Apple cluster — quiet, unified memory
- Mac Studio M3 Ultra (Boss, up to 512 GB unified — typical Dream config ~192 GB) + 5× Mac Mini M4 Pro 64 GB
- Network: 10GbE switch · Ollama REST API
- Total hardware: ~€19,000–21,000 (confirm Boss tier on apple.com/ee before you buy)
- Power draw: ~0.66 kW · ~25 dB (quiet)
- Unified memory: Boss + 5×64 GB — 70B+ on Boss, 32B-class on each Mini
OPENCLAW LOCAL AI NETWORK · 10GbE · OLLAMA REST API
⚡ OPENCLAW BOSS
Mac Studio M2 Ultra
192 GB Unified Memory · 24-core CPU · 76-core GPU
Qwen 2.5 72BDeepSeek-R1 70BOLLAMA_NUM_PARALLEL=8
10 Gigabit Ethernet · REST API · Ollama 0.0.0.0
Marketing
Mac Mini #1
M4 Pro · 64GB
Qwen 2.5 32B
Sales
Mac Mini #2
M4 Pro · 64GB
Qwen 2.5 32B
Dev
Mac Mini #3
M4 Pro · 64GB
DeepSeek Coder 32B
Finance
Mac Mini #4
M4 Pro · 64GB
Qwen 2.5 32B
Ops
Mac Mini #5
M4 Pro · 64GB
Qwen 2.5 32B
Apple vs PC (honest trade-offs)
- Apple: ~€19–21k hardware · ~€95/mo electricity · ~25 dB
- PC: ~€30k hardware · ~€670/mo electricity · ~50–60 dB
- PC Boss VRAM bandwidth ~1.8 TB/s vs M3 Ultra ~819 GB/s — PC can be faster per token on 70B, at higher cost and noise
- Both: same Jerry + OpenClaw orchestration — departments are independent machines on LAN
No more “8×4090 NVLink” myth
This PC path uses one 96 GB workstation GPU for the Boss and RTX 5090 in each department — realistic, shippable, and technically honest.
Mac wins quiet ops and power bill; PC wins raw Boss bandwidth if you accept cost, heat, and build complexity.