What Is Private AI?

How is private AI different from ChatGPT, Copilot or Gemini?

Public AI tools — ChatGPT, Microsoft Copilot, Gemini — run on the vendor’s shared, multi-tenant infrastructure. Your prompts and documents travel to their platform, are processed under their terms, and the vendor decides where data flows, which model version you get, and when it changes. Enterprise plans improve the contract around that arrangement — better data-handling terms, admin controls, sometimes a promise not to train on your data. What they don’t change is the architecture.

Private AI changes the architecture: the model, your data, the logs and the access rules all sit inside a boundary your organisation defines and governs.

Private GPT, private LLM, local AI, self-hosted AI — are these all the same thing?

Mostly, yes — they’re overlapping names for the same idea, each emphasising a different aspect:

• "Private LLM" or "private GPT" — the model component: a large language model running under your control instead of a public service.
• "Self-hosted AI" and "local AI" — emphasise where it runs: infrastructure you operate.
• "On-premise AI" (or on-prem) — your own physical servers specifically.
• "Air-gapped AI" — the strictest variant, with no outbound internet connection at all.
• "Confidential AI" — usually refers to hardware-level protections around data in use.
• "Sovereign AI" — widens the lens from one organisation to a jurisdiction: keeping AI capability, data and governance within national (or iwi, or organisational) control.

A real deployment usually combines several of these: a private LLM, self-hosted in your tenancy or on-premise, configured to whatever isolation level your obligations require.

Does private AI mean building our own model from scratch?

No — and this is the most common misconception. Modern private AI deployments start from a strong existing open-weight or licensable model, then make it useful and safe with architecture: a retrieval layer (RAG) that grounds answers in your organisation’s approved documents, access controls that decide who and what it can see, and logging so every interaction is reviewable.

You’re not competing with frontier labs on model training; you’re controlling where an already-capable model runs and what it’s allowed to do. That’s design and engineering work measured in weeks, not a research programme measured in years.

How does a private AI deployment actually work?

A working deployment has four layers:

• The model — supplies language ability.
• The retrieval layer — connects it to your knowledge (policies, procedures, matters, records) so answers are grounded in what your organisation actually knows rather than the model’s general memory.
• The governance layer — enforces boundaries: which users can ask what, which documents are in scope, which actions require a human.
• The operations layer — keeps it healthy: monitoring, logging, deliberate model updates.

Sovata’s four-step delivery process — define the problem, set the boundaries, prepare the knowledge, deploy and operate — takes most organisations from workshop to working system in 6–12 weeks.

Where can private AI run? On-premise vs private cloud vs air-gapped

• Your own servers (on-premise) — the strongest control position: model, data and logs in your server room. Suits organisations with existing infrastructure or strict residency rules.
• A local data centre — dedicated infrastructure hosted onshore in your country, giving physical data residency without operating hardware yourself.
• Your own cloud tenancy — a dedicated, governed environment in an onshore cloud region: cloud convenience, but inside a boundary you control rather than a vendor’s multi-tenant platform.
• Air-gapped — any of the above with no outbound internet access, for defence-grade, health or other high-sensitivity contexts.

The right answer is driven by your obligations and your existing infrastructure, not ideology — an air-gapped deployment nobody maintains is worse than a well-governed tenancy. This is exactly what a readiness workshop settles.

Is private AI more secure than public AI tools?

It removes an entire category of risk — your data leaving infrastructure you govern — and it makes your existing security controls apply to the AI, instead of requiring you to assess a third-party platform you can’t inspect.

But honesty matters here: a badly configured private deployment is not automatically safer than a well-configured public one. Security still depends on access design, monitoring and governance. What private AI gives you is the ability to enforce those things yourself, with an audit trail, rather than trusting a vendor’s word.

What does private AI cost?

Public tools are cheaper to start and harder to predict: per-seat and usage-metered pricing scales with adoption in ways finance teams can’t forecast. Private AI inverts that: a higher upfront design-and-deployment cost, then running costs that are typically fixed and predictable.

Which is cheaper overall depends on scale, usage pattern and how much your sovereignty requirements are worth. For light usage with no strong data sensitivity, a public enterprise plan is often genuinely the better buy — and a credible private AI provider will say so. For organisations with heavy predictable usage or real residency obligations, the calculation flips.

Who actually needs private AI?

• Organisations handling regulated or client-confidential data — health providers, law firms, accountants, financial advisers — where sending records to a vendor platform conflicts with professional and legal obligations.
• Government and public-sector bodies expected to keep citizen data onshore and auditable.
• Māori organisations and iwi, for whom data sovereignty is a matter of governance over taonga, not just compliance.
• Enterprises with usage-metered AI bills growing faster than the value, or automations that broke when a public model changed overnight.
• Any organisation whose security team keeps vetoing AI projects — usually a sign the architecture, not the ambition, is wrong.

What is sovereign AI, and how does it relate to private AI?

Sovereign AI is private AI viewed at the level of a jurisdiction: the principle that a country, sector or people should be able to use AI without ceding control of its data, infrastructure and rules to offshore platforms.

In New Zealand this conversation spans the Privacy Act 2020, public-sector expectations about onshore citizen data, and Māori data sovereignty — the recognition that data about Māori is taonga and belongs under Māori governance. A private AI deployment is the practical mechanism: it’s how an individual organisation makes its AI use sovereign in fact, not just in policy.