← Back to Course|Hardware Guide

🖥️ Buying a Computer for AI

If you only use cloud AI (Claude, ChatGPT), any modern computer works fine. This guide is for people who want to run AI models locally — which requires specific hardware considerations. We cover Mac vs Windows, RAM requirements, and specific recommendations by budget.

⚡ TL;DR — The Quick Answer

For local AI: Get an Apple Silicon Mac

The M-series chip's unified memory architecture makes it the best value for running large models locally. An M3 MacBook Pro with 36 GB RAM runs Llama 70B smoothly — something that would cost $3,000+ in a Windows GPU setup.

Just using cloud AI? Any modern laptop works

Claude, ChatGPT, Gemini, and Copilot run in a browser. Any laptop made in the last 4 years with 8 GB RAM and a decent internet connection handles them perfectly. No special hardware needed.

🍎 Mac vs 🪟 Windows for Local AI

Factor🍎 Mac (Apple Silicon)🪟 Windows (+ NVIDIA GPU)
Local model speed✅ Excellent (unified memory)✅ Excellent (dedicated VRAM)
Running 7B models✅ Any M-chip Mac (16 GB)✅ Any RTX 3060+ (8 GB VRAM)
Running 70B models✅ 36–48 GB unified RAM❌ Needs 2× A100 or H100 (~$30K+)
Cost for 70B capable💰 ~$2,500 (M3 Pro 36 GB)💰 $10,000–$30,000+
Battery life✅ Excellent (12–18 hrs)⚠️ Poor during GPU tasks
CUDA ecosystem (PyTorch)⚠️ MPS backend (some gaps)✅ Full CUDA support
Ollama support✅ Native, very fast✅ Works via NVIDIA CUDA
Cloud AI (Claude/Copilot)✅ Same as any machine✅ Same as any machine
Price/performance for local AI✅ Best value overall⚠️ High GPU cost premium

🍎 Apple Silicon — Recommended Configurations

Apple's M-series chips use unified memory — the CPU and GPU share the same RAM pool. This is why a 36 GB M3 Pro beats most Windows machines at local AI: the full 36 GB is available to the model.

Entry (~$1,300)
MacBook Air M3 — 16 GB RAM
Runs Llama 3.2 3B, Mistral 7B, phi3:mini comfortably. Good for learning and everyday AI assistance.
Good start
Mid (~$2,000)
MacBook Pro M3 — 18–24 GB RAM
Runs Llama 3.1 8B smoothly. Can run 13B models at acceptable speed. Good for developers.
Recommended
Pro (~$2,500–3,000)
MacBook Pro M3 Pro — 36 GB RAM
Runs Llama 3.1 70B at usable speed. Near GPT-4 quality locally. Best value for serious local AI work.
Best for local AI
Power (~$3,500–6,000)
MacBook Pro M3 Max — 48–128 GB RAM
Runs 70B models fast. Can run multiple models simultaneously. For ML engineers and researchers.
For professionals
Desktop (~$1,600+)
Mac Mini M4 Pro — 24–64 GB RAM
Best performance per dollar for a desktop AI workstation. Add your own monitor.
Best desktop value

🪟 Windows with NVIDIA GPU

Windows with a dedicated NVIDIA GPU is the other serious option for local AI. The advantage: full CUDA support for PyTorch/TensorFlow development and fine-tuning. The limitation: VRAM is the bottleneck (separate from system RAM), and 70B models need massive VRAM.

RTX 3060 (12 GB VRAM)
~$300Entry point
Llama 7B–13B well. Best budget GPU for local AI.
RTX 4070 (12 GB VRAM)
~$600Good mid-range
Llama 7B–13B fast. Better throughput than 3060.
RTX 4080 (16 GB VRAM)
~$1,000Recommended GPU
Llama 13B–34B models. Solid for development.
RTX 4090 (24 GB VRAM)
~$2,000Best consumer GPU
Llama 70B (quantized). Best consumer GPU for AI.
2× RTX 4090 (48 GB VRAM)
~$4,000+Enthusiast
Full 70B models at good speed. Serious AI rig.
⚠️ Key Windows limitation: VRAM is separate from system RAM. A 24 GB RTX 4090 cannot use your 64 GB system RAM for models. Compare: a 36 GB M3 Pro Mac can use all 36 GB for a model.

🐧 Linux (for Power Users)

If you're comfortable with Linux, Ubuntu + NVIDIA GPU is the most flexible setup for AI development. Full CUDA support, no licensing restrictions, best PyTorch performance. Same GPU recommendations as Windows apply. Ollama runs natively. The tradeoff is setup complexity — not recommended for beginners.

🎯 Our Recommendations by Use Case

📌 Learning AI, using cloud tools only
Any modern laptop with 8 GB RAM. Even a $700 laptop works.
📌 Developer using Copilot/Claude Code + occasional local models
MacBook Air M3 16 GB (~$1,300). Best all-rounder.
📌 Developer who wants serious local AI capability
MacBook Pro M3 Pro 36 GB (~$2,500). Run 70B models locally.
📌 ML engineer / fine-tuning / CUDA training work
Windows/Linux + RTX 4090 24 GB (~$2,000 GPU). Full CUDA ecosystem.
📌 Desktop workstation for AI research
Mac Mini M4 Pro 64 GB (~$2,400). Best local AI desktop value.
→ Run Local Models (Ollama)→ Free Local AI Guide