The phrase “AI sovereignty” has been bouncing around Canberra policy circles for the past year. It sounds good in speeches. But when I ask people what it actually means, the answers range from vague to contradictory.

So let me have a go at defining it and arguing why it matters.

What AI Sovereignty Means

AI sovereignty for Australia means three things.

Compute sovereignty. Australia needs domestic AI compute infrastructure so that our most sensitive AI workloads don’t depend on infrastructure controlled by foreign governments or companies. This doesn’t mean every AI model needs to run on Australian servers. It means we have the capacity to run critical workloads domestically when we need to.

Model sovereignty. Australia needs the capability to develop and maintain AI models trained on Australian data, reflecting Australian values and requirements. Relying entirely on American and Chinese foundation models for critical applications creates dependency risks.

Data sovereignty. Australian data used for AI purposes should remain under Australian jurisdiction and legal frameworks. When Australian health data trains an AI model on servers in Virginia, we’ve lost a degree of control that matters.

Why It Matters

The practical argument is straightforward. Every critical dependency on foreign AI infrastructure is a risk.

If the US and China enter a technology conflict that restricts AI access, Australian businesses and government agencies running on American AI platforms face disruption. That’s not hypothetical. We’ve seen technology access used as a geopolitical tool repeatedly.

If a major AI platform changes its terms of service, pricing, or data handling practices, every Australian organisation using that platform has to accept the changes or migrate. Migration from cloud AI platforms is expensive and time-consuming. The power is entirely with the platform provider.

If Australian data leaves Australian jurisdiction for AI processing, our ability to regulate how that data is used diminishes. Data protection laws have limited effect when the data sits on servers in another country operated by companies under another country’s legal system.

And there’s the economic argument. AI is the most important technology of this generation. Countries that develop domestic AI capability capture the economic value of AI development. Countries that only consume imported AI services pay for capability without building it.

What Australia Actually Has

Let’s be honest about our starting position. It’s better than the pessimists suggest but worse than the optimists claim.

On compute, Australia has limited AI-specific infrastructure. We have data centres, but purpose-built AI compute at the scale needed for training large models is sparse. The government has announced investment in this area, and several private sector initiatives are underway, but we’re years behind the US and China.

On models, Australian institutions have developed good AI models for specific applications. CSIRO’s Data61 does excellent applied AI research. Australian universities produce world-class AI researchers (who then mostly leave for better-funded overseas positions). What we lack is the ability to train large foundation models. That requires compute infrastructure we don’t have.

On data, Australian organisations generate vast amounts of valuable data. We have extensive government datasets, strong research data infrastructure, and increasing amounts of industry data. But much of this data is fragmented, poorly catalogued, and underused for AI development.

What We Should Do

Invest in domestic AI compute. The announced government investment in AI compute is a start but needs to be accelerated. Australia should have at least one world-class AI compute facility capable of training large models. This is infrastructure investment, not a research grant. It should be treated with the same strategic importance as defence infrastructure.

Build Australian foundation models. We should develop foundation models trained on Australian data for Australian applications. Not to compete with GPT or Claude on general capability, but to have models that understand Australian English, Australian institutions, Australian law, and Australian cultural context. These don’t need to be the biggest models. They need to be the most appropriate for Australian use cases.

Create a national AI data infrastructure. Make Australian datasets usable for AI development while protecting privacy. This means standardised data formats, cataloguing, access frameworks, and privacy-preserving techniques like the federated learning approaches CSIRO is developing.

Retain AI talent. The brain drain of Australian AI researchers to overseas positions is a genuine problem. We can’t match Silicon Valley salaries, but we can offer world-class research infrastructure, interesting problems, quality of life, and career paths that don’t require relocation. The government’s proposed AI research fellowships are a good start.

Regulate with sovereignty in mind. AI regulation should include provisions for domestic capability. Government AI procurement should preference domestically developed solutions where they meet requirements. Critical infrastructure AI should be required to run on Australian infrastructure.

The Cost of Inaction

If Australia doesn’t invest in AI sovereignty, we become a technology consumer entirely dependent on foreign providers. Our most sensitive data processes through foreign infrastructure. Our economic activity generates value captured overseas. Our national security depends on continued access to technology controlled by others.

That’s not a position any serious country should accept for a technology this important.

The investment required is significant but proportionate to the stakes. The National Reconstruction Fund already has manufacturing and technology mandates. The defence budget already includes technology investment. Redirecting a fraction of these existing allocations toward AI sovereignty would make a meaningful difference.

Australia has the research talent, the data, and the economic scale to build meaningful AI sovereignty. What we need is the strategic decision to do it and the discipline to follow through. Every year of delay makes the gap harder to close.