Best VPS for Self-Hosted AI: Hetzner vs DigitalOcean vs Vultr vs Hostinger
Best VPS for Self-Hosted AI: Hetzner vs DigitalOcean vs Vultr vs Hostinger
You want to self-host your AI. You need a VPS. But which one?
I’ve spent 6 months testing VPS providers for running local LLMs (Ollama, Llama 3.1, Qwen). Here’s what I found.
What Matters for Self-Hosted AI
Running an LLM on a VPS is different from running a web server. The bottlenecks are:
- RAM — The #1 factor. 8GB minimum for 7B-13B models. 16GB for 14B+ models.
- CPU cores — More cores = faster token generation. 2 vCPUs minimum.
- CPU generation — Newer = better. AVX2 support is required for Ollama.
- Disk I/O — Model loading time. NVMe SSD is a must.
- Network — For downloading models (7-40GB each). Good bandwidth matters.
- Price — You’re replacing a $20/month subscription. The VPS should cost significantly less.
The Contenders
1. Hostinger KVM2 — Best Budget Pick
| Spec | Value |
|---|---|
| vCPU | 2 |
| RAM | 8 GB |
| Storage | 100 GB NVMe |
| Bandwidth | 1 TB |
| Price | ~€6/month |
| Affiliate | Get Hostinger |
Pros:
- Best price-to-RAM ratio
- 8GB RAM runs Llama 3.1 8B comfortably
- NVMe SSD for fast model loading
- Good for single-user setups
Cons:
- Only 2 vCPUs (slower inference than 3-4 core options)
- Limited to smaller models (7B-13B)
- No GPU options
Best for: Most people. If you’re running a single-user AI assistant, this is the sweet spot.
2. Hetzner Cloud CPX21 — Best Value
| Spec | Value |
|---|---|
| vCPU | 3 (shared) |
| RAM | 8 GB |
| Storage | 80 GB |
| Bandwidth | 20 TB |
| Price | ~€7/month |
| Affiliate | Get Hetzner |
Pros:
- 3 vCPUs = noticeably faster inference
- Insane bandwidth (20 TB!)
- Excellent API and CLI tools
- EU-based (GDPR friendly)
- Easy snapshots and backups
Cons:
- Shared CPU (can be noisy neighbor issues)
- Only 80 GB disk (enough, but tight if you run many models)
- German company (some may prefer US-hosted)
Best for: Power users who want the best performance per euro. The extra vCPU core makes a real difference in token generation speed.
3. DigitalOcean Basic — Best for Beginners
| Spec | Value |
|---|---|
| vCPU | 2 |
| RAM | 4 GB (or 8 GB for $48/month) |
| Storage | 80 GB |
| Bandwidth | 5 TB |
| Price | $24/month (4GB) / $48/month (8GB) |
| Affiliate | Get $200 credit on DigitalOcean |
Pros:
- Easiest setup experience
- Best documentation and tutorials
- Managed database options
- Great community
Cons:
- Expensive for AI workloads — $48/month for 8GB RAM
- 4GB is barely enough for 7B models
- No shared CPU option
Best for: Complete beginners who want the smoothest experience and don’t mind paying more. But honestly, for AI workloads, the price doesn’t make sense.
4. Vultr — Best for GPU
| Spec | Value |
|---|---|
| vCPU | 1-4 |
| RAM | 4-16 GB |
| Storage | 25-320 GB |
| Price | $6-96/month |
| GPU | From $0.11/hr |
| Affiliate | Get Vultr |
Pros:
- GPU instances available (A100, L40S)
- Flexible pricing (pay by the hour)
- Multiple data center locations
- Bare metal options for serious workloads
Cons:
- GPU instances are expensive ($80-300+/month)
- Complex pricing structure
- Not the cheapest for CPU-only inference
Best for: Users who need GPU acceleration for larger models. If you’re running 70B models or need fast inference, this is your pick.
Performance Benchmarks
I tested Llama 3.1 8B inference on each provider (CPU only):
| Provider | Plan | RAM | Tokens/sec | Model Load | Score |
|---|---|---|---|---|---|
| Hostinger | KVM2 | 8GB | 10-12 t/s | 18s | ⭐⭐⭐⭐ |
| Hetzner | CPX21 | 8GB | 13-16 t/s | 15s | ⭐⭐⭐⭐⭐ |
| DigitalOcean | Basic | 4GB | 6-8 t/s* | 22s | ⭐⭐⭐ |
| Vultr | 2 vCPU | 8GB | 11-14 t/s | 16s | ⭐⭐⭐⭐ |
* DigitalOcean 4GB tested with Phi-3 mini instead of Llama 3.1 8B (too little RAM for 8B)
Winner: Hetzner CPX21 for best performance, Hostinger KVM2 for best value.
My Recommendation
| Use Case | Recommendation | Why |
|---|---|---|
| Single user, budget | Hostinger KVM2 | €6/month, 8GB RAM, simple |
| Single user, best perf | Hetzner CPX21 | €7/month, 3 vCPUs, faster |
| Multiple users | Hetzner CPX31 | €13/month, 4 vCPUs, 16GB |
| Need GPU | Vultr GPU | Pay by the hour when needed |
| Absolute beginner | DigitalOcean | Great docs, but expensive for AI |
For most people reading this: Get the Hostinger KVM2. It’s €6/month, has 8GB RAM, and runs Llama 3.1 8B at a perfectly usable speed. The savings over cloud AI subscriptions pay for the VPS in 3 days.
Want the Full Setup Guide?
I wrote a complete guide that walks you through setting up your entire AI stack on any of these VPS providers. Step-by-step, copy-paste configs included.
Get the Complete Guide — €49 →
Questions?
Drop a comment below or reach out on X. I’m happy to help you pick the right VPS for your use case.
Disclosure: Some links in this post are affiliate links. I earn a small commission at no extra cost to you. I only recommend providers I’ve personally tested.
Want to self-host your own AI?
Get the complete guide with copy-paste configs, step-by-step instructions, and 30-day support.
Get the Guide — €49 →