Run Ollama and Open WebUI on a rented GPU: your own ChatGPT for an evening
Tutorial · Published 30 October 2026 · prices on this page are live
Ollama runs open language models on a GPU with one command; Open WebUI puts a ChatGPT-style chat window in front of it. On PowerPod the two come as one ready-made software choice, so the whole thing takes about ten minutes from an empty account to a conversation with Llama, Qwen or Gemma running on a GPU you rent by the second.
Which GPU you need
What decides it is the model’s size in memory. Ollama’s default downloads are 4-bit versions, which need roughly these amounts of GPU memory with room for a normal chat context. The table shows the cheapest card on the board today with at least that much memory and what four hours of it costs.
| For | USD | INR | NPR |
|---|---|---|---|
| Llama 3.1 8B or Gemma 2 9B, about 6 GB in use | USD 0.32 | INR 32 | NPR 48 |
| Qwen 2.5 14B or Phi-4 14B, about 10 GB in use | USD 0.32 | INR 32 | NPR 48 |
| Qwen 2.5 32B or Gemma 2 27B, about 20 GB in use | USD 0.72 | INR 68 | NPR 104 |
| Llama 3.1 70B or Qwen 2.5 72B, about 42 GB in use | USD 1.56 | INR 148 | NPR 232 |
Step by step
- Create an account and add credits. Google or an email address; then Credits, where the smallest top-up is USD 5.00, INR 200 or NPR 200. There is no ID upload.
- Pick the software, then the GPU. Choose Ollama + Open WebUI (chat), then a card with enough memory for your model from the live board. Every offer shows the host’s country, machine and measured reliability. Press Rent.
- Wait for Running. About a minute on a warm host, a few minutes on a cold one. Then open the app: Open WebUI appears in your browser. The first visit asks you to create the admin login for your own instance.
- Pull a model. In the model picker, type a name such as
llama3.1:8borqwen2.5:14band choose “pull from Ollama.com”; the download takes a minute or two on a datacentre connection. The same works from a terminal:ollama pull llama3.1:8b ollama run llama3.1:8b - Chat. Pick the model at the top of the chat and talk to it. Upload a PDF to ask questions about it; Open WebUI handles the retrieval. Nothing you type leaves the machine you rented.
- Stop when you are done. GPU billing ends at once. The disk, with the models you pulled, is kept at the storage rate shown before renting, so the next evening starts without downloads; destroy the instance to stop paying for the disk too.
What an evening costs
Four hours of a 13B-class model served for chat, on the cheapest hosts today:
| GPU | Typical time | USD | INR | NPR |
|---|---|---|---|---|
| RTX 3090 | 4 h | USD 1.20 | INR 112 | NPR 176 |
| RTX A5000 | 4 h | USD 1.48 | INR 140 | NPR 220 |
| RTX 4090 | 4 h | USD 2.40 | INR 224 | NPR 352 |
| RTX A6000 | 4 h | USD 2.60 | INR 244 | NPR 384 |
| A100 SXM4 | 4 h | USD 2.60 | INR 244 | NPR 384 |
Tips
- A model that only just fits runs slowly and may fail with a long conversation. Leave a quarter of the memory free: for a 20 GB model pick a 24 GB card, for 42 GB pick 48 GB.
- Ollama answers an API on the same machine, so a script or app can use the model you are chatting with; Open WebUI also offers an OpenAI-compatible endpoint under its settings.
- Set a timer or an idle stop on the machine when you rent it, and a daily spending limit on the account, so a forgotten tab cannot run all night.
- Want the model to speak your language better, or follow your own documents? Fine-tune it first: the Llama fine-tuning guide and the guide to teaching a model a new language walk through it.
Questions
Do I need to install anything on my computer?
No. Open WebUI runs in your browser; the model runs on the rented GPU. A terminal over SSH is available if you want it, but nothing in this guide needs it.
Which model should I start with?
Llama 3.1 8B or Qwen 2.5 14B: both are quick on a 24 GB card and good at everyday writing, code and questions. Move to a 32B or 70B model when the answers are not careful enough, and rent a bigger card for that evening only.
Is my conversation private?
The model runs on the machine you rented, and Open WebUI stores the chat history on its disk, not with any AI company. Hosts are verified operators; for sensitive material pick a datacentre host and destroy the instance when you are done.
How much does it cost to leave it running?
By the second while the machine runs, at the hourly price locked when you rented it (USD 0.18 · INR 17 · NPR 26 today for the cheapest 24 GB card), and only the disk while it is stopped.