Private AI server
A private AI server: your own ChatGPT, on your own machine.
Ollama runs open models such as Llama, Qwen, Gemma and DeepSeek-R1 on the server, and Open WebUI gives you a ChatGPT-style chat in your browser. With a local model, your prompts and chats stay on your server: no AI company sees them. One click installs both on a Lite 8 for $12 a month, which runs 8B models without a GPU; a Lite 16 at $20 runs 14B. Billed by the hour, paid in crypto with no ID.
Welcome bonus: we double your first deposit. Deposit $100, spend $200. Your server is set up once we have reviewed and approved your order; delete it any time.
Plans and prices
Private AI server plans, sized by the model you want.
All prices in USD, without VAT. Customers in the EU pay their country’s VAT on top. No setup fees, no contracts.
The one-click app needs 8 GB. Lite 8: models up to 8B. Lite 16: up to 14B. Lite 32: up to 30B, by our sizing rule below. Billed per started hour, capped monthly.
| Plan | vCPU | RAM | Storage | Traffic (bandwidth) | Hourly | Monthly | Keep a year | Action |
|---|---|---|---|---|---|---|---|---|
| Lite 8 | 8 | 8 GB | 160 GB NVMe | 3 TB | $0.018 | $12 | $120save $24 | Get it Lite 8 |
| Lite 16 | 16 | 16 GB | 320 GB NVMe | 4 TB | $0.03 | $20 | $220save $20 | Get it Lite 16 |
| Lite 32 | 16 | 32 GB | 640 GB NVMe | 6 TB | $0.06 | $40 | $440save $40 | Get it Lite 32 |
| Lite 64 | 24 | 64 GB | 1,280 GB NVMe | 8 TB | $0.12 | $80 | $880save $80 | Get it Lite 64 |
| Lite 96 | 32 | 96 GB | 1,920 GB NVMe | 10 TB | $0.18 | $120 | $1,320save $120 | Get it Lite 96 |
| Lite 128 | 40 | 128 GB | 2,560 GB NVMe | 12 TB | $0.24 | $160 | $1,760save $160 | Get it Lite 128 |
| Lite 192 | 48 | 192 GB | 3,840 GB NVMe | 14 TB | $0.36 | $240 | $2,640save $240 | Get it Lite 192 |
| Lite 256 | 64 | 256 GB | 5,120 GB NVMe | 16 TB | $0.48 | $320 | $3,520save $320 | Get it Lite 256 |
Bandwidth: every server has a 10 Gbps port. Traffic is the monthly outbound allowance, pooled across all your servers. Incoming traffic is free, and past your pool it is $2 per TB.
Free model fit tool
Which AI models fit on your server?
A model has to fit in memory: system RAM on a Lite, GPU memory on a Forge. Pick a plan and see every popular model it runs, from Ollama’s own download sizes plus room for the system.
Biggest model that runs fully
Qwen3 14B
- Qwen3 0.6B (qwen3:0.6b, 523 MB): Fits in RAM
- DeepSeek-R1 1.5B (deepseek-r1:1.5b, 1.1 GB): Fits in RAM
- Llama 3.2 1B (llama3.2:1b, 1.3 GB): Fits in RAM
- Qwen3 1.7B (qwen3:1.7b, 1.4 GB): Fits in RAM
- Llama 3.2 3B (llama3.2:3b, 2 GB): Fits in RAM
- Qwen3 4B (qwen3:4b, 2.5 GB): Fits in RAM
- Gemma 3 4B (gemma3:4b, 3.4 GB): Fits in RAM
- Mistral 7B (mistral:7b, 4.4 GB): Fits in RAM
- DeepSeek-R1 7B (deepseek-r1:7b, 4.7 GB): Fits in RAM
- Llama 3.1 8B (llama3.1:8b, 4.9 GB): Fits in RAM
- Qwen3 8B (qwen3:8b, 5.2 GB): Fits in RAM
- DeepSeek-R1 8B (deepseek-r1:8b, 5.2 GB): Fits in RAM
- Gemma 3 12B (gemma3:12b, 8.2 GB): Fits in RAM
- DeepSeek-R1 14B (deepseek-r1:14b, 9 GB): Fits in RAM
- Phi-4 14B (phi4:14b, 9.1 GB): Fits in RAM
- Qwen3 14B (qwen3:14b, 9.3 GB): Fits in RAM
- gpt-oss 20B (gpt-oss:20b, 14 GB): Too big
- Gemma 3 27B (gemma3:27b, 17 GB): Too big
- Qwen3 30B (qwen3:30b, 19 GB): Too big
Speed: we don’t publish figures we haven’t measured. On CPU, answers come more slowly than from ChatGPT, and smaller models answer faster. Check any model on your server with ollama run qwen3:4b --verbose, which prints the speed in tokens a second.
What private means
What stays on your server, and what doesn’t.
“Private” gets stretched on AI hosting pages. Here’s exactly where your data goes.
| What | Where it goes |
|---|---|
| Prompts and chats with a local model | Stay on your server. Open WebUI keeps them in its data folder on your disk. |
| Model downloads | Fetched from Ollama’s library, which sees which model your server asks for. |
| A hosted model you connect (OpenAI, Anthropic and others) | Those prompts go to that provider, under its terms. |
| Web search, if you turn it on | Search queries go to the search service you pick. |
| The server | Runs on our hardware, like any host. See what that means in our hosting privacy guide. |
Read more in what your host still sees.
Local models, local chats, your server.
