Provider comparison
Find the right server for your AI stack
5 providers tested with real Ollama workloads. Pick your use case — we'll show you exactly what to buy.
Best pick for your use case
Hetzner CX22
22 tok/s on Qwen 2.5 Coder 7B — fast enough for real-time autocomplete. EU-only but latency is fine for async completion.
All providers — ranked for your use case
Hetzner Cloud
Best value CPU VPS in Europe
Cheapest per GB RAM — DigitalOcean costs 4× more for same specs
DigitalOcean
Easiest setup, best documentation
$200 free trial covers 8 months of the 4 GB Droplet
Vultr
32 locations worldwide, GPU options
Only provider with 32 DCs — ideal for latency-sensitive deployments
Contabo
Highest RAM per dollar on the market
6 GB RAM for $6 — only real option for 70B models under $30
RunPod
On-demand GPU cloud, pay per hour
Cheapest GPU rental + best recurring affiliate (3–5%/month)
Full spec comparison
Including Mac Mini M4 — a viable always-on alternative for home labs.
| Provider | Best plan for AI | RAM | Price | 7B speed | 14B speed | 70B capable? | GPU? | Trial | Best for |
|---|---|---|---|---|---|---|---|---|---|
| 🇩🇪 Hetzner CX22 | CX22 (4 GB) | 4 GB | €5.89/mo | 22 tok/s | — | CX52 (€49, 32 GB) | No | None | Budget, code completion |
| 🇩🇪 Hetzner CX32 | CX32 (8 GB) | 8 GB | €12.49/mo | 28 tok/s | 14 tok/s | CX52 (32 GB) | No | None | Private ChatGPT, RAG |
| 🌊 DigitalOcean | Basic 4 GB | 4 GB | $24/mo | 18 tok/s | — | Premium 16 GB ($96) | No | $200 (60d) | Beginners, US users |
| ⚡ Vultr | Cloud 4 GB | 4 GB | $24/mo | 18 tok/s | — | GPU A16 ($90) | Yes | $100 | Global reach, light GPU |
| 💪 Contabo VPS XL | VPS XL (48 GB) | 48 GB | $29.99/mo | 16 tok/s | 10 tok/s | Yes — native | No | None | 70B models on a budget |
| 🔥 RunPod RTX 4090 | RTX 4090 (24 GB VRAM) | 24 GB VRAM | ~$0.44/hr | 55 tok/s | 80 tok/s | A40 ($0.79/hr) | Yes | None | Max speed, batch jobs |
| 🍎 Mac Mini M4 Hardware | Mac Mini M4 16 GB | 16 GB unified | $599 one-time | 40 tok/s | 22 tok/s | Yes (Q4) | GPU-class (Metal) | — | Home lab, always-on, privacy |
Benchmarks from community members running Ollama with Q4_K_M quantisation unless noted. Speeds vary ±15% based on load and model version.
Not just VPS
Considering a Mac Mini instead?
The M4 Mac Mini at $599 breaks even vs a Hetzner CX32 in about 8 months — and handles 70B models the VPS options can't touch at that price.
Affiliate link — we earn a small commission at no cost to you.