Buy a workstation with 2 to 4 NVIDIA RTX PRO 5000 Blackwell 72 GB cards in Europe

Who a workstation with RTX PRO 5000 72 GB cards suits

We build workstations with two, three or four NVIDIA RTX PRO 5000 Blackwell 72 GB cards to order and deliver them anywhere in the EU. Each card brings 72 GB of GDDR7 with ECC at 1,344 GB/s and 300 W, so a four-card tower has 288 GB of GPU memory at 1.2 kW of board power. You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day.

The 72 GB card suits a company whose models sit between 48 and 96 GB per copy. A 4-bit 70B model fits one card with room for a small team, a 32B model in FP8 leaves about 29 GiB of cache on each card, and gpt-oss-120b runs over a pair. Typical buyers are ML teams that keep one model per card for several projects, research departments that need a 70B model on site without a server room, and engineering offices that combine rendering or simulation with a local assistant. The card is also the only RTX PRO 5000 that NVIDIA lists for vGPU, which matters for a deskside host running graphics virtual machines on Red Hat KVM.

NVIDIA's datasheet gives the 48 GB and 72 GB versions the same 14,080 CUDA cores, 1,344 GB/s and 300 W in a dual-slot card of 4.4 × 10.5 inches with one 16-pin connector, three NVENC and three NVDEC engines. Where a model fits both, they run at the same speed; Puget Systems measured them less than 1 per cent apart in Topaz Video's Starlight Mini model, as our comparison of the RTX PRO 5000 with 48 GB or 72 GB reports. The extra 24 GB pays only when the workload uses it.

Two, three or four cards per workstation

The table sets the three build sizes side by side. 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. For the 72 GB card we take 71.7 GiB as the driver figure, in proportion to the 95.6 GiB reported on a 96 GB card.

NVIDIA RTX PRO 5000 Blackwell datasheet (memory, 1,344 GB/s, 300 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 57.5 GiB per card; NVIDIA publishes no figure for GDDR7.

Two cards are the usual start: one model per card, or one 70B model in FP8 over the pair. Three cards suit a team that runs a split model and keeps one card free for experiments. Four cards make sense when several models must stay loaded at once, or for two TP=2 pairs. The RTX PRO 5000 has no NVLink, so cards exchange data over PCIe 5.0, and our article on one model on several GPUs over PCIe or NVLink explains what that costs.

Towers whose makers list four RTX PRO 5000 cards

We checked the makers' own documents on 24 October 2026. Only some of them state a count for the RTX PRO 5000, and the 72 GB version is not always named.

Dell Precision 7960 configurator (Canada), Lenovo PSREF ThinkStation PX specification of 9 October 2026 and HP Z8 Fury G6i specifications, read 24 October 2026; Puget Systems Quad GPU LLM Server page, read by our research on 24 October 2026.

Dell's configurator for the Precision 7960 offers the card as a single, dual, triple and quad option, with 1,400 W or 2,200 W chassis. Two of its four GPU slots run at PCIe 4.0 x16, which halves the link bandwidth of those two cards to the processor; that matters for a model split across all four, less for one model per card. Lenovo's specification for the ThinkStation PX states "up to four" for the RTX PRO 6000 Max-Q and gives no count for the RTX PRO 5000, so the PX needs a confirmed count for the RTX PRO 5000 before an order. HP's current Z8 Fury G6i page lists four RTX PRO 6000 Max-Q and no RTX PRO 5000.

We supply the cards for these towers and also build four-card workstations to order on Threadripper PRO or Xeon W platforms, as our AI servers page describes. Send us the tower your company standardises on, or the platform you prefer, and our team confirms the card count, slots 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. For four 72 GB cards that is 24 cores and 576 GB; two cards need 12 cores and 288 GB, which we round to 384 GB as eight 48 GB modules to fill eight channels. AMD lists the Threadripper PRO 9975WX with 32 cores, 350 W, 128 PCIe 5.0 lanes and eight memory channels, so four cards at x16 take 64 lanes and leave the rest for NVMe and the network card.

What fits on one card and on the whole build

The counts below apply our users-per-card rule: conversations of 8,192 tokens that fit beside the weights in an FP8 KV cache, at 1.25 GiB per conversation for Llama 3.3 70B, 1.0 GiB for Qwen3-32B and 0.625 GiB for Qwen3-14B. 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 61.5 GiB usable per card and ECC off; weights from the published checkpoints as in our model guides: Qwen3-14B FP8 15.2 GiB, Qwen3-32B FP8 32.0 GiB, Llama 3.3 70B NVFP4 39.8 GiB and FP8 67.7 GiB, gpt-oss-120b 60.8 GiB. Memory arithmetic, not measured throughput.

gpt-oss-120b loads on one card only on paper: 0.7 GiB is left, which holds no conversation and disappears if ECC takes memory. Llama 3.3 70B in FP8 needs a pair. As a planning aid, the example values in our guide to a private ChatGPT server by company size put about 20 requests in flight at peak for 500 staff and 80 for 2,000, so four cards running a 70B model in FP8 as two pairs cover the second case by memory.