Setting up
Set up your private AI server in three steps.
- 1. Deploy with the appPick Ollama + Open WebUI when you deploy a Lite 8 or larger. Our install keeps Ollama’s API on the server itself, where only Open WebUI talks to it: Ollama doesn’t check local requests, so it should never face the internet.
- 2. Create your account firstIn Open WebUI the first account becomes the administrator. Open it as soon as your server is ready and sign up. After that, new sign-ups can’t chat until you approve them.
- 3. Pull a modelFrom Open WebUI’s admin settings, or over SSH with
ollama pull qwen3:8b. Then point a domain at the server and put HTTPS in front.
Running it for more than 50 people? Open WebUI’s licence keeps its branding above that. For an API without the chat, see Ollama on a VPS.
One click, one account, one model.
Faster answers
When a GPU is worth it.
A GPU answers faster, but only for models that fit in its memory. A Forge L40S has 8 GB of NVIDIA L40S memory: enough for 8B models fully on the GPU, with bigger ones split between GPU and its 32 GB of RAM. A Forge L4’s 2 GB holds only the smallest models. The tool above shows both.
For most private chat on 8B and 14B models, a Lite does the job. See GPU servers when speed matters more than price.
CPU for value, GPU for speed.
What people run
What people use a private AI server for.
A private ChatGPT
Everyday questions, with nothing sent to an AI company.
Sensitive drafts
Contracts, code, client notes: kept on your disk.
A team chat
Accounts for colleagues, approved by you.
An API for your apps
Ollama’s OpenAI-compatible API, for your own scripts.
Model testing
Try Qwen, Llama and Gemma side by side, by the hour.
Agents with a local model
Self-hosted agents that never call out.
Your use case is on the list? Lite 16 is $20 a month, or $0.03 an hour.
The rules
What you pay for, and what you don’t.
Every rule is on this page before you pay.
The server: billed per started hour at 0.15% of the monthly price, never more than the monthly price. How hourly billing works.
The models: free to download from Ollama’s library. Each model has its own licence; check it before commercial use.
Messages: no per-message or per-token fees with local models. The server is the only limit.
Try it without the risk
The worst case is an hour and a few cents.
You shouldn’t have to take any host’s word for it. Here is what protects your money. Refund policy
- Delete any timeBilling stops at the hour you delete the server. No notice period, no cancellation form.
- A monthly capHourly charges stop at the plan’s monthly price. The rest of that month is free.
- The calculator is a promiseIf your bill is ever higher than our calculator for the same plan and hours, we credit the difference to your balance.
- Your balance waits for youCredit you don’t spend stays on your account for the next server, in any of our countries.
- Paused, not deletedRun out of credit and hourly servers pause, kept for 7 days. Top up and they come back.
- People who answerOur target is a first support reply within 20 minutes, 24/7, and 99.99% uptime with credit back if we miss. SLA
Deposit $100 and you have $200 to spend.
1. Create an account with an email address, a password and your country (for VAT). No name, no passport, no phone number.
2. Add funds from $20 in BNB, BTC, ETH, SOL, TRX, USDC, USDT, XRP or Monero. Each invoice shows one exact amount on one network, and the price holds for 30 minutes. A first deposit of $100 or more is matched 100% in bonus credit, up to $250; a smaller one gets +25%.
3. Your balance updates once the network confirms. You can close the page; we watch the chain for you. Coins, networks and confirmation times.
4. Deploy. Pick a country and an operating system. Your server is set up once we have reviewed and approved your order, and billing starts only when it is ready.
Locations
Available in 15 countries.
Bulgaria VPSSofia
Germany VPSFrankfurt
Iceland VPSReykjavik
Japan VPSTokyo
Lithuania VPSŠiauliai
Luxembourg VPSLuxembourg
Moldova VPSChișinău
Netherlands VPSAmsterdam
Pakistan VPSIslamabad
Romania VPSBucharest
Singapore VPSSingapore
Sweden VPSStockholm
Turkey VPSIstanbul
Ukraine VPS*Kyiv
USA VPSSan Francisco
* Ukraine: Service may be disrupted by the war. Keep backups in another country.
Private AI server FAQ
Private AI server questions, answered straight.
Do I need a GPU?
No. Ollama runs on CPU with system RAM. A Lite 8 runs 8B models; a GPU only makes answers faster.
How much RAM do I need?
The model’s size plus room for the system. Our tool says: 8B on a Lite 8, 14B on a Lite 16, 30B on a Lite 32. Check a plan.
How fast is it?
Slower than ChatGPT on CPU, faster on smaller models. We haven’t measured our plans yet, so we don’t quote a number: ollama run --verbose shows yours.
Is it really private?
With a local model, your prompts and chats stay on your server. Connecting a hosted model or web search sends those requests out. The full list.
Who can sign up to my Open WebUI?
The first account becomes the admin, so create yours first. Later sign-ups wait for your approval.
Can I use ChatGPT or Claude models in it too?
Yes: Open WebUI connects to OpenAI-compatible APIs as well. Those chats go to that provider.
How do I pay?
In crypto: BTC, ETH, USDT, USDC, SOL, BNB, XRP or XMR, with an email and a password and no ID.
The rest of the keep
Same account, same balance, six more ways to build.
Your AI, on your server.
Ollama and Open WebUI on a Lite 8, $12 a month.
Top up in crypto (deposit $100, spend $200 on your first one), pick a country. Your server is set up once we have reviewed and approved your order. No card, no ID, no contract.
