Australia's sovereignty question has moved from storage to AI

Earlier this year, a single export directive issued in Washington forced a leading AI lab to switch off two of its newest models for every customer worldwide, Australian businesses included, for almost three weeks.

Australian organisations have spent a decade getting comfortable with data sovereignty, and most now consider the matter settled once their data is stored onshore. However, that comfort is misplaced. Knowing where your data is stored tells you very little about who controls the AI that reads it, where that AI runs, or whether a decision made in another country can take it away. For a growing number of businesses, those questions now decide whether operations keep running.

The rules still focus on where data is held

Australia’s rules largely reflect the older view. Data residency requirements, cloud assurance frameworks and most procurement checklists were designed around where data is held, and they do that job reasonably well. They say far less about what happens when a software vendor passes customer records to someone else’s model for a few seconds to draft a reply or score a lead.

Accountability, meanwhile, is moving towards the customer. The draft second tranche of Privacy Act reforms would make the business that controls personal information primarily responsible for it, including when a supplier does the processing. APRA already expects banks, insurers and super funds to see past their suppliers to the providers those suppliers depend on. The obligation sits with the customer, while the visibility often sits with a vendor several steps removed.

Small firms and large ones feel it differently

A 20-person accounting practice or a regional manufacturer rarely has a procurement team to read an AI provider’s terms. New AI features turn up inside software the business already pays for, and with them comes every downstream decision its vendors have made. The business carries the risk without ever having chosen it.

Large enterprises face the reverse problem. They may run dozens of software platforms, each embedding a different outside model. Their data residency paperwork can be immaculate while the reasoning over that data happens offshore, inside systems they have never assessed and could not easily leave.

Sovereignty now has three layers

AI has split one question into three. The first is data: where it is stored and whose law can compel access to it. The second is the model: who controls the AI working on that data, and whether access can be withdrawn. The third is compute: who controls the chips everything runs on, which governments now treat as an instrument of national policy.

Australia has a mature conversation about the first, a young one about the second and barely any about the third. The exposure is greatest in the layers nobody is asking about.

The market’s response has been to put a ‘sovereign’ label on almost everything. Some of those offerings are a local region of a global platform that remains answerable to the laws of its home country. Heavy-handed localisation rules can also backfire, because compliance costs fall hardest on smaller providers and can leave buyers more reliant on the same few incumbents the rules were meant to reduce.

The test of any sovereignty claim is whether it reduces your dependence on parties you cannot see and widens your options if you need to leave. Applying it starts with knowing whether a vendor runs its own infrastructure or resells someone else’s, which model reasons over your data, and who operates it. Leaders should also map what would stop working if an outside provider withdrew access, check that their data and workflows can move elsewhere, and look at how the vendor makes money, since advertising or data licensing can pull against a customer’s interests.

Why owning the stack matters

Checks should favour vendors that control more of what sits beneath their software. A provider that runs its own data centres, rather than renting capacity from a hyperscaler, takes a whole layer of third-party legal obligations out of the chain. One that builds its own AI models and runs them on its own infrastructure lets customers apply AI to their records without passing them to an outside lab, or depending on that lab’s terms and the rules of its home country. Business models matter too. A vendor that doesn’t rely on advertising or outside investors has fewer incentives that pull against its customers.

For a small business, choosing a vendor like that means getting the benefits of AI without inheriting a chain of unseen suppliers. For an enterprise, it means fewer outside providers to map, fewer points where a decision made elsewhere can interrupt operations, and a clearer answer when the board or the regulator asks who controls the data.

No vendor is beyond the reach of every government or insulated from a global chip market that no single company controls, and anyone promising total sovereignty is overselling. What every organisation can control is how many outside parties stand between its data and the tools that depend on it. Australian businesses should start counting them.

Image credit: iStock.com/chiewr