Sovereign AI What It Means for Australian Organisations

It depends on what you mean by sovereignty. Storing data in an Australian cloud region means the data at rest is subject to Australian jurisdiction and law enforcement processes. However, if you are processing that data through an AI model API where inference runs on US-based infrastructure (even if the AWS region is in Sydney), your prompts and context are being transmitted to and processed on US infrastructure. The model inference step is where most AI data sovereignty risk resides. For genuine sovereignty, the model inference must also occur on Australian infrastructure, not just the data storage.

The Privacy and Other Legislation Amendment Act 2024 did not introduce AI-specific provisions, but several of its changes directly affect how organisations must handle AI processing of personal information. The increased civil penalties (up to $50 million or three times the benefit obtained) apply to serious or repeated interferences with privacy, which could include systematic use of personal information in AI systems without appropriate authority. The new requirement for automated decision-making notifications where AI significantly affects individuals creates additional obligations for AI systems that produce consequential decisions.

Microsoft 365 Copilot processes data through Microsoft's Azure OpenAI Service, which means AI inference occurs on Microsoft infrastructure. Microsoft's data residency commitments are that data in Microsoft 365 is stored in Australia for Australian tenants with data residency commitments, but the AI inference processing for Copilot may involve different infrastructure depending on the service tier. For Australian organisations with APP 8 obligations, the question is whether Microsoft's terms constitute an adequate assurance that the overseas recipient will not breach the APPs. This is a legal question specific to your organisation's risk assessment and may require legal advice.

As of 2025, there is no single government-endorsed list of sovereign AI providers. The closest equivalent for government agencies is the DTA's Hosting Certification Framework and the ASD's Certified Cloud Services List for government-grade cloud infrastructure. AI providers deploying on certified infrastructure are using a sovereign foundation, but certification of the infrastructure does not automatically certify the AI system built on top of it. For organisations requiring assured sovereign AI, we recommend a combination of using HCF-certified infrastructure and obtaining a written data processing agreement that explicitly commits to Australian-only data processing.

AUKUS creates specific technology transfer and information sharing obligations that are still being defined as of 2025. What is known is that technology related to the nuclear submarine program and certain advanced capabilities areas must be handled with the same rigour as classified defence information. For defence companies engaged in AUKUS work, the principle is that any data related to advanced capabilities must be handled on infrastructure that cannot be compelled to disclose to a foreign government without an Australian authorisation process. A private LLM deployed on air-gapped, on-premises hardware is the architecture most clearly consistent with these emerging requirements.

When evaluating sovereign AI vendor claims, ask for: written documentation of where model inference occurs (not just data storage), written commitment that your prompts and context are never transmitted to overseas AI providers, details of the underlying model and whether any API calls are made during inference, the physical location of all infrastructure used in the inference pipeline, and whether the vendor will accept an audit right to verify these claims. Vague assertions about "Australian data sovereignty" without specific technical documentation should be treated with scepticism.