Define the budget goal first

For learning, almost any supported PC with enough free storage can be worth trying. For private chat and short summaries, 16 GB RAM can work, but 32 GB is better. For coding help and document workflows, a budget machine can become frustrating unless it has enough memory and a decent GPU.

For a small team, a budget PC is risky as the main server unless the workload is light and someone owns backups, updates, and security.

Cheapest path: use your current machine

The cheapest useful local AI setup is the computer you already own. Install Ollama or LM Studio and run a small model.

ollama run llama3.2:3b

If that is slow, try:

ollama run gemma3:1b

This test tells you whether the current machine is close enough or whether you need RAM, SSD space, GPU power, or a different path.

Budget desktop path

A used or refurbished desktop is often better value than a budget laptop for local AI because desktops are easier to upgrade. Look for:

  • 32 GB RAM or cheap RAM upgrade support.
  • 1 TB SSD or space to add one.
  • Standard power supply options if adding a GPU.
  • Enough case airflow.
  • Supported Windows or Linux path.
  • A GPU slot if you plan to upgrade later.

The risk is that used office desktops may have weak power supplies, cramped cases, limited GPU clearance, or proprietary parts. Check before buying.

Budget GPU reality

For language models, VRAM matters. A cheap GPU with too little VRAM may not improve much. A used GPU can be good value, but only if it is healthy, supported, and has enough VRAM for the models you want.

Rough planning:

  • 4 GB VRAM: small models and learning.
  • 8 GB VRAM: better small-model experience.
  • 12 GB VRAM: stronger budget target.
  • 16 GB+ VRAM: better headroom but usually higher cost.

Do not buy a used GPU without checking power requirements, physical size, seller condition, and driver support.

Budget laptop path

A budget laptop is convenient but less ideal for local AI. It may have soldered RAM, limited VRAM, weak cooling, and no upgrade path. If you need portability, buy more memory up front and accept smaller models.

For local AI, a cheap 8 GB laptop is usually a poor new purchase. A 16 GB laptop is a better minimum. A 32 GB laptop is much more useful, but can push the price out of true budget territory.

Budget Mac path

Apple Silicon can run local models well when unified memory is sufficient. The problem is that entry-level memory and storage can become limiting. If buying for local AI, treat 16 GB unified memory as the minimum practical tier and 32 GB as a better target.

A small Apple desktop can be a quiet local AI machine, but upgrades are limited after purchase. Price the memory and storage you actually need, not only the base model.

Cloud AI versus budget PC

Sometimes the cheapest local AI PC is not a PC. If you only need occasional high-quality answers, a subscription or API may cost less than hardware. If you need private drafts, offline use, repeated experiments, or local control, budget hardware can make sense.

Compare the whole cost:

  • Hardware purchase.
  • RAM and SSD upgrades.
  • Electricity.
  • Your setup time.
  • Backups.
  • Optional cloud API usage.

If you buy a $900 budget PC to avoid a $20 monthly subscription, the payoff period is long unless you also need local privacy or offline use.

Used and refurbished buying checks

Used hardware can make local AI affordable, but it shifts risk onto you. Before buying, confirm the exact CPU, RAM capacity, SSD size, GPU model, VRAM amount, power supply wattage, and return policy. Ask whether the system has been stress tested. For a GPU desktop, confirm that the case has airflow and that the power supply has the right connectors.

Avoid vague listings such as "AI ready" without real specs. For language models, a clear 32 GB RAM and 12 GB VRAM listing is more useful than a flashy product photo. Also check noise and power draw if the machine will sit in an office.

If you are buying for a small business, warranty and replacement speed may matter more than the lowest used price. A cheap machine that fails during a project is not cheap.

Upgrade order

Upgrade in this order:

  1. SSD space if your system drive is nearly full.
  2. RAM if the machine supports it and memory pressure is obvious.
  3. GPU only after you know which model sizes you want.
  4. Whole system replacement if the PC is unsupported, non-upgradeable, or thermally weak.

Do not start with the GPU if the machine has 8 GB RAM, a tiny SSD, and poor cooling.

Red flags

Avoid machines with unsupported Windows versions, soldered low memory, tiny SSDs, no room for upgrades, old GPUs with weak driver support, or proprietary power supplies that block a GPU upgrade.

Be careful with used gaming PCs. They can be good value, but the GPU, power supply, storage, and cooling condition matter.

Bottom line

The best budget local AI PC is usually an upgradeable desktop, not the cheapest laptop. Start on your current machine, aim for 32 GB RAM and a 1 TB SSD if buying, and treat GPU VRAM as the key budget tradeoff. Use cloud AI when the local hardware math does not make sense.