Why a philosophy?
Anyone can rent ChatGPT. Renting is fine — until the day it isn't. These six principles are a yardstick for telling a good AI setup from a fragile one: whether you're choosing a tool, vetting a vendor, or building your own system. Every page in this Academy assumes them.
1. Own the Stack
The most important decision in AI is not which model you use — it's who owns the computer it runs on. On infrastructure you control, you set retention policies, audit the logs, and decide what data leaves the building. Renting intelligence is convenient until it conflicts with compliance, privilege, or self-interest.
2. Zero Trust & Isolate by Default
Every customer, department, and project gets true isolation — separate containers, separate databases, separate memory — and nothing is trusted just because it's "inside." That's zero trust applied to AI: every request proves itself, every boundary is enforced. No shared inference pools that commingle prompts, no global embeddings that leak context between tenants. Isolation isn't paranoia; it's the minimum viable security model.
3. Data Dignity
Your documents, conversations, and embeddings deserve the same respect as financial or medical records. Good systems don't train on your data and don't retain your prompts. When you do use a hosted service, insist on Zero Data Retention (ZDR) — a guarantee that nothing is kept after your prompt is processed. Never assume it's the default: ZDR is usually a setting you enable or a contract you sign. What you generate belongs to you, lives on your servers, and follows your retention policy — not a third-party provider's.
4. Honest Trade-offs
AI is not magic, and anyone selling it as such is selling you something. Self-hosted inference gives total control but demands operational expertise. Managed private platforms absorb that labor at a premium. Public APIs are cheapest — and place your data in someone else's trust model. Demand real numbers for each option and choose for yourself.
5. Built for the Real World
Most organizations have 10–500 users, finite IT budgets, and real compliance obligations. A sensible AI design assumes no on-site GPU cluster, no dedicated ML engineer, and no appetite for six-figure experiments. A 40-person firm should be able to run private AI on a single modest server with predictable monthly costs.
6. Transparent by Design
Open-source components. Public licenses. Documented configuration. No proprietary black boxes, no vendor lock-in. If you ever decide to move your AI elsewhere, you can — because the components, the data, and the configuration are yours.
The "how" behind these principles: the Sovereign AI track turns them into practice — what it means to own the model, the data, the infrastructure, and the policy, plus a decision checklist for what belongs where.