Best VPS for Ollama and local AI in 2026
Ollama's throughput on a VPS depends on three variables: how much of the model fits in RAM, how fast each vCPU is, and whether the CPU supports AVX-512. The provider shortlist below is built around real benchmarks on identical models, not vendor specs.
The two real questions
1. Will the model fit in RAM?
If the model + context + working memory exceeds physical RAM, Ollama spills to swap or refuses to load. Throughput collapses by 10-100x. Sizing rule of thumb (quantized Q4_K_M):
| Model | RAM at idle | RAM under load (4096 ctx) |
|---|---|---|
| 3B | 2.5 GB | 3.5 GB |
| 7-8B | 5 GB | 6.5 GB |
| 14B | 9 GB | 11 GB |
| 24B | 14 GB | 17 GB |
| 32B | 20 GB | 24 GB |
| 70B | 43 GB | 50 GB |
Add ~2 GB headroom for the OS, n8n, anything else co-hosted. Round up to the next plan tier.
2. How fast per token?
Per-vCPU clock speed and the SIMD instruction set matter more than core count. Ollama uses llama.cpp under the hood; llama.cpp is highly optimized for AVX-512 on x86 and Apple Silicon's NEON. AMD EPYC chips (Zen 4 / Zen 5) with AVX-512 outperform older Intel Xeon chips on the same model.
Real benchmark numbers
Test conditions: identical model (llama3.1:8b Q4_K_M), identical 500-token prompt, single concurrent request, CPU-only. Sustained tokens/sec averaged over 5 runs.
| Provider / plan | vCPU / RAM | CPU class | Tokens/sec |
|---|---|---|---|
| Inferno 16 GB | 4 / 16 GB | AMD EPYC (AVX-512) | 11.2 |
| Hetzner CPX41 | 8 / 16 GB | AMD EPYC | 10.8 |
| Hetzner CX42 (Intel) | 8 / 16 GB | Intel Xeon (no AVX-512 exposed) | 6.4 |
| Contabo 16 GB | 6 / 16 GB | Older Xeon | 4.9 |
| Vultr High Frequency | 4 / 16 GB | Intel Xeon (high clock) | 9.1 |
| DigitalOcean Premium | 4 / 16 GB | Intel Xeon | 7.5 |
Two patterns: AMD EPYC with AVX-512 leads, and provider choice matters more than vCPU count at this workload. A 4-vCPU EPYC beats an 8-vCPU Xeon without AVX-512 by 40-70%.
The tested shortlist
1 — Inferno VPS (high-RAM AMD EPYC plans)
Why it wins for Ollama: AMD EPYC with AVX-512 on the larger tiers, the fastest per-token throughput in our CPU-only test. NVMe storage means model loading is quick even on cold starts. Eight locations across EU and Asia.
Best plan: 16 GB for 7-14B models, 32 GB for 24-32B models. Going higher than 32 GB makes more sense only if you also need to co-host n8n, Postgres, multiple agents, etc.
2 — Hetzner CPX (AMD line)
Why it's good: CPX22-44 use AMD EPYC and benchmark almost identically to Inferno's EPYC plans. Falkenstein / Helsinki / Nuremberg locations.
Watch-outs: CX (Intel) line is meaningfully slower for LLM inference. Read the spec page and pick CPX explicitly.
3 — Vultr High Frequency
Why it's good: the fastest single-thread Intel chips. Without AVX-512 it still falls behind EPYC, but it's the best Intel option for Ollama.
Watch-outs: roughly 2× the EU pricing of Inferno or Hetzner. Only worth it if you specifically need Vultr's global location footprint.
4 — Hetzner Dedicated (AX-line)
Why it's good: when you outgrow VPS plans (32B+ models, multi-model setups), Hetzner's dedicated AX line with Ryzen 7950 or EPYC chips and 64-128 GB RAM is a leap in price/performance vs cloud equivalents.
Watch-outs: month-long contracts, manual setup, less elastic than VPS.
5 — GPU options (for completeness)
If you need real interactive speed (50+ tok/s) or want to run 70B-class models, CPU is the wrong path. Look at Vast.ai (cheap consumer GPUs from independent hosts), RunPod (similar), Lambda (cleaner SLA), or Hetzner's dedicated GPU servers. Out of scope for this CPU-focused page; the AI silo's ComfyUI guide touches on GPU sizing.
Avoid for Ollama
- OpenVZ VPS providers. Often can't access certain syscalls; llama.cpp's optimizations don't engage. You'll measure half the throughput of equivalent KVM.
- Burstable cloud instances (AWS t-series, GCP e2). Sustained throughput gets throttled; Ollama's continuous inference burns through CPU credits in minutes.
- Plans where vendor doesn't disclose the CPU model. Without knowing whether you get EPYC or old Xeon, throughput is a gamble.
Concurrent users and worker sizing
Single-model loaded in RAM, multiple parallel requests: llama.cpp can serve concurrent prompts but each new request slows the others proportionally (KV-cache sharing helps a little; CPU contention dominates). Rough rule:
- 1 concurrent user: full throughput as measured above.
- 2-3 concurrent users: ~60-70% of single-user throughput per user.
- 5+ concurrent users on a single CPU box: noticeable lag for everyone. Move to multiple instances (one per user) or pay for a GPU.
Picking a location
For an Ollama API that's only called by other services on the same box (Open WebUI on the same VPS, n8n on the same VPS), location doesn't matter — internal calls are loopback.
For an Ollama API consumed by external clients (your laptop, an n8n on a different host), pick a location close to the caller. Inferno's EU footprint covers most cases; Tokyo if your callers are in Asia.
Recommended: Inferno VPS (16-32 GB EPYC plans) for Ollama
AMD EPYC with AVX-512 gave the best CPU-only throughput in our tests. 16 GB for 8-14B, 32 GB for 24B+, with NVMe model loading and DDoS-included networking.
What to do next
- Picked a VPS? Install Ollama in 10 minutes.
- Open WebUI as a chat front-end.
- Build n8n AI agents with your local Ollama.
- LocalAI if you also need audio / image / embeddings on one server.
Author: Ivo Pärnaste — Tallinn. Last tested: 2026-05-20 with identical Ollama 0.4.6 benchmark suite across six providers.