context length
• Whether users sit at the machine with monitors or reach it through virtual machines or remote sessions
• The tower your company standardises on, if any, and the operating system or hypervisor
• The room: circuit rating, cooling and where the tower will stand
• Delivery site, invoicing entity and timeline

Write these points in the form below, and we reply with a configuration and a written quote within one business day.

FAQ