Speed per user follows memory bandwidth. By our arithmetic, one card generating Llama 3.3 70B in NVFP4, which reads about 40.6 GB per token, is held to about 33 tokens per second for a single user; that is a ceiling from the specifications, not a measurement. By default vLLM's source disables its custom all-reduce on more than two PCIe-only GPUs, so a TP=4 layout falls back to NCCL, while TP=2 pairs keep the faster path; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576), as our article on RTX PRO 6000 tensor parallelism without NVLink describes. We found no published measurement of tensor-parallel scaling on four RTX PRO 5000 cards. Our guide to four RTX PRO 5000 72 GB in one workstation covers the layouts in more depth.

MIG halves of 36 GB and vGPU on Red Hat KVM

NVIDIA's product page gives the RTX PRO 5000 up to two MIG instances, and the datasheet sizes them at 36 GB each on the 72 GB card, so four cards could offer eight isolated 36 GB GPUs. NVIDIA's MIG user guide, updated on 13 October 2026, lists only the 48 GB version so far, so read the profile names on the delivered card with nvidia-smi mig -lgip. MIG needs Linux, driver 575.51.03 or later and the display mode switched to compute; NVIDIA staff named DisplayModeSelector 1.76 or later for the 72 GB card in May 2026. The switch turns the card's display outputs off. A 36 GB half holds a 32B model in 4-bit with cache for about nine conversations, but not the same model in FP8.

The 72 GB card is on NVIDIA's vGPU list from vGPU 20.2 of August 2026, a production branch that NVIDIA supports until March 2027. Of the hypervisor pages we read on 24 October 2026, only NVIDIA's validated platforms for Red Hat Enterprise Linux with KVM list the card, with the graphics editions vWS and vPC, time-sliced or MIG-backed; the VMware vSphere and Ubuntu support matrices of 29 September 2026 do not list it. For virtual machines on vSphere, the RTX PRO 6000 Server Edition and the RTX PRO 4500 Server Edition are the cards to plan; our guide to MIG and vGPU: VMs per GPU compares them.

Power on 230 V and the room

All three builds stay on one 230 V circuit by our estimate: the cards, a 350 W processor and 15 per cent for memory, drives, fans and power-supply losses.

Our estimate from NVIDIA's 300 W board power and the 350 W TDP of the Threadripper PRO 9975WX, plus 15 per cent, at full load; a 16 A circuit planned at 80 per cent carries 2.94 kW. The same figures as for 300 W Max-Q builds.

A four-card tower turns about 2 kW into heat, roughly 6,800 BTU/h, which belongs in a technical room or an office with air conditioning. Large workstation supplies reach their full rating only at 200 V or more, which every European socket provides, and Dell's 2,200 W supply uses a C19 cord rather than the 10 A C13 cord. The cards are blower designs that exhaust through the slot bracket, so leave free space behind the tower. Our article on RTX PRO 6000 Max-Q power on 230 V covers circuits, cords and UPS sizing for the same 300 W class, and our guide to a GPU server in an office covers noise and heat in shared rooms.

RTX PRO 5000 72 GB or RTX PRO 6000 Max-Q

Both cards are 300 W blower designs in two slots, so the same four-card tower takes either. On the datasheets the Max-Q build is ahead on memory and bandwidth for the same board power.

NVIDIA datasheets, MIG user guide (13 October 2026) and vGPU supported GPU list (2 October 2026); conversations by our users-per-card rule.

A Max-Q card holds Llama 3.3 70B in FP8 on its own; the 72 GB card does not. MIG gives no advantage to the RTX PRO 5000 either: a Max-Q splits into two 48 GB instances, larger than the 36 GB halves, or into four of 24 GB. What the RTX PRO 5000 72 GB offers and the Max-Q lacks is vGPU on Red Hat KVM. If your models fit 72 GB per copy and that point does not matter, the decision comes down to the price of the two configurations, which we quote side by side. Our pages to buy a 4× RTX PRO 6000 Max-Q workstation and on two RTX PRO 5000 72 GB or one RTX PRO 6000 Max-Q carry the comparison further. When the models need less than 48 GB per copy, our page to buy an RTX PRO 5000 48 GB workstation is the other option.

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, with MIG instances set up if you plan them. The card has compute capability 12.0 and NVIDIA's datasheet lists CUDA 12.8, so containers built on older CUDA releases need checking first. 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: models, users, data, 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. The workstation 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 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 you can repeat on delivery.

What to send us for a quote

• The models, with quantisation and context length, and whether one model is split or each card runs its own
• How many people use the machine at peak, and whether through one service, MIG instances or virtual machines
• Other work on the same machine: rendering, simulation, fine-tuning, data preparation
• 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