Ethos / Private AI Foundation

Start small.
Stay sovereign.

One secure local AI node that runs open models privately and reaches frontier APIs only when policy allows.

Private by defaultModel-neutralOffice-friendlyUpgradeable

$10–15KComplete installed starter system

The recommendation

One machine. Useful headroom.

01 / COMPUTE

M3 Ultra

Mac Studio

256 GB unified memory, 2 TB SSD and built-in 10GbE. Quiet enough for the office; large enough for serious local inference.

02 / LOCAL MODELS

70B–120B

Practical private tier

Run useful quantized open-weight models locally. Smaller 8B–32B models remain fast and inexpensive for routine agents.

03 / OPERATIONS

~270 W

At maximum tested load

No server room, liquid cooling or specialist datacenter operations required. Typical mixed use will be lower.

Sources: Apple Mac Studio specifications · Apple power and thermal data

Simple bill of materials

Everything needed to start safely.

Component	Purpose	Budget
Mac Studio M3 Ultra · 256 GB · 2 TB	Local AI compute	$6K–$8K
8 TB encrypted Thunderbolt NVMe	Fast model storage	$700–$1.5K
Encrypted NAS · 20–40 TB usable	Versioned backup	$2K–$4K
1,500 VA pure-sine UPS	Clean power + shutdown	$600–$1.1K
Firewall + managed switch	Isolated private network	$500–$1.5K
Complete starter system		$10K–$15K

Model-neutral by design

Local first. Frontier when approved.

Ethos work

Apps + agents

Research
Documents
Internal workflows

→

Control point

Private model gateway

Identity + permissions
Data classification
Budgets + audit trail
Model aliases, not vendor lock-in

→

Destinations

Local + approved cloud

Open weights stay local
Closed frontier via API
Sensitive data never leaves

Gateway reference: LiteLLM · Local runtimes: Ollama · llama.cpp · MLX

Clear boundaries

What this gives us — and what it does not.

What we gain

Private local inference for confidential work
Freedom to swap open-weight models
Controlled access to closed frontier APIs
Encrypted storage and versioned backups
A reusable agent, evaluation and data layer

What we are not buying

A datacenter or GPU cluster
High-concurrency production serving
Frontier-model pretraining infrastructure
Local access to closed proprietary weights
A machine capable of hosting 2.8T K3-class models

Recommended path

Buy once. Measure before scaling.

Install the private nodeFileVault, isolated network, encrypted backup and UPS.

Put one gateway in frontLocal models by default; approved frontier routes with limits.

Operate for 90 daysMeasure quality, latency, concurrency, privacy exceptions and API cost.

Scale only on evidenceAdd a second Mac for resilience or rent NVIDIA capacity before buying it.

Decision: approve a $10K–$15K sovereign AI starter node and keep every model replaceable.

Security reference: Apple FileVault and Secure Enclave

Start small.Stay sovereign.