One shared GPU server or a GPU workstation per person: a team of 5 to 30 compared

Ten workstations or one server: the decision for a team

For a data science or ML team of 5 to 30 people, one shared GPU server suits most of the work once the team shares models, data and a model service, and once more than a few people need a GPU at the same time. A workstation per person suits people who need a local display, work offline or need a whole 72 or 96 GB card to themselves all day. Many teams of 10 end up with both: two or three workstations and one server with four cards.

The comparison below uses a team of 10 and two concrete builds we supply. On one side, ten workstations, each with one NVIDIA RTX PRO 5000 Blackwell 72 GB. On the other, one server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards, the inference node on our AI servers page, split with MIG into 16 instances of 24 GB.

NVIDIA RTX PRO 5000 Blackwell datasheet and RTX PRO 6000 Blackwell Server Edition product brief, as quoted in our workstation and server guides; NVIDIA MIG user guide (13 October 2026) for the 1g.24gb profile; power by our estimates below.

Our guide to rack server, tower or workstation for AI compares the chassis types for given cards. This page answers the count question: how many machines a team needs, and where they stand.

What each person gets: memory and compute

A workstation gives its user the whole card. The RTX PRO 5000 72 GB has 14,080 CUDA cores, 72 GB of GDDR7 with ECC and 1,344 GB/s at 300 W, and by our users-per-card rule 61.5 GiB of it holds weights and cache. That runs Qwen3-32B in FP8 with cache for 29 conversations of 8K, or Llama 3.3 70B in NVFP4 with cache for 17, as our guide to the RTX PRO 5000 72 GB workstation works out.

On the server, MIG splits each RTX PRO 6000 Server Edition into up to four isolated instances of 24 GB, two of 48 GB or one of 96 GB, and NVIDIA's MIG user guide lists them as 1g.24gb, 2g.48gb and 4g.96gb. A 1g.24gb slice has a quarter of the card's compute and memory. By our division of NVIDIA's figures, that is about 6,000 CUDA cores and about 400 GB/s per person, against 14,080 cores and 1,344 GB/s on the workstation card. A 24 GB slice holds a model of up to about 14B parameters in FP8, such as Qwen3-14B with 15.2 GiB of weights, with some room for cache; a 32B model in FP8 needs a 48 GB instance.

NVIDIA datasheets as quoted in our RTX PRO 5000 and RTX PRO 6000 guides; slice figures are our division of the card's 24,064 CUDA cores and 1,597 GB/s by four, not NVIDIA figures.

So the workstation is the faster machine per person for the job one person runs alone, such as training a vision model on one GPU. The server wins on the jobs one person cannot run on a desk card: a 96 GB card for a day, two cards for a fine-tuning run, or a model that the whole team queries. Our RTX PRO 6000 MIG profiles show what fits each slice size.

Layouts for ten people on one four-card server

MIG is set per card, so the four cards need not look alike. Three layouts cover most teams of ten:

MIG profiles from NVIDIA's MIG user guide (13 October 2026); gpt-oss-120b with 19 conversations of 32K on one card by our users-per-card rule, as in our four-card server guide.

In the second layout, one card runs a shared model for the whole team, gpt-oss-120b with room for 19 conversations of 32K by our rule, and the other three give twelve people a 24 GB slice for notebooks and experiments. On ten workstations, that model does not fit one 72 GB card with any usable cache; each person would run a smaller model of their own.

A split model runs on whole cards. NVIDIA's MIG guide states that "NCCL is currently not supported with MIG", so a two-card fine-tuning job needs two cards out of MIG mode. We supply this server with the operating system, drivers and CUDA installed on request; send us the team size, the models and how often people need a whole card, and our team replies with a layout and configuration.

Sharing the server: JupyterHub, Kubernetes and Slurm

Two common ways to hand the slices out are JupyterHub on Kubernetes and Slurm.

JupyterHub on Kubernetes. The Zero to JupyterHub guide shows a profile that requests a GPU through extra_resource_limits with nvidia.com/gpu: "1", and it needs NVIDIA's device plugin on the node. With NVIDIA's GPU Operator, the "mixed" MIG strategy exposes each slice size as its own resource, such as nvidia.com/mig-1g.24gb, so a user picks "24 GB notebook" or "whole card" from the profile list. A slice is held while the user's notebook server runs and returns to the pool when it stops. The layout is changed with the nvidia.com/mig.config node label, and the GPU Operator's documentation of 14 October 2026 states that the MIG Manager "requires that no user workloads are running on the GPUs being configured".

