Self-hosted AI on a VPS: the complete series
The hub for the self-hosted AI series.
This series explains how to run AI tools on your own server, for beginners. It starts with what self-hosted AI is and what it can realistically do, walks through your first setup, shows what fits on each plan size, and then covers useful projects: searching your own documents, automating work, speech, and AI agents that carry out tasks for you. It finishes with how to keep all of it safe and private.
The honest summary before you start: small AI models on a VPS are slower and less capable than the big hosted services, and the large models need far more memory than most plans have. Self-hosting shines for privacy, for always-on tools and agents, and for learning. Many setups mix both, running the tools on your server and calling a hosted model through an API for the hard thinking.
4 GB to 8 GB VPS plans in stock, the sweet spot for self-hosted AI
- 4 GB KVM VPS Special3 CPU · 4 GB RAM · 60 GB disk · 7 TB bandwidth · New York, Los Angeles, Dallas$59.99/yrabout $5.00/moView at RackNerd
- 6 GB KVM VPS Special6 CPU · 6 GB RAM · 100 GB disk · 12 TB bandwidth · New York, Los Angeles, Dallas$89.99/yrabout $7.50/moView at RackNerd
- 8 GB KVM VPS Special7 CPU · 8 GB RAM · 150 GB disk · 20 TB bandwidth · New York, Los Angeles, Dallas$119.99/yrabout $10.00/moView at RackNerd
- Ryzen-4GB (Linux)2 CPU · 4 GB RAM · 30 GB disk · 3 TB bandwidth · New York, Chicago, Seattle$265.57/yrabout $22.13/moView at RackNerd
In stock when this page loaded. See all 9 matching plans
What is in this guide
- The parts of the series
- Where to start
- A few words you will see
- Which plan for which job
- Local model or API?
- Six rules for every AI project
- What this series does not cover
- When to step up
The parts of the series
| Part | Guide | What it covers |
|---|---|---|
| 1 | Self-Hosting AI Explained for Beginners | What self-hosted AI is, what it is good for, and its honest limits |
| 2 | Self-Hosted AI Tools Explained: A Beginner's Map | The main kinds of tools and how they fit together |
| 3 | Your First Self-Hosted AI in 15 Minutes | A first working setup, step by step |
| 4 | What AI Can You Run on a 1-2 GB VPS? | What fits on small plans, mostly tools that call an API |
| 5 | What AI Can You Run on a 4 GB VPS? | Small local models and comfortable AI apps |
| 6 | What AI Can You Run on a 6-8 GB VPS? | Larger local models and several AI tools together |
| 7 | What AI Can You Run on a Dedicated Server? | When a whole machine makes sense for AI |
| 8 | Self-Hosted AI vs Paying for an API: What It Really Costs | Comparing server costs with pay-per-use model services |
| 9 | Search and Question Your Own Documents With a Private AI (RAG Explained) | Asking questions about your own files |
| 10 | Automate Your Work With AI: n8n Workflows on a Small VPS | Workflows with AI steps that run on their own |
| 11 | Transcribe and Speak: Self-Hosted Speech AI on a VPS | Turning speech into text and text into speech |
| 12 | Keeping Your Self-Hosted AI Safe and Private | Security and privacy for every AI tool in the series |
| 13 | Running a self-hosted AI agent on a VPS | Agents such as OpenClaw and Hermes Agent that carry out tasks for you, and how to run them safely |
Where to start
- You are new to all of this. Read parts 1 and 2, then try part 3.
- You want to know what your plan can run. Go to the part for your plan size: 4, 5, 6 or 7.
- You are deciding whether it is worth it. Read part 8 before buying anything.
- You have a project in mind. Documents: part 9. Automation: part 10. Speech: part 11. An assistant that does things for you: part 13.
- You already run AI tools. Read part 12, and part 13's security rules if you run an agent.
- You have never set up a server. Start with your first hour on a new RackNerd VPS.
A few words you will see
| Term | What it means |
|---|---|
| Model | The AI itself. Larger models are more capable and need more memory. |
| Local model | A model that runs on your own server. |
| API | A way to use a hosted model over the internet, billed by use. |
| Quantised | A model shrunk to use less memory, with some loss of quality. |
| RAG | Retrieval-augmented generation: letting a model answer from your own documents. |
| Agent | An AI that plans and carries out tasks using tools, not just answers questions. |
| Ollama | A popular tool for running local models. |
Which plan for which job
A rough starting point, not a promise. Memory needs vary a lot with the model, the settings and the tool. The part guides give more detail.
| Job | A sensible starting plan | Read |
|---|---|---|
| A chat interface using a hosted model through an API | 1 GB to 2 GB | Part 4 |
| Workflows with AI steps, using an API | 2 GB | Part 10 |
| A light AI agent such as Hermes Agent, using an API | 2 GB to 4 GB | Part 13 |
| A small local model for simple tasks | 4 GB | Part 5 |
| Questions about your own documents | 4 GB with an API, more with a local model | Part 9 |
| Speech to text for short recordings | 4 GB | Part 11 |
| OpenClaw, or any agent that drives a browser | 8 GB | Part 13 |
| Larger local models, or several AI tools together | 6 GB to 8 GB | Part 6 |
| Bigger models and heavy use | A dedicated server | Part 7 |
RackNerd's control panel can install some of these tools in one click, including Ollama, AnythingLLM, LibreChat, n8n, OpenClaw and Hermes Agent. A one-click install erases the server first. RackNerd's one-click apps explained covers what each one does and how to secure it.
