Australia has been handed a warning from Washington: access to frontier AI is not a product guarantee. It is a geopolitical privilege. The big question is what will Canberra do about it?
Anthropic has abruptly disabled its most advanced Claude Fable 5 and Mythos 5 models after a US Government export-control directive ordered the company to suspend access for foreign nationals, including people outside the US and foreign nationals inside the US. Anthropic said the order cited national security concerns and fears that safeguards could be bypassed for cybersecurity vulnerability discovery.
Anthropic argues it has a vast range of guardrails built-in that will prevent harmful usage, but nevertheless, the law is the law, and right now Australians (along with billions of others worldwide) can not use Anthropic’s advanced models.
Implicitly, Australian companies, universities, public agencies, and defence-adjacent users assumed they could simply rent the world’s best intelligence from US vendors forever. After all, we can use many, many SaaS apps, many hypervisor services, and so on. It was previously unthinkable Australia would effectively be on a ban list like those user agreements of years ago that read “you certify you are not a citizen or resident of Iran, North Korea, Russia …” – well, those may as well include Australia now.
The USA may be an ally, and may well be a trusted trading partner. But they are not a sovereign supplier.
Europe has understood this for some time. The EU is pushing AI gigafactories through EuroHPC to build large-scale compute for training, fine-tuning and deploying advanced AI systems, explicitly because compute shortages threaten Europe’s strategic autonomy. France’s OVHcloud is now planning frontier AI models to become another European alternative to Mistral, with CEO Octave Klaba arguing Europe must master AI technology itself.
The mood shift is broader than AI models. Europe is also rethinking dependence on US hyperscalers for defence and critical workloads, with warnings that cloud reliance can create a de facto “kill switch” risk. In other words, sovereignty is no longer just about where data sits. It is about who controls the model, the cloud, the management plane, the chips, the energy supply and the legal jurisdiction.
Australia has started moving, but cautiously. The Federal Government’s National AI Plan sets out ambitions around opportunity, shared benefits and safety, while the APS AI Plan focuses on trust, people and tools for public-sector adoption. More concretely, Canberra has issued expectations for data centre and AI infrastructure developers, saying projects should serve Australia’s national interest, support the energy transition, use water responsibly, create local skills and jobs, and strengthen research and local capability.
That is a useful start. It is not yet a sovereign AI strategy.
Australia does not need to build a ChatGPT-killer tomorrow. But it does need sovereign AI resilience. That means at least five things.
First, Australia needs domestic AI compute that is available to researchers, startups, critical industries and government, not just hyperscaler tenants. If AI factories consume Australian power and water, some portion of their capacity should support Australian capability.
Second, government procurement should require model portability. Agencies should avoid hard-wiring critical workflows to a single foreign frontier model. Architectures should support Anthropic, OpenAI, Google, Mistral, open-weight models and local models through interchangeable orchestration layers.
Third, Australia should invest in fine-tuned national models for Australian law, health, education, government services, cybersecurity and public-sector knowledge. The point is not always to beat frontier models; it is to own models that cannot be switched off by another country’s export decision.
Fourth, Australia should work with Europe and trusted Indo-Pacific partners on open-weight, auditable and sovereign-deployable AI systems. Europe’s AI gigafactory push and Mistral-style open model strategy offer a useful template.
Fifth, Australia needs enforceable policy, not just guidance. Data centre expectations are welcome, but sovereignty depends on hard requirements around access, auditability, security, resilience, local benefit and continuity of service.
The Anthropic decision may be reversed or softened through a trusted-partner arrangement; G7 leaders are already discussing broader allied access. But the precedent remains. If access to frontier AI can disappear overnight because of another country’s national-security decision, Australia must treat AI as strategic infrastructure.
It’s easy to simply treat this discussion as “America bad”, but the reality is simpler and a lot more serious: every nation acts in its own interest when technology becomes strategic.
Australia should do the same.