Slurm. The Slurm documentation (version 26.05) supports MIG since release 21.08: AutoDetect=nvml in gres.conf, and the slices listed as GPUs in slurm.conf, optionally typed by profile. It also states that Slurm "expects MIG devices to already be partitioned, and does not support dynamic MIG partitioning". A day-and-night layout under Slurm is therefore a planned change by the administrator, not something a job requests.

Our guide to the GPU server for a university or research lab covers Slurm partitions and quotas in more detail, and GPU sharing on Kubernetes with MIG, time-slicing and MPS compares the sharing modes. When several departments share one pool, quotas and chargeback follow the methods in our guide to a shared GPU platform for departments. MIG on bare metal needs no licence; MIG-backed compute vGPU for virtual machines needs NVIDIA AI Enterprise per GPU.

Idle cards, power and noise

A workstation card works only when its owner does. Over lunch, in meetings, on holiday and at night, the card at the desk waits. On the server, a stopped notebook frees its slice for someone else, and a queue can run long jobs at night on whole cards.

Our estimates: workstation with one RTX PRO 5000 72 GB (300 W card + 350 W for processor and the rest) × 1.15, as in our RTX PRO 5000 workstation guide; server with four RTX PRO 6000 Server Edition and two AMD EPYC 9355 at 280 W plus 15 per cent, as in our four-card server guide. These are full-load figures; a measurement on the delivered system confirms or corrects them.

At full load, ten workstations draw about twice the power of the server, by our estimates, and every watt ends up as heat in the office. The server needs a rack position with 32 A or three-phase at 600 W per card, or stays on a 230 V 16 A feed at the 450 W cap, as our guide to GPU server power on 16 A, 32 A and three-phase explains. It does not belong next to desks: Dell's technical guide for its eight-card PowerEdge XE7745, for example, says the server is "not recommended for deployment in acoustically sensitive environments". Our guide to a GPU server in an office covers noise and heat for teams without a server room.

Data, security and administration

Where the data lives changes with the build. On one server, training data, models and notebooks sit in one place, with one set of access rights, one backup and one log of who opened what. On ten workstations, each person keeps copies on a local drive, and each workstation that holds personal or confidential data is one more device in scope for your security rules. Which rules apply to your data is a question for your data-protection officer; the hardware choice decides how many places they have to cover.

Administration scales with the number of machines. NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as fixed driver versions; on the server that is one update, on ten workstations ten, each with its own CUDA, container runtime and framework versions. A team that wants identical environments for everyone gets them more easily from one server with shared container images.

The server is also a single point of failure. If it stops, the whole team stops, while a failed workstation affects one person. Two smaller servers, or a server plus a few workstations, spread that risk.

When a mix wins

A mix of a few workstations and one server fits most teams of ten. The workstations go to the people whose work needs them: a local display for visualisation, 3D or interactive data exploration, work on the move or in a lab without network access, or a whole 72 GB card for daily work that would otherwise block a server card. Everyone else works on the server, and the whole team uses its model service.

For a team of 10, that can be two workstations with one RTX PRO 5000 72 GB each and a four-card server in the "model service plus slices" layout: twelve 24 GB slices for eight people, with room to grow, and a shared gpt-oss-120b. A team that wants 96 GB at the desk takes an RTX PRO 6000 Max-Q workstation, as our guide to the best workstation GPU for AI compares.

From 5 to 30 people

The number of slices scales with the cards, two to eight per server. Not everyone in a team runs a GPU job at the same moment, so the slice count can sit below the head count; how far below depends on how the team works, and a month of usage data from a pilot answers it better than a rule.

Slice counts from four 1g.24gb instances per RTX PRO 6000 Server Edition (NVIDIA MIG user guide); team splits are our planning examples, not measured.

From about 20 people, two four-card servers are often better than one eight-card server: they can stand on two feeds, one can be maintained while the other runs, and a failure takes out half the capacity, as our comparison of one 8-GPU server or two 4-GPU servers works out. A small team of three to five that mainly prototypes can start on DGX Spark systems instead, as our guide to one DGX Spark for a team describes.

What we supply

You can buy both builds from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two to eight RTX PRO 6000 Server Edition cards and workstations with RTX PRO Blackwell workstation cards, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component, and install the operating system, drivers, CUDA and a container runtime on request. The four-card configuration and its variants are on our page to buy a 4× RTX PRO 6000 Server Edition server. Models, RAG and the shared model service for the team are our Private AI/ML service; describe your team, its models and its data, and we propose the split between server and workstations.

FAQ