The Hardware
Apple Mac mini — one node of the AI Subfactory
Apple M4 Pro
64 GB unified memory
20-core GPU
16-core Neural Engine
Thunderbolt 5
Product image © Apple Inc.
Inside the Mini — silicon at work
Animated cutaway (cover off) · driven by the simulated telemetry below · component layout representative, not exact
Memory read
Memory write
Thunderbolt · RDMA
NVMe I/O
Modelled savings vs. cloud API pricing
$0
▲ modelled accumulation · not a measured figure
Loading today's win…
Data Sovereignty
100%
0 bytes left the building today
Tokens Processed Today
0
Avg Time-to-First-Token
— ms
vs ~850ms typical cloud RTT
Cluster Uptime
99.97%
24 / 7 dedicated agent box
Rate Limits Hit
0
unlimited local throughput
Hardware Payback
—%
of $8,000 cluster recouped
Apple Silicon Telemetry — simulated stream, cluster-wide
Throughput per node (tok/s)
4 nodes · rolling 60s window
| Time | qwen3:32b | deepseek-r1:32b | deepseek-r1:70b | nomic-embed |
|---|
Compute utilization
Cluster average · GPU vs Neural Engine
| Time | GPU % | ANE % |
|---|
Unified memory bandwidth
DRAM read / write · GB/s, cluster sum
| Time | Read GB/s | Write GB/s |
|---|
Power & thermal envelope
Separate scales by design — never dual-axis
Power draw (W)
Peak node temperature (°C)
Cluster Topology
Why Local — the argument, made visible
Cost per 1M tokens
Local marginal cost vs. blended cloud API rate
Response latency
On-device inference vs. typical cloud round trip
Sovereignty rate
100%
all inference on-devicetarget 100%
Hardware payback progress
—%
$8,000 cluster costbreakeven ≈ 18–24 mo
Concept console. Every number on this page is simulated; none of it is live telemetry from the subfactory. The panels are shaped for the real feed (per-node tokens/sec, GPU and Neural Engine utilisation, memory bandwidth, power and temperature per node, streamed from the machines themselves). When that collector ships it will appear in the build log first, and this caption will change. Cost, latency and payback figures are planning assumptions, not measurements.