Want to run AI locally on your own PC – no subscriptions, no cloud costs, no data leaving your machine? This guide covers exactly what hardware you need to do it properly.
Why Run AI Locally at All?
The cloud AI services are convenient – you open a browser, type a prompt, and get an answer. But there are three real problems with relying on them entirely: cost, privacy, and control.
ChatGPT Plus costs £20 a month. Claude Pro is similar. Midjourney starts at £9 a month. Stack a few of these together and you’re spending £500-£600 a year on AI subscriptions – for tools that can slow down at peak times, change their pricing, or restrict what you can do without warning.
Privacy is the bigger issue for a lot of people. Everything you type into a cloud AI service is processed on someone else’s server. For personal use that might be fine. For business work, client data, or anything sensitive – it’s a real concern.
Run AI on your own hardware and nothing leaves your machine. No subscriptions, no usage caps, no data shared with third parties. Your hardware, your models, your rules.
The ability to run AI locally used to require serious technical knowledge. That’s changed dramatically in the last two years. Tools like Ollama, LM Studio, and ComfyUI have made local AI genuinely accessible – install, download a model, run. The hardware is the main consideration now.
The Key to Running AI Locally: VRAM
If you take one thing from this article, let it be this: VRAM is everything for local AI. Not your CPU speed, not your RAM, not your storage. The amount of video memory on your graphics card determines what models you can run, at what speed, and at what quality.
AI models are loaded into GPU memory to run. If the model is too large to fit in your VRAM, one of two things happens: it either falls back to your system RAM (which is much, much slower – we’re talking 10x slower in some cases) or it simply refuses to load at all.
“More VRAM means bigger models, faster inference, and more ambitious workflows. It’s the one spec you can’t easily compensate for elsewhere.”
How Much VRAM Do You Actually Need?
The answer depends entirely on what you want to run. Here’s a practical breakdown:
| Use Case | Min VRAM | Recommended | Notes |
|---|---|---|---|
| Small LLMs (7B params) – Llama 3.1 7B, Mistral 7B | 6-8GB | 8-12GB | Good for chat, summaries, basic coding assist |
| Mid LLMs (13B params) – Llama 3.1 13B, Mistral NeMo | 10-12GB | 16GB | Noticeably smarter responses, still fast |
| Large LLMs (34B params) – Qwen 2.5 32B, Mistral Large | 20-24GB | 24GB+ | Approaching GPT-4 quality for many tasks |
| Image Gen – Stable Diffusion 1.5 / SDXL | 6-8GB | 12-16GB | More VRAM = higher resolution, faster generation |
| Image Gen – FLUX Full / High Resolution | 16GB | 24GB+ | FLUX is the current state-of-the-art model |
| AI Video Generation – Wan2.1, CogVideo | 16GB | 24-32GB | Very demanding – more VRAM = longer clips, better quality |
| Very Large LLMs (70B+) – Llama 3.1 70B | 40GB+ | 48GB+ or dual GPU | Flagship hardware only |
Common Mistake: Many people buy an 8GB GPU thinking it’ll handle serious AI work – then discover it’s too limited for anything beyond basic 7B models. If budget allows, always go higher on VRAM. You can’t upgrade it later without replacing the whole card.
Which GPU Do You Need to Run AI Locally?
For local AI, NVIDIA is the clear choice. AMD cards technically work, but NVIDIA’s CUDA platform has far better software support across all the major local AI tools. If you’re building a machine specifically for AI, NVIDIA isn’t just recommended – it’s essentially required for the best experience.
Current Best Options by Budget
- RTX 5070 (12GB) – ~£550-650: Entry AI – good for 7-13B LLMs and SDXL image generation. Solid starting point.
- RTX 5080 (16GB) – ~£900-1,100: Good mid-range – faster than the 4090 but less VRAM. Great for image gen and mid LLMs.
- RTX 4090 (24GB) – ~£1,400-1,600: Strong mid-range – excellent all-rounder. Handles FLUX, 34B LLMs, video generation.
- RTX 5090 (32GB) – ~£1,900-2,200: Recommended flagship – best single-GPU AI card available. 32GB handles almost everything.
The RTX 5090 is the current gold standard for a single-GPU AI workstation. 32GB of VRAM handles 70B LLMs at quantised quality, full FLUX image generation at high resolution, and AI video generation without compromise. It’s expensive – but it’s the one card that won’t leave you wishing you’d bought more VRAM six months later.
Want a Chester-built AI workstation specced around your needs? See our AI Workstations page →
System RAM for Running AI Locally
System RAM matters more for AI than it does for gaming. Here’s why: when a model is too large for your VRAM, it offloads layers to system RAM. The more RAM you have, and the faster it is, the less painful that offloading is.