2 GB VPS plans in stock, enough for AI tools that use an API
- 2 GB KVM VPS Special2 CPU · 2 GB RAM · 35 GB disk · 5 TB bandwidth · New York, Los Angeles, Dallas$35.99/yrabout $3.00/moView at RackNerd
- Ryzen-2GB (Linux)2 CPU · 2 GB RAM · 20 GB disk · 2 TB bandwidth · New York, Chicago, Seattle$222.37/yrabout $18.53/moView at RackNerd
- KVM-2GB3 CPU · 2 GB RAM · 75 GB disk · 2 TB bandwidth · Los Angeles, Dallas, New York$247.08/yrabout $20.59/moView at RackNerd
In stock when this page loaded. See every plan
Local model or API?
| Local model on your VPS | Hosted model through an API | |
|---|---|---|
| Privacy | Your data stays on your server | Your prompts go to the model provider |
| Quality | Smaller models, simpler tasks | The most capable models |
| Speed on a VPS | Slow on CPU, especially longer answers | Fast |
| Cost | The server, which needs more memory | A small server, plus pay-per-use model charges |
| Best for | Private, simple tasks and learning | Hard tasks, agents and anything that needs quality |
Part 8 works through the costs. Whichever you choose, set a spending limit on any API account before you start.
Six rules for every AI project
- Keep control panels private. Bind AI tools to
127.0.0.1and reach them through an SSH tunnel or VPN. See setting up a WireGuard VPN. - Never leave a tool without a login, and turn off open sign-up.
- Protect API keys. Keep them in files readable only by the app, with a spending cap on each.
- Give agents their own server and only the tools they need. See part 13.
- Read and test anything AI writes before you rely on it.
- Check licences. Models and tools have their own terms, especially for business use.
Part 12 and a beginner's guide to securing your VPS cover the rest.
What this series does not cover
- Training your own models. That needs graphics cards and far more resources than a VPS.
- Graphics card servers. RackNerd's VPS plans run on CPU. Local models are therefore slower than on a machine with a suitable graphics card.
- Exact model recommendations. Models change every few months. The guides describe families and sizes, and you should check each project's current documentation.
- Guarantees of speed or quality. Test with your own tasks.
Cautions for every plan: swap is a safety net and not extra RAM, unmetered does not mean unlimited, fair-use limits apply to CPU and bandwidth, and RackNerd's acceptable use policy covers what your AI tools do.
When to step up
If a local model is too slow or keeps running out of memory, either move up a plan size or switch that task to an API. For agents, 8 GB gives room for browser tasks. For larger local models and heavy use, a dedicated server is the next step. How to read VPS specs explains which specs matter for AI.
8 GB VPS plans in stock, for agents and larger local models
- 8 GB KVM VPS Special7 CPU · 8 GB RAM · 150 GB disk · 20 TB bandwidth · New York, Los Angeles, Dallas$119.99/yrabout $10.00/moView at RackNerd
- KVM-8GB6 CPU · 8 GB RAM · 220 GB disk · 5 TB bandwidth · Los Angeles, Dallas, New York$385.59/yrabout $32.13/moView at RackNerd
- Ryzen-8GB (Linux)3 CPU · 8 GB RAM · 75 GB disk · 5 TB bandwidth · New York, Chicago, Seattle$395.17/yrabout $32.93/moView at RackNerd
In stock when this page loaded. See every plan
Dedicated servers in stock with the most memory, for bigger models
- [UT] Dual Intel Xeon E5-2695 V3 - 512GB RAM - 10x 1.6TB SSD56 threads · 512 GB RAM · 16 TB disk · 50 TB bandwidth$579.00/moView at RackNerd
- 10Gbps - Dual Xeon E5-2640 V3 Dedicated32 threads · 256 GB RAM · 5 TB disk · 100 TB bandwidth$299.00/moView at RackNerd
- Dual Intel Xeon E5-2650 v2 | 256GB RAM | 4x 2TB SSD32 threads · 256 GB RAM · 8 TB disk · 100 TB bandwidth$349.00/moView at RackNerd
In stock when this page loaded. See all 71 matching plans
To compare what is available now, start with the cheapest RackNerd VPS guide and check the sales page.