Buy a workstation with 2 to 4 NVIDIA RTX 6000 Ada Generation cards in Europe

When an RTX 6000 Ada workstation still makes sense

You can buy a workstation with two, three or four NVIDIA RTX 6000 Ada Generation cards from us, built to order and delivered anywhere in the EU; the card is available to order. Each card has 48 GB of GDDR6 with ECC at 960 GB/s and 300 W, so four cards give 192 GB of GPU memory at 1.2 kW of board power. Our team sizes the configuration to your workload and budget and replies within one business day.

For a new AI workstation without constraints, the Blackwell cards at the same 300 W are the better choice, and we say so further down. The RTX 6000 Ada is still the right card in three situations:

• An existing Ada fleet. A company that runs Ada workstations with one validated image, driver and set of containers adds the same card and keeps one configuration. Container images built only for compute capability 8.9 run on it unchanged.
• Certified applications. Where a CAD, simulation or visualisation vendor certifies specific cards for its software, its list decides. If that list names the RTX 6000 Ada and not yet the RTX PRO Blackwell card you would buy instead, Ada keeps you inside the vendor's support.
• vGPU on a workstation-class card. NVIDIA lists the RTX 6000 Ada for vGPU from release 15.2, including VMware vSphere. Of the current RTX PRO Blackwell workstation cards only the RTX PRO 5000 72 GB is on that list, and on Red Hat KVM only.

A fourth case is mixing. Some teams add Blackwell cards next to their Ada cards; our guide to mixing Ada and Blackwell GPUs explains what one driver covers and why each model should stay on cards of one generation.

The card and the builds

NVIDIA RTX 6000 Ada Generation datasheet (document 25046) and product page, read on 24 October 2026; 16-pin connector from our RTX 6000 Ada vs L40S comparison.

Two cards give 96 GB at 600 W, three 144 GB at 900 W, four 192 GB at 1,200 W. With 300 W, a blower cooler and a 4.4 × 10.5 inch dual-slot board, the RTX 6000 Ada has the same power and dimensions as the RTX PRO 5000 and the RTX PRO 6000 Max-Q, so a tower built for four Ada cards is mechanically and electrically in the same class as one for four of those Blackwell cards. Check the maker's list for the exact card in either case.

Towers whose makers list four RTX 6000 Ada

Lenovo PSREF ThinkStation PX specification of 9 October 2026, Dell Precision 7960 configurator (Canada) and HP Z8 Fury G6i specifications, read on 24 October 2026; Supermicro from our RTX 6000 Ada vs L40S comparison, read on 10 October 2026.

Lenovo's specification for the ThinkStation PX states "Supports up to four NVIDIA RTX 6000 Ada Generation" and marks the card as available in specific countries and regions. The PX carries two Xeon Scalable processors with two PCIe 5.0 x16 slots on each. Dell's configurator offers the Precision 7960 with one, two, three or four RTX 6000 Ada, with 1,400 W or 2,200 W chassis. Dell's specifications list two full-height PCIe 5.0 x16 slots and two full-height PCIe 4.0 x16 slots, so in a four-card build at least two cards sit in Gen4 slots; that costs an Ada card nothing, because the card itself is PCIe 4.0. Supermicro's SYS-741GE-TNRT takes up to four double-width GPUs on two redundant 2,000 W power supplies and lists the RTX 6000 Ada among them. HP's current Z8 Fury G6i page lists only the RTX PRO 6000 Max-Q for four cards.

We supply the cards for these towers and build four-card workstations to order on Threadripper PRO or Xeon W platforms. Send us the tower model and the cards your teams run today, and our team checks slots, power and cables before quoting.

What we put in the workstation

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.

By that rule four 48 GB cards need at least 24 cores and 384 GB of system memory, two cards 12 cores and 192 GB. Each Ada card takes 16 PCIe 4.0 lanes, and a 128-lane workstation processor has room for four cards plus NVMe and network.

What two and four RTX 6000 Ada run

The counts follow our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card, less the weights, divided by the FP8 KV cache of one 8,192-token conversation: 0.625 GiB for Qwen3-14B, 1.0 GiB for Qwen3-32B, 1.25 GiB for Llama 3.3 70B. GDDR6 cards lose 6.25 per cent of memory with ECC enabled, as NVIDIA's CUDA Best Practices Guide states, so the table gives both cases.

Our estimates: 37.3 GiB usable per card with ECC on, 40.0 GiB with ECC off in brackets, from 44.8 and 47.8 GiB of driver-reported memory as in our 48 GB GPU comparison; weights Qwen3-14B FP8 15.2 GiB, Qwen3-32B FP8 32.0 GiB, Llama 3.3 70B FP8 67.7 GiB. Memory arithmetic, not measured throughput.

FP8 is the precision to plan with on Ada. The card has FP8 Tensor Cores but no FP4, so NVFP4 checkpoints belong on Blackwell; for other 4-bit formats, check that your serving engine supports the checkpoint on Ada. A 70B model in FP8 does not fit one card and is cramped on a pair; it becomes usable over all four cards. Models up to 14B run one copy per card with room to spare, which suits a team that keeps several small models loaded.