Even if you have a 32GB VRAM GPU, running large multi-modal workflows – image generation alongside an LLM, for example – can eat system RAM quickly. 32GB is the minimum for serious AI work. 64GB is comfortable. 128GB is the right spec for flagship builds.
RAM Speed Matters Too: Faster RAM (DDR5-6000 vs DDR5-4800) measurably improves model offloading speeds when VRAM fills up. Worth speccing properly from the start.
Storage – Fast NVMe Is Non-Negotiable
AI models are large files. A 7B model might be 4-8GB. A 70B model can be 40GB or more. You’ll accumulate several over time – different models for different tasks. You need the storage to hold them, and you need it to be fast.
Why fast storage matters: every time you load a model, it reads from your drive into memory. A slow drive means 30-60 second load times on large models. An NVMe Gen4 or Gen5 drive loads the same model in 5-10 seconds. That adds up significantly if you’re switching between models regularly.
- Minimum: 1TB NVMe Gen4 – enough to get started, will fill up quickly
- Recommended: 2TB NVMe Gen4 – comfortable for most users
- Flagship: 4TB NVMe Gen5 – fast loads, plenty of space, future-proof
CPU – Important, But Not the Priority
The CPU matters less for AI inference than you might expect. The GPU does the heavy lifting. That said, a weak CPU creates bottlenecks in data preprocessing and multi-tasking – if you’re running an LLM in the background while doing other work, you’ll notice a slow CPU.
For a dedicated AI workstation, a Ryzen 9 9900X or 9950X (or Intel Core i9 equivalent) is the right call. High core counts help with preprocessing tasks, and the memory bandwidth of modern high-end CPUs complements the GPU well.
What Does a Complete AI Workstation Look Like?
Here’s how a well-specced local AI machine comes together at three budget points:
Entry AI Build – From £1,500
GPU: RTX 5070 – 12GB VRAM
CPU: Ryzen 7 9700X
RAM: 32GB DDR5
Storage: 2TB NVMe Gen4
Best for: 7-13B LLMs, SDXL image generation, getting started with local AI
Recommended AI Build – From £2,500
GPU: RTX 5080 – 16GB VRAM
CPU: Ryzen 9 9900X
RAM: 64GB DDR5
Storage: 4TB NVMe Gen5
Best for: Mid LLMs, full FLUX image gen, AI video, serious creative workflows
Flagship AI Build – From £4,000
GPU: RTX 5090 – 32GB VRAM
CPU: Ryzen 9 9950X
RAM: 128GB DDR5
Storage: 4TB NVMe Gen5
Best for: 70B LLMs, maximum resolution image gen, AI video, no compromises
What Software Do You Actually Use?
The hardware question answered, here’s a quick map of the software ecosystem – all free, all actively maintained, all running locally on your machine:
- Ollama – the simplest way to run LLMs locally. Install, pull a model, chat via terminal or a web UI. Works with Llama, Mistral, Gemma, Qwen, and dozens more.
- LM Studio – a polished desktop app for running and chatting with local LLMs. Great for non-technical users who want a clean interface.
- ComfyUI – the most powerful tool for Stable Diffusion and FLUX image generation. Node-based workflow builder with enormous community support.
- Automatic1111 – simpler than ComfyUI, still very capable for image generation. Better starting point for beginners.
- Jan – a clean, simple desktop LLM client. Good alternative to LM Studio.
Hardware Only: At Major Custom Computing, we supply and build the hardware. Software installation is down to you – but all of the tools above are well documented with active communities, and most install and run in under 30 minutes on the right hardware.
Run AI Locally – The Short Version
Running AI locally is genuinely viable in 2025 – the software has caught up with the hardware, and the hardware has never been better. The key points:
- VRAM is the bottleneck – buy as much as your budget allows. 12GB minimum, 24-32GB for serious work.
- NVIDIA only – AMD works, but NVIDIA’s CUDA support is significantly better across all local AI tools.
- RAM matters more than for gaming – 32GB minimum, 64GB comfortable, 128GB flagship.
- Fast NVMe storage – model load times depend on it. Gen4 minimum, Gen5 preferred.
- CPU is secondary – high-end Ryzen 9 for premium builds, but not the priority.
If you’re in Chester or the surrounding area and want a machine built specifically for local AI work – properly specced around your actual workload, hand-assembled and tested – that’s exactly what Major Custom Computing does.
View our AI Workstations Chester page → or get in touch with Mat directly.
Need a Build or Repair?
Talk to Mat directly — honest advice, no call centres, no faff.