Australia has been offered a pragmatic new definition of AI sovereignty. A recent report from the United States Studies Centre argues that, because Australia cannot reproduce America’s frontier AI stack, it should pursue “sovereignty through managed dependence”. That is, it should pursue privileged access to US models and infrastructure, supported by alliance and commercial commitments.
The report is right about two important things. Australia cannot simply buy its way to frontier AI self-sufficiency. Nor should it assume that data centres located here are sovereign when overseas companies control the models, cloud platforms, chips and operational decisions.
But managed dependence addresses only part of the problem.
Better and more reliable access is worth securing, but access remains permission. Sovereignty requires the capacity to keep operating when the terms of that permission change.
For Australia, the realistic objective is not control over every layer of the AI stack. It is sovereign operational capability: sufficient control over compute, models, data, skills and deployment to preserve national choice for critical functions.
Frontier access can be delayed, restricted, repriced or withdrawn. Australia should not assume that a close alliance guarantees timely or equivalent access to the most capable systems. Treaties, government arrangements and commercial contracts can reduce disruption risks, but none can remove the United States’ authority to change its export controls or national-security priorities.
Managed dependence is therefore an important strategy for frontier access, but it is not a complete account of sovereignty.
The report frames Australia’s choice too narrowly. Australia does not have to reproduce the American frontier AI stack to build capabilities that reduce the consequences of dependence.
Sovereign AI does not require autarky. It requires control over enough of the stack to protect sensitive data, sustain critical services, adapt and evaluate systems, allocate compute to national priorities and move between providers without losing the capacity to operate.
Australia imports critical components of its defence, energy and telecommunications systems. The presence of foreign technology does not make domestic control meaningless. The objective is not to eliminate every dependency, but to avoid those that are so concentrated and difficult to reverse that another government or company can determine whether an essential Australian capability continues.
Not all dependencies create the same risks. A foreign-designed chip installed in Australia may still depend on overseas firmware, software, licences, maintenance and future supply. But it can provide materially greater operational control than a remotely delivered service whose availability may change immediately.
Chip dependence affects future supply and refresh cycles. Dependence on an overseas API can interrupt present operations. Neither provides complete autonomy, but they are not strategically equivalent.
The report’s cost comparison also warrants scrutiny. It considers the Australian Academy of Technological Sciences and Engineering’s proposal for between $4.5 billion and $5 billion over five years, before invoking an estimate that sovereign frontier AI could cost US$500 billion.
These figures price fundamentally different objectives.
ATSE proposes publicly supported AI factories, national compute, secure inference, datasets, domain-specific models, workforce development and access for government, researchers, startups and small businesses. It is not a proposal to reproduce the scale and full capabilities of OpenAI, Anthropic or Google DeepMind.
The US$500 billion figure comes from an exploratory estimate for a four-year multinational project intended to create a frontier laboratory capable of competing with the leading American developers. The problem is not necessarily that the estimate is wrong. It is that it prices a different undertaking.
This does not establish that ATSE’s design or price tag is correct. It means the case against it requires a genuine assessment of the capability it would create, rather than comparison with the cost of pursuing a radically more ambitious objective.
That assessment should be demanding. Public compute can be poorly procured, underutilised or quickly overtaken. Domestic programs can become expensive infrastructure projects disconnected from real workloads. Nominally sovereign systems may remain dependent on foreign software and expertise.
Any Australian investment should therefore be workload-led, contestable and regularly tested for technical relevance, utilisation and public value.
The report’s claim that sub-frontier capability “answers no strategic need” is less convincing.
Some national-security tasks may require immediate access to the most capable system available. That supports negotiating preferential frontier access. It does not follow that every strategically important task is a contest for the highest general-purpose benchmark score.
In healthcare, agriculture, public administration, defence and critical infrastructure, the most valuable system may be the one that is secure, auditable, locally adaptable and continuously available. A domestically controlled model need not be the world’s best general-purpose model to support a specialised Australian workload or provide a credible fallback.
In May 2026, Epoch AI estimated that, on its composite Epoch Capabilities Index, the most capable open-weight models had lagged frontier closed models by an average of approximately four months since January 2026. The measure aggregates performance across multiple benchmarks and should not be read as a uniform four-month gap across every task. But a time-limited capability lag is not equivalent to strategic uselessness.
For some workloads, frontier performance will be essential. For others, task performance, security, affordability and assured availability will matter more. A serious national strategy must distinguish between them.
Comparable countries are already pursuing blended approaches.
The United Kingdom has committed up to £2 billion through 2030 to develop a modern public compute ecosystem combining purpose-built supercomputers with cloud capacity and emphasising supplier diversity, security and portability. Canada has launched a C$2 billion five-year strategy spanning public computing infrastructure, commercial AI data-centre projects and compute access for Canadian businesses.
Neither country is pursuing autarky — both recognise that domestic capability can complement international access.
Australia should adopt the same portfolio logic.
The first track should be managed access: agreements with the United States, frontier laboratories and hyperscalers covering timely model releases, reserved domestic capacity, secure deployment, continuity, pricing and transition arrangements.
The second should be selective sovereign capability: Australian-controlled public compute, secure inference for critical workloads, locally held open-weight and licensed models, domain-specific systems, independent assurance, technical skills, portability and tested exit plans.
This capability is also necessary to strengthen Australia’s bargaining position. Australia can offer land, energy, grid infrastructure, critical minerals and favourable conditions for data-centre investment. But hosting infrastructure does not guarantee strategic control or economic value.
Unless these arrangements secure Australian compute access, local research and industry participation, workforce development and critical-service continuity, Australia could supply the land, energy and public infrastructure while intellectual property, pricing power and high-value activity remain offshore.
Australia would become an AI tenant in its own economy.
The practical test is straightforward. If foreign frontier access were interrupted, could Australia continue its most important services, protect sensitive data, operate fit-for-purpose models and shift priority workloads while alternatives were established?
If not, dependence may be better managed, but sovereign capability has not been secured. Australia needs both American frontier access and Australian-controlled capability. Access can be negotiated, but agency has to be built.
Dr Alex Antic is the faculty head of AI Strategy at UNSW Canberra, and deputy director of the UNSW AI Institute
Do you know more? Contact James Riley via Email.