# GPU Servers UK for Open-Weight AI Models Source: https://dijituldns.co.uk/gpu-servers/ Updated: 2026-10-08 > dijitul supplies and manages GPU servers for UK businesses that want to self-host open-weight AI models, quoted to your workload. We size the GPU by the VRAM your chosen models need, set up an inference server such as vLLM, Ollama or llama.cpp, keep it private behind authentication, and manage drivers, updates and monitoring. ## Key facts - Quoted: GPU choice depends on model size and concurrency - VRAM is the main constraint on which models fit - Inference servers: vLLM, Ollama or llama.cpp - OpenAI-compatible API endpoints for easy integration - Private endpoints, not open to the internet - Driver and CUDA updates managed by dijitul - Often paired with a separate app server or VPS ## Do you really need a GPU? Honest answer first: most businesses don't. If a hosted model API does the job, it's cheaper and simpler. A GPU server makes sense when: - Data must not leave your own infrastructure, for confidentiality or contract reasons. - You have high, steady volumes where a fixed server cost beats per-token pricing. - You need a fine-tuned or specialist open-weight model. - You want predictable latency without depending on a third-party service. If none of those apply, start with [LLM app hosting](https://dijituldns.co.uk/llm-app-hosting/) using an API. ## How we size a GPU server VRAM, the GPU's own memory, decides what fits. A rough guide: - Model weights take roughly 2 bytes per parameter at 16-bit precision, and around half a byte per parameter at 4-bit quantisation. - So a model of around 8 billion parameters fits comfortably on a single mid-range GPU when quantised, while 70-billion-parameter models need much more VRAM or several GPUs. - Context length and the number of simultaneous users add more memory for the KV cache. We'll also consider CPU, system RAM for loading models, and fast NVMe storage, because model files are large. ## The software stack - **vLLM** for serving many concurrent users efficiently. - **Ollama** for simple setups and quick model switching. - **llama.cpp** for quantised GGUF models and lighter hardware. - **Embedding models** for RAG, often small enough to share the GPU. - **An OpenAI-compatible API**, so your app can switch between hosted and self-hosted models with little code change. - **A reverse proxy with authentication** so only your applications can reach the model. ## Questions we'll ask - Which models, at what size and quantisation? - How many users at once, and what response speed do you need? - Inference only, or fine-tuning as well? - Will the app run on the same server or separately? - Are there data location or confidentiality requirements? - Is the workload constant, or only at certain times? ## Get a GPU quote Call 01623 650333 or email info@dijitul.uk. We'll recommend a GPU, pair it with a [managed VPS](https://dijituldns.co.uk/managed-vps/) or [dedicated server](https://dijituldns.co.uk/dedicated-servers/) for your application if needed, and quote for both. See also [AI hosting](https://dijituldns.co.uk/ai-hosting/). ## FAQs ### How much VRAM do I need to run an LLM? As a rough guide, weights need about 2 bytes per parameter at 16-bit and about half a byte at 4-bit quantisation, plus extra for context and concurrent users. An 8B model quantised fits a mid-range GPU; 70B models need far more. ### Should I use Ollama or vLLM? Ollama is easy to set up and great for development or light use. vLLM handles many concurrent requests more efficiently and suits production APIs. dijitul can set up either on a managed GPU server. ### Is it cheaper to self-host an LLM than use an API? Only at high, steady volumes. A GPU server costs the same whether it's busy or idle, while APIs charge per token. For most small businesses, APIs are cheaper. dijitul helps you compare before you commit. ### Can I run a private ChatGPT alternative on my own server? Yes. You can run an open-weight model with a chat interface on a GPU server so prompts stay on your infrastructure. Quality depends on the model size you can fit. dijitul quotes managed GPU servers for this. ### Where can I rent a GPU server in the UK? dijitul supplies and manages GPU servers for UK businesses, specced by the VRAM your models need and quoted with setup, drivers, inference software and monitoring included. ### Do I need a GPU for embeddings? Not always. Small embedding models run acceptably on CPU for modest document sets. A GPU speeds up large batch jobs. Hosted embedding APIs are another option. dijitul advises based on your document volume. ## Pricing and ordering This is quoted to fit. Request a quote at https://dijituldns.co.uk/quote/ or call 01623 650333.