Bandwidth sets the speed per user. By our arithmetic, Qwen3-32B in FP8 reads about 34.4 GB per token, which limits one RTX 6000 Ada to about 28 tokens per second for a single user, against 39 on an RTX PRO 5000 and 52 on an RTX PRO 6000 Max-Q. These are ceilings from the specifications, not measurements. Splitting a model over four cards runs over PCIe 4.0, and by default vLLM disables its custom all-reduce on more than two PCIe-only GPUs, so TP=4 falls back to NCCL; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576). Our comparison of 48 GB GPUs: RTX 6000 Ada, RTX 5880 Ada, RTX PRO 5000 and L40S sets the 48 GB cards side by side.

vGPU, NVIDIA AI Enterprise and drivers

NVIDIA's list of GPUs supported by vGPU, updated on 2 October 2026, shows the RTX 6000 Ada from release 15.2. NVIDIA's support matrix for VMware of 29 September 2026 lists it for VCF 9.1, VCF 9.0 and ESXi 8.0 from Update 3 P06, with RTX Virtual Workstation, Virtual PC and Virtual Applications. The card ships with its display outputs on, and its datasheet notes that the display ports are not active when vGPU software is used; NVIDIA's release notes require display-off mode for vGPU, set with the displaymodeselector tool, so every card bound for a vGPU host needs that step. The RTX 6000 Ada has no MIG, so virtual machines share a card in time slices. Our article on GPUs in VMware vSphere: passthrough or vGPU covers the host side.

The NVIDIA AI Enterprise support matrix for release 7.8 lists the RTX 6000 Ada Generation, on the condition that the card runs in a system listed on NVIDIA's certified systems page; ask for the listing of the exact workstation model before you order. NVIDIA's data-centre drivers 580.178.04 and 595.91.07 of 3 August 2026 list the RTX 6000 Ada beside the RTX PRO Blackwell cards, so one installed driver covers both generations. Ada support arrived with CUDA 11.8, and the card has compute capability 8.9.

Power on 230 V

Our estimate at full load from NVIDIA's 300 W board power and a 350 W workstation processor such as the AMD Threadripper PRO 9975WX (32 cores, default TDP 350 W), plus 15 per cent for memory, drives, fans and power-supply losses; the upper figure for four cards allows for more memory and drives. A 16 A circuit planned at 80 per cent carries 2.94 kW.

The figures are the same as for four RTX PRO 5000 or four Max-Q, since all three cards draw 300 W. One 230 V 16 A circuit carries one four-card tower, and its roughly 2 kW of heat calls for a technical room or an air-conditioned office. Large workstation supplies reach their full rating at 200 V or more; Dell's 2,200 W unit for the Precision 7960 uses a C19 cord, and Lenovo's power configurator for the PX recommends a dedicated circuit for each power supply. Our article on RTX PRO 6000 Max-Q power on 230 V covers the same 300 W class in detail.

When Blackwell is the better buy

NVIDIA datasheets, MIG user guide (13 October 2026) and vGPU supported GPU list (2 October 2026); conversations of 8K tokens by our users-per-card rule, RTX 6000 Ada with ECC on (off in brackets).

All three cards draw 300 W and take two slots, so the comparison is not about the power supply. Choose Blackwell when the model needs more than 48 GB per copy, when you plan NVFP4 checkpoints, when MIG should split cards between users, or when speed per user matters: the RTX PRO 5000 48 GB has the same memory with 40 per cent more bandwidth, and the RTX PRO 6000 Max-Q has twice the memory and holds Llama 3.3 70B in FP8 on one card. In Puget Systems' roundup of 18 December 2025 the RTX PRO 5000 48 GB was 11 per cent ahead of the RTX 6000 Ada in V-Ray GPU. Lenovo states up to four Max-Q for the PX, and Dell's Precision 7960 configurator offers four RTX PRO 5000 or four Max-Q, so a company standardised on either tower can move to Blackwell without changing the tower model.

Our comparison of the RTX PRO 6000 Blackwell and the RTX 6000 Ada goes through the generation change card by card, and four RTX 6000 Ada or four RTX PRO 6000 Max-Q compares the four-card builds. The Blackwell workstations are on our pages to buy an RTX PRO 5000 48 GB workstation and to buy a 4× RTX PRO 6000 Max-Q workstation; the smaller Ada cards are covered on our page to buy an RTX Ada workstation.

Software on delivery

On request the workstation arrives with your operating system or company image, the NVIDIA driver branch your Ada fleet runs, CUDA and a container runtime with the NVIDIA Container Toolkit. NVIDIA recommends its open GPU kernel modules for Ada, and requires them for Blackwell, so a workstation that may later take a Blackwell card is best set up on the open modules from the start. 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 covers pinning the branch.

Ordering in the EU

1. You describe the workload and the fleet: cards and drivers in use, applications and their certification, users, room and timeline. Sizing is free of charge.
2. We propose the configuration, with a Blackwell alternative where it fits, and check cards, tower, power supply and cooling.
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.

What to tell us for a quote

• The Ada cards, driver branch and images your teams run today
• The applications and whether their vendor certifies specific cards
• Models and users, or the vGPU profiles and hypervisor if virtual machines are planned
• The tower your company standardises on, if any
• Circuit rating, cooling and where the tower will stand
• Delivery site, invoicing entity and timeline

List these points in the form below, and we reply with a configuration and a written quote within one business day, with a Blackwell option alongside where it suits the work.

FAQ