Our current turnaround time is 7-10 working days
What are you building for?
Tailored recommendationsPC builds across the site will be filtered to match your chosen use-case.

AI Workstations

Hand-built AI workstations for LLM, Stable Diffusion & ML.

From RTX 5090 inference rigs to dual RTX 6000 Pro Blackwell studio workstations with 192GB combined VRAM. Specified, built and supported by people who actually run these workloads.

  • 5-year warranty
  • Hand-built in Stevenage
  • Up to dual RTX 6000 Pro Blackwell
  • Up to 192GB VRAM
  • 24-hour burn-in testing
Configure an AI WorkstationTalk to a Specialist

Choose your AI workload

Three core configurations. Each is a starting point — every CREATE workstation is configurable down to the cooler, fan and RAM kit.

From £3,499

LLM Inference & Fine-Tuning

Run Llama 3, Mistral, Mixtral and DeepSeek locally. From single-RTX-5090 inference rigs to dual RTX 6000 Pro Blackwell training stations.

VRAM from 24GB · View configurations →

From £3,199

Stable Diffusion & Generative Image

SDXL, FLUX, Stable Video and ComfyUI workstations tuned for sustained GPU throughput, with VRAM headroom for high-res and batch workflows.

VRAM from 24GB · View configurations →

From £4,499

Machine Learning & Data Science

PyTorch and TensorFlow training rigs with ECC RAM options, NVLink-capable GPU pairs and high-capacity NVMe storage for dataset throughput.

VRAM from 32GB · View configurations →

AI GPU comparison · which card fits which model

A practical guide to matching NVIDIA Blackwell GPUs to local AI workloads. Model-size figures assume 4-bit quantisation with a sensible context budget; FP16 / training requires roughly 2× the VRAM.

AI GPU comparison · which card fits which model
 VRAM (GB)AI TOPS (FP4)Max practical LLM (Q4)Tier
RTX 5060 Ti
Blackwell · 16GB GDDR7
167597B (Q4)Entry
RTX 5070 Ti
Blackwell · 16GB GDDR7
161,40613B (Q4)Mid
RTX 5080
Blackwell · 16GB GDDR7
161,80113B (Q4)Mid-High
RTX 5090
Blackwell · 32GB GDDR7
323,35270B (Q4)High
RTX 6000 Pro Blackwell
Pro · 96GB GDDR7 (ECC)
964,000180B (Q4) / fine-tune 13BPro
2× RTX 6000 Pro Blackwell
Pro · 2× 96GB (192GB total)
192 (combined)8,000405B (Q4) / fine-tune 70BStudio

AI TOPS figures published by NVIDIA at the Blackwell launch. Real-world throughput depends on framework, batch size, quantisation method and cooling — we test every build under sustained load before dispatch.

AI workstation FAQs

Common questions from teams and individuals specifying a CREATE AI workstation.

How much VRAM do I need for local LLM inference?

As a rule of thumb a 4-bit quantised model needs roughly half its parameter count in GB of VRAM, plus headroom for context. A 7B model fits comfortably in 8GB; a 13B in 16GB; a 70B model needs ~40GB and is the sweet spot for a single RTX 5090 (32GB) with offload, or a single RTX 6000 Pro Blackwell (96GB) for unconstrained context. For 180B+ models we'd recommend dual RTX 6000 Pro.

Is the RTX 5090 better than the RTX 4090 for AI workloads?

Yes. The RTX 5090 has 32GB GDDR7 (vs 24GB GDDR6X on the 4090), substantially higher memory bandwidth, and Blackwell tensor cores with FP4 support — practical AI throughput is 1.7–2.2× a 4090 on most modern frameworks, and the extra 8GB VRAM lets you run 70B-class models that simply do not fit on a 4090.

Should I get a consumer RTX 5090 or a professional RTX 6000 Pro?

Choose the RTX 5090 if you need the best price-per-token for inference and SDXL workflows up to 70B. Choose the RTX 6000 Pro Blackwell when you need 96GB VRAM in a single card (fine-tuning, very long context, large image batches), ECC memory for unattended training, or to put two cards in one chassis. Most teams running production AI in-house go RTX 6000 Pro; most prosumers and developers go RTX 5090.

Do you support NVLink for multi-GPU training?

Blackwell consumer cards (RTX 5080/5090) do not support NVLink. The RTX 6000 Pro Blackwell supports peer-to-peer and large-pool memory aggregation in supported chassis — we configure these in workstation cases with appropriate PSU headroom (typically 1600W Platinum) and air or hybrid cooling.

What CPU and RAM should I pair with an AI workstation?

For single-GPU inference an Intel Core Ultra 9 285K or AMD Ryzen 9 9950X with 64GB DDR5 is plenty. For training and multi-GPU we move to AMD Threadripper PRO (24+ cores, 8-channel ECC RAM) and 128–256GB. Storage matters too — we spec at least 2TB Gen5 NVMe for hot datasets and a second 4TB+ NVMe for model weights.

What about local AI assistants and RAG workflows?

An RTX 5090 with 32GB VRAM happily runs Ollama, LM Studio, llama.cpp, vLLM and AnythingLLM with 70B-class models at usable speeds (10–20 tok/s). Pair it with 64–128GB system RAM for vector databases (Qdrant, Chroma) and you have a fully local RAG stack with no API costs and no data leaving your premises — popular with legal, medical and engineering clients.

Do you offer support for AI workloads after delivery?

Yes — every CREATE AI workstation includes our 5-year hardware warranty (parts and labour). On AI-tier builds we also include a one-on-one onboarding session covering driver setup, CUDA/cuDNN, your chosen framework (PyTorch / TensorFlow / Ollama / ComfyUI) and basic performance tuning. Bespoke ongoing support is available on request.

How long does an AI workstation take to build?

Current turnaround is 7–10 working days from order. Studio-tier builds (dual RTX 6000 Pro, Threadripper PRO) typically run 10–14 working days because of additional 24-hour burn-in and validation testing.

Why CREATE PCs for AI

Specified by builders who run the workload

We pick the cooler, PSU, NVMe and RAM kit for sustained AI throughput — not just to pass a quick benchmark.

Up to dual RTX 6000 Pro Blackwell

We routinely build single- and dual-GPU professional workstations with 96GB and 192GB VRAM pools.

5-year warranty (parts + labour)

Longer than PCSpecialist, longer than Chillblast’s combined collect-and-return + labour coverage.

24-hour burn-in before dispatch

Every system is stress-tested under load for at least 24 hours — Prime95, OCCT and sustained GPU workloads.

AI-focused onboarding

Optional one-on-one setup session covering CUDA, cuDNN, PyTorch / Ollama / ComfyUI and your specific framework.

UK hand-built, fully cable-managed

Built in Stevenage, photographed at every stage, packaged so they arrive perfectly.