Sovereign AI for New Zealand and Australian businesses.
What the phrase means once the marketing is taken out, the three options a business actually has today, when each fits, and what we are doing about it ourselves. Published 5 November 2026.
The short answer
Sovereign AI means being able to answer three questions about your AI without asking a vendor: where the model runs, where your data rests, and who can switch it off. A New Zealand or Australian business has three options today. A frontier model such as Claude pinned to an Australian cloud region keeps processing onshore. A private open weight model on hardware you own keeps everything inside your building. And the third is the one we are building ourselves: our own open weight model, starting on a Mac mini, as the proof that a business can run a capable model on its own premises.
What sovereign AI means, in three questions
Governments use the phrase about national capability. For a business it is simpler and more practical: how much of your AI could carry on tomorrow if a vendor, a cloud region or a foreign law changed its mind today.
Where does the model run?
The machine that turns your prompt into an answer has a location. It might be a vendor's data centre on another continent, a cloud region in Sydney, or a computer in your own building. That location decides which laws apply to the processing and who can see it happen.
Where does the data rest?
Prompts, documents and answers are stored somewhere, for some length of time, under someone's terms. Residency is about that storage. A vendor can process in one country and store in another, so the two questions have to be asked separately.
Who can switch it off?
A hosted model can be repriced, retired or made unavailable to you by a decision you were not part of. A model you run on your own hardware cannot. This is the question most residency conversations skip, and it is the one that matters most over five years.
The three options a business has today
Vendor facts below are the ones our comparison pages hold, checked against each vendor's own documentation on the dates shown there. Vendors change their plans often; check the date before you rely on a line.
A frontier model pinned to an Australian region
Claude through Amazon Bedrock in an Australian region processes every request onshore, inside an AWS boundary the business controls. Claude's own apps do not list Australian residency; ChatGPT Enterprise can store data at rest in Australia for eligible new customers, but OpenAI does not offer Australian processing.
Where Australian processing is accepted and you want frontier quality now. Common for New Zealand businesses.
Runs in Australia; rests in Australia; the vendor can still change the model.
A private model on infrastructure you own
An open weight model on hardware in your building or in a New Zealand data centre, run by you or by us, with a frontier model kept for the hard cases. Nothing leaves your boundary.
Where the rule says onshore or in-house: health information, some government and legal work, client contracts that name the country. Also where volume is high enough that owning beats metering.
Runs where you put it; rests where you put it; nobody can switch it off but you.
Our own open weight model, as we build it
The Engine: Sentry AI's own open weight model stack, starting on a Mac mini in our office, running our own operations and our daily eval. We publish what it manages and what it cannot on a dated build log.
Not available to buy today. Watch the log if you want to know, from real work, how much of a business's daily AI can run on a machine on a desk.
Runs on our desk; rests on our desk; the proof we are building towards on-premises custom models.
You rarely have to pick one. The common shape is a frontier model for the hard, open-ended work and a smaller model you control for the daily, repeated work, with a daily eval checking that the smaller one is still doing its share well. The open weight models page covers where that line usually falls.
Data sovereignty in New Zealand
New Zealand has no general law that keeps business data onshore. The Privacy Act 2020 governs cross-border disclosure rather than forbidding it, and processing in Australia is accepted by most New Zealand organisations and their customers. Three things raise the bar: health information, some government and legal work, and client contracts that name the country. For those, only a model running in New Zealand meets the rule, and today that means your own hardware or a New Zealand data centre.
Whether New Zealand should build the data centres that would make onshore frontier AI ordinary is a live argument, split between a climate camp and a growth camp. We wrote up the data centre debate and the middle path both sides skip. Until that settles, the practical answer for most businesses is Australian processing for frontier work and a local model for the data that cannot travel.
What we do about it
Today, and what we are building towards
Today we do two of the three options for clients. We deploy Claude pinned to an Australian region on infrastructure the client controls, and we build private LLM infrastructure: open weight models on hardware a business owns, with a frontier model for the hard cases.
The third option is ours to prove. The Engine is our own open weight model stack, starting on a Mac mini in our Auckland office, running our own operations and our daily eval first. It is the proof that a business can run a capable model on its own premises, and it is published as a dated build log. On-premises custom models are what it builds towards, not something we sell today.
Frequently asked questions
What does sovereign AI mean for a business?
That you can answer three questions about your AI without asking a vendor: where the model runs, where your data rests, and who can switch it off. The more of those answers are inside your own borders and your own control, the more sovereign the setup. It is a spectrum, not a badge.
Is Claude sovereign if we use it in Australia?
Claude's own apps do not list Australian data residency. Claude through Amazon Bedrock in an Australian region does process every request onshore, inside an AWS boundary the business controls. That answers where the model runs and where the data rests; the vendor can still change the model, so the third question is only partly answered.
Does data sovereignty in New Zealand mean the data has to stay in New Zealand?
Not by law for most businesses. New Zealand's Privacy Act 2020 governs cross-border disclosure rather than banning it, and Australian processing is widely accepted here. Health information, some government work and some client contracts set a stricter bar. Where the rule says onshore, only a model running in New Zealand, on your own hardware or in a New Zealand data centre, meets it.
Is an open weight model on our own hardware the most sovereign option?
Yes. The weights are yours to run, the machine is yours, and nothing leaves the building. The trade is capability and upkeep: a small local model is usually enough for narrow, checkable work, while long multi-step reasoning still favours a frontier model. Most businesses end up with both.
What is Sentry AI doing about it?
Two things. Today we deploy Claude pinned to an Australian region and build private LLM infrastructure on hardware a client owns. For the longer term we are building the Engine, our own open weight model stack, starting on a Mac mini in our office, to run our own operations and our daily eval. Its progress is published on the build log, and on-premises custom models are what that work builds towards.
Does sovereign AI cost more?
Usually it costs differently. A hosted API meters every token and can be repriced; hardware is paid once and then costs power and someone's time. Which is cheaper depends on volume and on how often the model gets the job right first time. We work it out as cost per completed task before recommending anything.
Which of the three answers does your business need?
In a short call we will look at the work, the data and the rule you have to meet, and say which option fits, including when a business plan on Claude or ChatGPT is simply enough.
Ask an AI about this page
Opens the assistant with this page loaded: read it, summarise it, cite it.