Everything I run — Ollama for local LLMs, ComfyUI for image generation, n8n for workflow automation, OpenClaw for my AI agent — lives on a single Mac mini M4 with 32GB of unified memory. It costs roughly $8-10/month in electricity and replaced $80-100/month in cloud subscriptions.
Why the Mac Mini M4
The Mac mini's secret weapon is unified memory architecture. Unlike a PC where CPU and GPU have separate memory pools, the M4 shares everything. Every byte of your 32GB is available for AI inference. A gaming PC with 32GB of system RAM but only 8GB of GPU VRAM can't touch what this machine does.
It runs silently, draws 15 watts idle, fits in your palm, and costs $799 for the 32GB configuration. For an always-on AI server, nothing else comes close on value.
Recommended Build — Under $1,000
Mac mini M4, 32GB unified memory: $799
Ollama + ComfyUI + n8n + OpenClaw: $0 (all open source)
Monthly electricity: ~$8-10
Why 32GB not 16GB: When macOS swaps model weights to SSD, performance drops 97% — from 40 tokens/sec to under 0.3. That cliff is at 16GB the moment you load a model plus any other tool. 32GB gives you breathing room to run multiple AI services simultaneously.
Step 1: Install Ollama (5 minutes)
Ollama makes local LLMs as simple as a terminal command. No Python setup, no Docker, no dependency management.
On the M4 with 32GB, Llama 3.1 8B runs at 35-45 tokens per second — well above comfortable reading speed. Start with three models: llama3.1:8b (general purpose), deepseek-coder-v2:16b (coding), and mistral:7b (fast lightweight tasks).
Step 2: Install ComfyUI for Image Generation (15 minutes)
ComfyUI runs Stable Diffusion locally. I use SDXL Base 1.0 for thumbnails, YouTube scenes, and social graphics — replacing a $30/month Midjourney subscription. Images generate in 45-60 seconds at 1344x768 on the M4. Not the fastest, but free and completely private.
Step 3: Install n8n for Automation (10 minutes)
n8n is the automation backbone connecting everything. Web interface on localhost:5678. I run 40,000-60,000 executions per month at zero cost. Your first workflow: daily email summary → Ollama summarization → Telegram notification. Build it in 20 minutes.
Step 4: Install OpenClaw (15 minutes)
OpenClaw transforms the Mac mini from a passive server into a proactive agent. It connects to Telegram and gives your AI access to tools — files, browser, shell, APIs. Full setup in my OpenClaw 101 guide.
Step 5: Configure Always-On Operation (10 minutes)
Disable sleep. Enable auto-restart after power failure. Enable SSH for remote access. Create launchd services for auto-start. Install Tailscale for remote access from anywhere. After this, your Mac mini runs headless — managed entirely from your laptop or phone.
What This Replaces
| Cloud Service | Monthly Cost | Local Replacement |
|---|---|---|
| ChatGPT Plus | $20/mo | Ollama (free) |
| Midjourney | $30/mo | ComfyUI (free) |
| n8n Cloud | $20/mo | Self-hosted (free) |
| AI APIs | $10-30/mo | Ollama API (free) |
| Total saved | $80-100/mo |
At $80-100/month saved, the $799 Mac mini pays for itself in 8-10 months. I'm 6+ months past break-even. Every month from here is pure savings.
Honest Limitations
32GB can't run 70B+ parameter models — need 64GB minimum. Image generation batches are slow versus a dedicated GPU. Fine-tuning requires cloud compute. For everything else — running LLMs, generating images, automating workflows, hosting agents — this machine handles it all without breaking a sweat.