Choosing where to run AI: shared hosting, a server of your own, or a GPU

Where you run AI depends on what you want it to do. If your website calls a hosted model such as ChatGPT or Claude, the model runs at the provider and your hosting only sends requests, so shared hosting is enough. If you want an agent or an automation that runs all day, it needs a server of its own. If you want to run an open model yourself, so you control where your data lives, you need enough memory for the model and, for anything beyond light use, a graphics processor (GPU). This guide walks through each route and the Noiz service that fits it.

Three ways to use AI from your hosting

Route 1: call a hosted model from your website

The simplest way to add AI to a site is to have your website's server call a model provider's API. A visitor's browser talks to your site; your site talks to the provider and passes back the answer.

The rule that matters is where the key lives. OpenAI's API documentation says never to expose your key in browser or app code. Anthropic advises keeping keys out of code and configuration files, in environment variables or a secrets store, and rotating them regularly. Keep the key on the server and the browser never sees it.

Which Noiz plan. This runs on shared hosting. Legacy runs PHP, so a short server-side script can make the call. Premium adds SSH access to your own account and your choice of PHP version, so you or your developer can build a fuller integration. Premium's dedicated PHP pool also keeps your site responsive while it waits for the provider to answer.

Route 2: run an agent on a server of your own

An agent is a program that works on its own: it watches a mailbox, answers in a chat channel, runs a workflow each hour, or carries out multi-step tasks with tools. It still calls a hosted model, but it needs to run continuously, keep its own files and connect where it needs to.

Shared hosting is built for websites, not for long-running programs. A Noiz Cloud server gives you a Linux server with full root, Ubuntu or Debian, and one dedicated IPv4 address, so the agent can run as a service you control. You run the operating system and the agent, and look after its keys and updates.

Route 3: run an open model yourself

Open-weight models can be downloaded and run on your own server, so prompts and data are not sent to an outside provider. Three open-source tools do most of this work:

For a chat interface on top, Open WebUI is widely used. Check its licence before you brand it for a client: from version 0.6.6 it carries its own licence with a branding clause, which its makers say is not an approved open-source licence.

How much memory a model needs

The model has to fit in memory: the GPU's memory if you use one, otherwise the server's own RAM, with room for the conversation on top. As a rough guide from Ollama's model library, as at October 2026:

• Small models (1 to 3 billion parameters) download at 1 to 2 GB.
• Mid-size models (around 8 billion) download at about 5 GB.
• Large models (around 70 billion) download at over 40 GB.
• The largest run to hundreds of gigabytes and need several data-centre GPUs.

CPU or GPU

Ollama and llama.cpp both run models on an ordinary processor with no GPU. That suits small models, light use and testing: summarising documents overnight, tagging records, or a tool used by one or two people. For a model that many people use at once, or a larger model that must answer while someone waits, a GPU is the practical choice.

Which Noiz service fits

Ask us where your AI should run

Tell us what you want AI to do, how many people will use it, and whether your data may leave your server. We will say which route fits. Open a ticket.

Sources

Checked 19 October 2026. Model sizes and tool versions change; these are as at October 2026.

Keys

• OpenAI API reference: authentication
• Anthropic: API key best practices

Running models

• Ollama
• Ollama: Linux
• Ollama: GPU support
• Ollama: FAQ
• Ollama library: Llama 3.1
• Ollama library: Llama 3.2
• llama.cpp
• vLLM documentation
• Open WebUI licence

Noiz

• Shared hosting
• Cloud servers
• Dedicated servers

Note: while some of this content was produced with the help of AI, the final editing has been done by a real human being.