Best VPS for AI & LLM Apps in 2026
Updated Aug 2026 · 8 plans benchmarked
Serving an open LLM — Ollama, LocalAI, or Open WebUI — is a RAM story first and a CPU story second. The model weights have to live in memory to run at all, so the winning box is the one that gives you the most GB of RAM and fast NVMe per dollar. Every pick here clears the 8 GB floor where small models become usable.
- RAM is the whole game — model weights must fit in memory, and 8 GB is where a 7B-class model becomes usable; 16 GB gives you headroom for context and concurrent requests.
- Fast NVMe storage so model weights (4–8 GB each) load quickly and hot reloads don't stall.
- A CPU with decent single-thread speed — inference is compute-bound on a GPU-less box, so steady clock matters more than core count for small models.
- Optional GPU support if you ever want to accelerate, and a provider with generous bandwidth for pulling weights and serving many requests.
The picks
The most RAM per dollar in the market — 8 GB NVMe for $4.95 is the cheapest way to get a 7B model in memory, though you trade some operational polish for it.
The value pick for CPU inference — the CX32 (4 vCPU / 8 GB) runs 7B-class models at ~5–10 tokens/s, and 20 TB of traffic means serving a busy API costs nothing extra.
The VPS-2 (4 vCPU / 8 GB) covers the same small-model workload at a low sticker price, with daily backups and anti-DDoS included by default.
When RAM or GPU needs outgrow the entry tier, Kamatera's custom sizing lets you dial up exactly what a bigger model wants — plus hourly billing to test a box for an afternoon.
The gentlest on-ramp for a first AI box: one-click images and the clearest docs mean less time on setup. You pay for that polish over Hetzner's raw price.
The picks, side by side
Prices verified Aug 2026 · How we benchmark →
Common questions
How much RAM do I need to run an LLM on a VPS?
Small models (3B) run in 4 GB; 7B-class models want 8 GB minimum and are much more usable at 16 GB. The model weights must fit in memory, and what's left over is what the context and your other services get.
Can I run an LLM on a CPU-only VPS?
Yes — Ollama, LocalAI, and Open WebUI all run without a GPU, just more slowly. Expect roughly 5–10 tokens per second for a 7B model on a 4-core box, which is fine for chat and automation, not for real-time generation at scale.
Is a GPU VPS worth it for self-hosted AI?
Only if you're serving many users or need low latency. A GPU gives 5–10x the tokens/second but multiplies the price. For a personal or small-team setup, a well-sized CPU box is usually the better value.