PerspectiveOwning your knowledge

Sovereignty is control, not location

Ask most AI suppliers whether their product is sovereign and you will hear the same reply: the data is hosted in Australia. It is a true answer. It is also an answer to the smallest of the questions that matter.

KYROGA consultant

Hosted in Australia is one answer to four questions.

Location tells you where the disks are. It does not tell you who can reach what is on them, who decides when the system changes, whether you can walk away with what it has learned, or whose law can compel access to it. A system can sit in a Sydney data centre and still fail every one of those.

The market has a name for selling the postcode as the whole answer: sovereignty washing. The fix is not a better postcode. It is a better definition.

Sovereignty is control, tested four ways.

Sovereignty is control, not location. I use four tests. A system is only as sovereign as its weakest one.

  1. Data

    Where does your knowledge sit, and who can reach it?

    What failing looks like
    Your documents and every question your people ask are copied into a supplier's index you can't inspect.

  2. Operational

    Who runs it, audits it, and can switch it off?

    What failing looks like
    The supplier's operations team decides when it changes, and your audit trail lives in their format.

  3. Model

    Can you change the model without rebuilding what you know?

    What failing looks like
    Your knowledge is entangled with one vendor's model, so changing provider means starting again.

  4. Jurisdictional

    Whose law can compel access?

    What failing looks like
    A local subsidiary of a foreign provider, subject to that country's disclosure laws.

The jurisdictional test is the one location answers least well. In June 2025, under oath before a French Senate inquiry, Microsoft France's director of public and legal affairs was asked whether he could guarantee French customer data would never be handed to US authorities without France's agreement. "No," he said, "I cannot guarantee that," adding that it had never happened. The data was in France. The control was not.

Why local is now a practical choice.

For years the honest objection to sovereign AI was capability. If you ran your own models, you ran worse ones. That objection has mostly expired. Even Y Combinator now asks founders to help companies own their intelligence, observing that "nearly every company building with AI rents its intelligence".

Open models have closed the gap with frontier models for about 95% of what organisations need AI for: finding and summarising what they already know, drafting, answering questions from their own records, and working through routine decisions. They run on hardware you own, with no per-token meter running.

The remaining 5% is real. The hardest reasoning, long multi-step agent work and some multimodal tasks still favour the frontier. The answer is not to pretend otherwise. It is to design the route: local where it matters, and a governed path to a frontier model where it pays, with your policy deciding what is allowed to make that trip.

Every question is a deposit.

The four tests are about risk. The bigger reason to care is value.

Every question your people ask an AI system is a deposit. It records what the business needed to know, what it already knew, and what it decided. In a rented system those deposits land in someone else's account. They improve someone else's product, and they stay behind when you leave.

In an owned system they compound. Each answer that gets checked and corrected becomes knowledge the next answer starts from. After a year, the gap between those two organisations is not the model they used. It is what each of them kept.

Six questions to put to any AI supplier.

Ask for the answers in writing. How readily a supplier gives them is part of the evidence.

  1. Where does our data sit, and which of your staff or subprocessors can reach it?
  2. Can the whole system run on our premises or devices, and what stops working if it does?
  3. If we change model provider, what do we have to rebuild?
  4. Can we export everything the system has learned, in a documented format, and restore it elsewhere?
  5. Which country's laws can compel you to hand over our data, including through a parent company?
  6. For any answer the system gives, can you show us its sources and replay how it was reached?

For APRA-regulated entities, questions 3 and 6 are no longer optional. APRA's April 2026 letter to industry on AI expects entities to understand their AI dependencies, including third and fourth parties, and calls for credible fallback processes where AI supports critical operations. It also notes that upstream dependencies such as foundation models are often opaque.

Start with the question, not the postcode.

None of this means every organisation needs to unplug from the cloud tomorrow. It means sovereignty is a design decision, made test by test, for the capability you are improving. Local by default; data leaves only where your policy allows it to.

The postcode is still worth asking about. It is just the first question, not the last.

Sources

Share this article

Read next

Keep what you learn.

BrainBuild brings what your organisation knows together, puts it to work on a real problem, and keeps it yours.

Or just talk to us