AI computing requirements: how much hardware do you need?

Published October 27, 2026 · 9 min read

Read in: Português · English · Español

Artificial intelligence has become the world's biggest consumer of computing power. Tech companies are building data centers the size of neighborhoods to train and run models. But what about you, if you just want to use AI in a project, a bot or an automation? Do you need a graphics card that costs thousands, or will a regular VPS do? This guide explains where all that hardware demand comes from and how many resources each kind of use really needs.

Why does AI need so much computing power?

A language model like the ones behind today's chatbots is, at its core, a huge collection of numbers called parameters (or weights). Small models have a few billion of them; the largest have hundreds of billions. To generate each word of an answer, the computer has to do math with practically all of those numbers.

That creates two main requirements:

• Memory: all the parameters must be loaded into memory (system RAM or the graphics card's memory) while the model is working.
• Memory speed: for every word it generates, the model has to read those parameters again. That is why, in AI, the bottleneck is usually memory bandwidth rather than the number of processor cores.

This is why GPUs (graphics cards) dominate AI: they run thousands of calculations in parallel and use memory that is much faster than a server's regular RAM.

Training an AI vs. using an AI

A model goes through two very different stages, and their costs are not even in the same league.

The good news: almost nobody needs to train a model from scratch. What you will do is inference, either by calling the API of a service such as ChatGPT, Claude or Gemini, or by running an open model (Llama, Qwen, Gemma, Mistral, DeepSeek and others) on your own server. There is also fine-tuning, which adapts an existing model to your data and sits in between: it needs a GPU, but far less than full training.

How much memory does an AI model need?

You can estimate it with a simple formula: number of parameters × bytes per parameter, plus some headroom for the conversation context. Bytes per parameter depend on quantization, a technique that stores the numbers with less precision to save memory:

• 16-bit (no quantization): 2 bytes per parameter.
• 8-bit: about 1 byte per parameter, with almost no noticeable quality loss.
• 4-bit: about half a byte per parameter. This is the most common format for running AI at home or on modest servers, with a small loss in quality.

In practice, for models quantized to 4 bits:

CPU or GPU: what changes in practice

An AI model can run on a processor alone, with no graphics card. The difference is how fast the words come out:

• CPU only (regular VPS): models with 1 to 4 billion parameters respond at an acceptable pace for bots and automations. 7 to 8 billion parameter models work, but usually generate only a few words per second, especially on shared vCPUs. Good for background tasks, not for a chat that needs to feel instant.
• GPU: the same answer comes out many times faster, and you can serve several people at once. For image generation (Stable Diffusion, Flux) and models above 13 billion parameters, a GPU is practically mandatory.

Some AI workloads are light and run fine on a CPU: embeddings (turning text into vectors for search and RAG), text classification, OCR and audio transcription with the smaller Whisper models.

Three ways to use AI in your project

1. An AI provider's API

You send the text and get the answer back, paying per use (per token). The most capable models on the market are only available this way. It requires no hardware beyond the server making the calls.

• Best answer quality and zero model maintenance
• Cost grows with volume
• Your data goes through the provider

2. An open model on a CPU-only VPS

A VPS with plenty of RAM runs small models for a fixed monthly cost, with your data under your control. It works well for summarizing text, classifying messages, extracting data from forms and answering simple questions in automated workflows.

• Fixed cost, no per-token billing
• Privacy: nothing leaves your server
• Slower and less capable answers than the large models

3. A GPU server

Needed for medium and large models at good speed, image generation and fine-tuning. GPU servers cost much more than a regular VPS and are usually rented by the hour from cloud providers that specialize in AI.

• Speed and capacity for larger models
• Expensive to leave running all month
• Makes the most sense for heavy use or short periods

Which VPS to choose for each AI use

Hostinger KVM — easy to start, data centers in several regions

For projects that call AI APIs and run n8n, bots and agents, Hostinger is the most practical option: a simple control panel, ready-made application templates and data centers in several regions, so you can pick the one closest to your users. Plans with more vCPU and RAM can also handle small models with Ollama.

Contabo — lots of RAM for the lowest price

If you plan to run local models on a CPU, memory is what matters most, and Contabo gives you more GB of RAM per dollar than most providers. Ideal for background AI tasks.

Hetzner — dedicated CPUs

Hetzner's dedicated vCPU plans avoid the performance swings of shared CPUs, which helps when the model runs on the processor all the time. Hourly billing is also handy for testing without commitment.

How to test an AI model on your VPS with Ollama

Ollama is the simplest way to run open models on a Linux server. On an Ubuntu server that is already secured:

# install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# download and chat with a small model
ollama run llama3.2:3b

# see how much memory the model is using
ollama ps

Ollama also provides a local API on port 11434, which you can call from n8n, a bot or your own code. Start with a small model, measure the speed, and only then try larger ones.

Tips to avoid wasting resources

• Use the smallest model that does the job. Tagging a message as "question" or "complaint" doesn't need a giant model.
• Prefer quantized models (4- or 8-bit): they fit in less memory and run faster.
• Create swap as a safety net, but don't rely on it to load the model: running AI from disk is extremely slow.
• Monitor RAM with htop or free -h. If the system starts killing processes for lack of memory, it's time for a bigger plan or a smaller model.
• Split the work: send what needs high intelligence to the API and the repetitive work to the local model.

Summary

AI needs a lot of computing power mainly because of memory: the whole model has to fit in it, and the faster it is, the faster the answers come out. Training models is data center work, but using AI is within reach of any project. For most people, a 4 to 8 GB VPS calling an AI API is the best value. Small models run fine on a CPU alone; large models, real-time chat and image generation call for a GPU.