Pathway 01 · Start wee
Home AI Starter
For: beginners, community projects and anyone who wants to find a useful task before investing in equipment.
The starting kit
- Your current laptop or desktop, with a supported OS and browser.
- A small folder of non-sensitive practice material.
- A notebook for useful prompts, mistakes and what you checked.
Cloud or local?
A browser-based tool does its computing on the provider’s systems. Check its current plan limits and data policy; don’t assume a free tier will stay unchanged. For local experiments, check your runtime’s requirements and try a small model first.
What not to buy yet
A GPU, server or “AI PC” label is not a prerequisite for learning. Keep your money until you can name a specific limitation in your current setup. If the machine is no longer receiving security updates, address that before installing new tools.
Move up when: you have a repeatable task and can explain why your existing machine cannot do it well enough.
Pathway 02 · One capable desk
AI Workstation
For: people exploring local models, coding assistance or creative tools who are willing to maintain their own software.
Two routes to investigate
- Apple Mac mini: investigate Apple-silicon support in your runtime. Compare unified-memory configurations, storage and ports. Memory is shared with the OS; check upgrade limitations before purchase.
- NVIDIA GeForce RTX desktop: investigate if your software supports the exact GPU and drivers. VRAM, power supply, cooling and case clearance matter more than a gaming label.
Neither route is a universal winner. A chat model, coding model and image workflow can need very different resources. Check the exact workload rather than comparing marketing badges.
A sensible order
- Choose one runtime and model. Read the current hardware requirements and licence.
- Estimate memory with headroom for the OS, context and other applications.
- Ask the supplier to confirm the complete system, including power and cooling.
- Keep a separate backup and test restoring a file.
Don’t assume: that model file size equals total memory needed, or that system RAM performs like GPU VRAM. This guide supplies no benchmark or guaranteed model-size limit.
Pathway 03 · The workshop grows
PSA-Style Home Lab
For: confident tinkerers learning how to separate services and recover when something goes wrong.
This is an educational architecture inspired by Pi, Ada and Wee Bot’s Cyber Croft. It is not a description of PSA Studio’s actual machines, network or services.
Give each box a job
- AI compute: reuse your workstation for model inference. You do not need a second GPU machine just to call it a lab.
- Automation: a Raspberry Pi 5 is one option for light services, subject to ARM support. Include suitable power, cooling, enclosure and storage. It is not a substitute for an AI workstation.
- Network: use a wired connection where practical. Keep an inventory and a simple diagram. No port forwarding or public exposure for your first experiments.
- Recovery: keep independent backups, a shutdown procedure and written restore steps. If choosing a UPS, check its wattage, runtime and safe-shutdown support against the actual equipment.
Always-on means ongoing work
Updates, electricity, heat, noise and failed storage are part of the project. Add one service at a time. Keep people in control of any automation that sends messages or changes data. Follow the manufacturer’s electrical guidance; never overload sockets or improvise mains wiring.
Move up when: you can restore your first service and explain how to stop it safely. Wee Bot’s desire for blinking lights is not a capacity plan.
The compatibility checklist
Use these questions with the exact supplier listing and current software documentation. These checks reduce mistakes; they don’t certify a complete build.
What to confirm before ordering| Check | Ask this | Common trap |
|---|
| Software & OS | Does this runtime support my OS, processor architecture and exact accelerator? | Assuming Windows, macOS, Linux and ARM all support the same features. |
|---|
| RAM & VRAM | How much memory does this model need, including context and runtime overhead? | Treating a model’s download size as its complete memory requirement. |
|---|
| Storage | Is there room for models, datasets, updates and backups? Are ports and cables suitable? | Filling the system disk, or buying a fast drive behind a slow connection. |
|---|
| GPU & power | Are drivers, case space, PSU capacity and the required connectors confirmed? | Assuming a graphics card fits because a motherboard has a matching slot. |
|---|
| Local privacy | What leaves the machine: prompts, telemetry, plugins or cloud fallback? | Believing “installed locally” automatically means offline. |
|---|
| Backups | Can I recover projects and settings after a failed disk or bad update? | Calling a second folder on the same disk a backup. |
|---|
Where to check software support
Read the current Ollama documentation or LM Studio documentation if considering those runtimes. These are research links, not tested recommendations. Also read the chosen model’s licence, usage restrictions and hardware notes.
The total-cost buying checklist
The Builder’s £ bands are your planning ceilings, not product prices. We do not claim that any complete system fits a particular band.
- Write the job down: “I want to run this model for this task” beats “I want the most powerful AI box”.
- Compare complete UK quotes: include VAT, delivery, returns, warranty and the exact configuration. For used equipment, check condition, support life and return rights.
- Add the bits: display, keyboard, cables, power supply, cooling, enclosure, storage, backup and any network equipment you actually need.
- Allow for running costs: optional software subscriptions and cloud API usage are separate from hardware. Check limits before connecting any billed service.
- Estimate electricity: average watts ÷ 1,000 × hours used × your electricity tariff per kWh. Measure actual draw if possible; rated PSU capacity is not average consumption.
- Keep headroom: leave a contingency instead of spending the entire ceiling on the main computer.
- Confirm before paying: save the listing and compatibility answers. Don’t rely on an AI-generated parts list as proof.
Independent guidance, not reviews: no hands-on testing, benchmark results, live pricing or affiliate relationship is claimed here. There are no active affiliate links. Future affiliate links must be labelled next to the buying link; PSA Studio may earn a commission from qualifying purchases. Product availability, specifications and terms can change.
Your first week: one useful win
- Choose one non-sensitive task and write down what a good result would look like.
- Install only from official sources. Follow current documentation; don’t paste commands you don’t understand.
- Try a small example, then check the result yourself. AI can be confidently wrong.
- Record what worked, back up your project and restore one file to prove the backup.
- Share a useful result with someone. Only then decide what to add next.
Pi: “Let’s make AI simple.” Ada: “Did you test that first?” Wee Bot: “Can we label the buttons?”