The definitive Australian guide to buying salvaged and broken-screen M1 Max MacBook Pros for headless local LLM inference. What to look for, what to pay, what to avoid.
Jump to ChecklistBroken-screen M1 Max MacBook Pros are the best value-for-money hardware for running local LLMs in Australia. The screen is useless to most buyers — but for headless SSH-based LLM inference, you don't need it. Here's why they're incredible value.
| Chip | Apple M1 Max — 10-core CPU, 32-core GPU |
| Memory bandwidth | 400 GB/s (faster than most desktop GPUs) |
| Memory options | 32GB or 64GB unified memory (crucial for LLM model size) |
| Power draw (load) | ~30–40W (vs 300W+ for equivalent GPU rig) |
| Neural Engine | 16-core — supports MLX framework |
| New retail price (2021) | $3,500–4,500 AUD (64GB config) |
| Salvaged price (2026) | $800–1,800 AUD (see price guide below) |
| LLM inference (Llama 3.1 8B Q4) | ~30–40 tokens/sec decode |
| LLM inference (Qwen2.5 14B Q4) | ~20–25 tokens/sec decode |
| LLM inference (Llama 3.1 70B Q2) | ~3–5 tokens/sec (64GB only) |
What you should actually pay for a salvaged M1 Max in Australia, based on condition and specs. Prices in AUD, sourced from eBay, Gumtree, and Facebook Marketplace trends.
Use this checklist when evaluating any salvaged M1 Max listing. If you can't verify these points remotely, ask the seller — or walk away.
If the trackpad is clicking oddly or the bottom case is bulging, the battery is swollen. This is a fire risk. It's repairable ($200–300 AUD at an Apple service provider), but factor that into your offer price. A swollen battery means the machine has been stored in poor conditions (heat, full discharge) — check for other issues.
A cracked screen that still partially works is actually better for setup — you can do the initial macOS setup visually before switching to headless mode. A completely dead screen means you'll need an external monitor for initial setup, then switch to SSH. Both are fine for LLM use, but a dead screen should mean a lower price.
If Find My Mac / Activation Lock is enabled and the seller doesn't know the Apple ID password, the machine is permanently locked. There is no workaround. Apple will not unlock it without proof of original purchase. This is the most common scam — "it works fine, just needs a reset" usually means it's Activation Locked. Always verify Activation Lock is OFF before paying.
If the seller says "only one port works," that's usually fine for headless use — you only need one port for power and one for an external monitor (during setup). But if NO ports work, the machine is unusable. Test both ports during inspection.
Some salvaged units have had their SSDs replaced with third-party NVMe drives. This can cause boot issues, kernel panics, and reduced performance. If the SSD has been replaced, verify it's a compatible module and that the machine boots reliably. Third-party SSDs in M1 Macs can cause intermittent issues that are very hard to diagnose.
Ask how the machine was used. If it was a video editing workstation running at full load for years, thermal paste may be degraded. For LLM inference (sustained but not extreme load), this is usually fine. But if you see throttling to <50% performance within 5 minutes of inference, the thermal management may need attention.
Once you've bought your salvaged M1 Max, here's how to turn it into a dedicated local LLM inference server.
Connect an external monitor via USB-C/HDMI. Boot the machine, complete macOS setup, and create a local admin account. Enable SSH: System Settings → General → Sharing → Remote Login → On. Note the machine's IP address from System Settings → Network.
From your main machine, SSH in: ssh user@[ip-address]. Install Homebrew: /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)". Then install Ollama: brew install ollama. Start the Ollama service: ollama serve (or set up as a launchd service for auto-start).
Download a model to test: ollama pull llama3.1:8b. Once downloaded, test it: ollama run llama3.1:8b "Hello, world!". You should see token generation at 30–40 tokens/sec on a 32GB M1 Max.
Close the lid — the machine will sleep. To prevent sleep with lid closed, use: brew install --cask amphetamine (GUI tool) or set via terminal: sudo pmset -a disablesleep 1. Alternatively, use a magnetic sleep sensor bypass plug (~$5 on eBay) to trick the lid sensor into thinking the screen is open.
Place the machine on a cooling stand or laptop cooler for sustained inference loads. The M1 Max will throttle at ~95°C. A $30 cooling pad keeps temps at 75–80°C and prevents thermal throttling during long generation runs. Power: use the original 140W charger — lower-wattage chargers may not sustain peak performance.
Create a launchd plist for Ollama to auto-start on boot. Install monitoring: brew install htop for quick checks. For remote monitoring, set up a simple Prometheus + Grafana stack, or use Ollama's built-in stats. Consider installing ollama-model-manager for easy model switching and deletion.
Ollama runs an OpenAI-compatible API on localhost:11434. To use it from other machines on your network, set OLLAMA_HOST=0.0.0.0 before starting Ollama. For external access, use a reverse proxy (Caddy or nginx) with TLS and authentication. This turns your salvaged M1 Max into a private LLM API endpoint.
Check out our M1 Max benchmarks and headless rig tutorials on GeniusTechLab.
Read the Benchmark Reviews