One powerful platform in Bangor, four mini data-centres on your site
No jargon required — here is what the investment actually buys, and how a business plugs into it.
Six layers, one service
From your laptop to the silicon in Bangor — and on to an edge unit at your premises. Hover each layer for the gist.
Your tools
The coding assistants, AI agents and business workflows you already use — pointed at the platform instead of an overseas API.
Access layer
Isolated project workspaces for teams, plus an OpenAI-compatible API for tools and automation — one key, per-workspace controls.
Models, managed
Leading open-weight models — Qwen, GLM, DeepSeek or suitable successors — hosted and updated by Bangor. Fine-tuning, RAG and agentic systems built on top.
The central platform
An enterprise Blackwell AI server at Bangor: eight RTX PRO 6000 GPUs, 768 GB of ECC GPU memory, a terabyte of system RAM — training, hosting and evaluation under one roof.
The edge layer
Four NVIDIA DGX Spark systems — Grace Blackwell, 128 GB unified memory each — encrypted, asset-tagged and loaned to partner sites on rolling trials before returning to Bangor.
Governance underneath it all
Bangor owns and manages infrastructure, security, access, training and support. Data, IP, security and equipment-loan agreements protect every party.
Specification, for the technically curious
Central AI platform — installed at Bangor University
- System
- Enterprise 6U Blackwell GPU server
- GPUs
- 8 × NVIDIA RTX PRO 6000 Blackwell Server Edition — 96 GB ECC GDDR7 each, 768 GB aggregate
- CPU
- 2 × Intel Xeon 6530P (64 cores)
- Memory
- 1,024 GB DDR5 ECC
- Storage
- 2 × 1.92 TB NVMe boot · 1 × 7.68 TB NVMe scratch
- Networking
- Dual-port 10GbE · NVIDIA ConnectX-7 400 Gb/s
- Power
- Redundant 3,000 W supplies
Edge AI systems — deployed on rotation
- Fleet
- 4 × NVIDIA DGX Spark
- Processor
- NVIDIA Grace Blackwell GB10
- Memory
- 128 GB unified, per system
- Storage
- 4 TB NVMe per system
- Networking
- 10GbE / ConnectX-7
- Custody
- Asset-tagged, encrypted, loan-controlled; returns to Bangor after each trial
- Pooling
- All grant-funded assets return to the central pool by 31 March 2028
If you can use today's AI tools, you can use the platform
The service exposes an OpenAI-compatible API. Coding assistants, agent frameworks and business workflows that speak that standard can point at Wales-hosted models by changing a base URL and key — no rewrite.
Teams that prefer not to touch a terminal work inside a project workspace with training and support from the Bangor team.
- One workspace per project, with its own access controls
- Standard APIs for chat, completions and embeddings-style workflows
- Usage logged per key — tokens, latency, throughput — so pilots can be measured, not guessed
Conceptual demonstration using an example endpoint — not a live service. Real endpoints, keys and model names are issued at project scoping.
The workload types the platform supports
LLM & VLM fine-tuning
Adapt open-weight language and vision-language models to your terminology, formats and documents.
RAG & knowledge work
Grounded question-answering over your own contracts, reports, manuals and records.
Agentic & multi-agent systems
Automated workflows that act across your systems — with human checkpoints and full observability.
ML & deep learning training
Conventional machine learning and deep learning training runs on 768 GB of GPU memory.
Computer vision
Inspection, detection and document understanding — central or on edge units at your site.
Time-series, sensors & simulation
Engineering AI, telemetry analytics and simulation workloads from environment to manufacturing.
Ready to put it to work?
See how participation works — or email the project lead directly.