A three-week US export block on foreign access to Anthropic’s most advanced models has reopened one of the hardest questions in Australian technology policy: should Canberra try to build a frontier AI model of its own?
Writing in The Conversation, Olivia Shen, Director of Strategic Technologies at the United States Studies Centre at the University of Sydney, argues the answer is no. Her case is blunt: the money does not add up, the odds of catching the leaders are poor, and Australia has better places to spend its limited advantage.
It is an unusually pointed intervention in a debate that has, until now, run mostly on anxiety.
What triggered the debate
The catalyst was a scare in June. As an earlier analysis in The Conversation set out, the US Department of Commerce ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models for foreign nationals, inside and outside the United States.
Anthropic shut off access within hours. Australian researchers and businesses that had wired those models into their products were, briefly, locked out.
Commerce lifted the controls less than three weeks later, restoring access first to a handful of trusted US organisations before widening it again. But the episode landed hard. It showed that access to the tools now underpinning parts of the Australian economy can be switched off in another capital, with little warning and no appeal.
That is the fear Shen is responding to. Her argument is not that the risk is imaginary. It is that building a sovereign frontier model is the wrong answer to a real problem.
Why the maths points away from a frontier model
Shen’s first pillar is cost. She notes the price of training a leading-edge model is forecast to exceed US$1 billion by 2027, a figure that puts genuine frontier development beyond the reach of a mid-sized economy competing on its own.
Her second pillar is track record. France, the United Arab Emirates, Japan, South Korea, India and Switzerland have all poured money into sovereign AI ambitions. None, she writes, has yet produced a model matching the best American or Chinese systems.
The lesson she draws is uncomfortable but hard to dismiss. Spending is not the same as competing, and a billion-dollar bet placed a step behind the frontier can be obsolete before it ships.
Her alternative is to play to Australian strengths: critical minerals, data-centre capacity, and the specialised, high-quality datasets that local institutions already hold. Rather than build a rival model, she argues, Australia should secure reliable access to the best ones, push for trusted-ally status in US licensing arrangements as the United Kingdom has done, and use its new AI Safety Institute to earn a seat at the table on evaluation and safeguards.
The counter-case Canberra is already making
Not everyone in government frames the problem the way Shen does. The politics of AI sovereignty have grown sharper this year.
Industry Minister Tim Ayres has warned that Australia cannot afford to be a cork bobbing on the ocean of other people’s technology, a line widely reported as the government sharpened its stance. Coalition figure Andrew Hastie has framed AI access as a strategic contest, and Defence Minister Richard Marles has called for Australia to build some agency of its own in the field.
That instinct is not baseless. The earlier Conversation analysis breaks sovereignty into four parts: data sovereignty, meaning data stored here and subject to Australian law; compute sovereignty, meaning in-country data centres under Australian control; model sovereignty, meaning capability not dependent on a foreign provider; and policy sovereignty, meaning the ability to set our own rules rather than inherit another country’s export controls.
The June block hit the last two directly. It is possible to accept Shen’s economics on model-building while still worrying that outsourcing the model layer entirely leaves Australia exposed the next time Washington reaches for the switch.
Our analysis
FluentSea’s view: Shen has the stronger argument on the narrow question, and the debate has been muddling two different things.
Building a frontier model and building sovereign AI capability are not the same project. A country can decline the first while investing heavily in the second, and that is roughly where the federal government’s own settings already sit.
The National AI Plan, released late last year, leans into infrastructure and adoption rather than model-building. It backs sovereign compute for the public service through a centralised GovAI hosting service, positions Australia as a data-centre hub, and funds the AI Safety Institute with an A$29.9 million commitment. None of that requires a homegrown Fable or Mythos.
The harder judgement is about resilience. A trusted-ally licensing deal is only as durable as the administration that grants it, and the June block was a reminder that whitelists can be redrawn. That argues for a hedge: local compute, portable systems that can move between model providers, and open-weight options for the workloads that must never go dark.
The trap Shen rightly names is the vanity project, a billion-dollar model built for prestige that arrives a generation behind. Australia should skip that. The quieter work of controlling its data, its compute and its rules is where sovereignty is actually won.
Why it matters for Australia
The stakes are practical. Australian firms, universities and agencies have already embedded foreign frontier models into products, research and services, which is exactly why the June outage stung.
For local industry, the payoff sits in the layer above the model: fine-tuning, safety evaluation, sector-specific applications and the data-centre and critical-minerals base that a global AI build-out needs. Those are jobs and skills Australia can plausibly hold onto.
The risk of getting the framing wrong is real too. A billion dollars chasing a model Australia is unlikely to win is a billion not spent on compute, skills and the safeguards that make dependence on someone else’s model survivable. Sovereignty, on this reading, is less about owning the model and more about never being caught flat-footed when access to it disappears.
Sources: The Conversation — Australia shouldn’t try to build its own frontier AI; The Conversation — The US government can shut off access to AI at will; Startup Daily — Why building sovereign AI in Australia doesn’t make sense; Department of Industry, Science and Resources — National AI Plan; Minister Tim Ayres — National AI Plan media release.
