TrackVPS › Private AI › Self-host LLM cost
What a self-hosted LLM costs per month
Published 21 October 2026Prices checked 4 November 2026By Can
You can run your own AI model on a rented server. The model size decides the RAM, and the RAM decides the price. Here is what five popular sizes cost today, from real plan prices.
We may earn a commission if you buy through our links. It never changes the prices or the order.
Monthly price by model and provider
Cheapest plan at each provider with enough RAM, at the price after any intro or promo (ex VAT). All four contract through an EU company on the link we use (see the EU cloud providers list). IONOS is left out because its ionos.com links sign you with IONOS Inc. (US). Q4 = the default 4-bit download on Ollama.
RAM needed = Ollama download size + context memory estimate + 1.5 GB for the system. Sizes from the ollama.com tag pages, for example qwen3:8b-q4_K_M (5.2 GB). Monthly cost uses the price after any intro, so a cheap intro never counts as the ongoing price. Plans where the provider page does not show that price are left out (Contabo Cloud VPS 4, Contabo Cloud VPS 16, Contabo Cloud VPS 18). Netcup prices need a 12-month term.
What you get for the money
These are normal VPS plans. There is no GPU, so the model runs on the CPU. For one person asking questions, a 3B to 8B model is usable. A 14B model works, but you wait longer for each answer. 32B and 70B fit in RAM on the bigger plans, but expect long waits.
We have not measured speed (tokens per second) yet. So we do not quote any. If you need fast answers for a team, you need a GPU server, and those cost a lot more.
How to pick a size
- Trying it out, simple chat: a 3B or 4B model. Any 8 GB plan has room. Cheapest: RackNerd at €6.68/month.
- Daily use for one person: an 8B model like Qwen3 8B or Llama 3.1 8B. It needs about 7.9 GB, so an 8 GB plan is tight. A 12 GB or 16 GB plan gives you headroom.
- Better answers, more patience: 14B (Qwen3 14B, Phi-4, DeepSeek-R1 14B). Plan for 16 GB or more.
Longer context costs RAM too. At 32k tokens, Qwen3 14B needs several GB more than at 8k. Use the calculator with Qwen3 14B at 32k to see the difference.
Private, but not magic
Self-hosting keeps your prompts off an AI company's servers. Your data sits on a server you rent. The hosting company still controls the hardware. If that matters, check who owns the provider. Our EU cloud providers list shows the company and its country for all 14 providers, and the CLOUD Act explainer says why the owner matters more than the datacentre.
Compare the providers
Contabo's 24-month promo prices are low, but its normal prices are higher (Cloud VPS 6 is €6.00 for 24 months, then €9.00), so we use the normal price. Read the catch first: Contabo review. For a German alternative, see Netcup vs Hetzner and Hetzner alternatives.
Questions
How much does it cost to self-host an LLM?
On 4 November 2026, the cheapest VPS we track that fits an 8B model (Qwen3 8B at Q4, 8k context, about 7.9 GB RAM) was OVHcloud VPS-2 at €7.21/month. A 14B model needs about 12.1 GB and starts at €16.80/month.
Can I run a 70B model on a VPS?
It fits in RAM on a 48 GB plan (€30.00/month on Contabo), but on CPU it will be very slow. For a 70B model you want a GPU server, which costs far more.
Is a self-hosted LLM private?
Your prompts stay on a server you rent instead of going to an AI company. The hosting company still runs the hardware, so pick the provider with care. Our EU cloud providers list shows who owns each one.
Do I need a GPU?
Not for small models used by one person. A normal VPS runs 1B to 8B models on the CPU. Answers come slower than ChatGPT. For many users, big models or fast answers, you need a GPU.