An AI deal with Washington could improve Australia’s access to frontier models. But Australia should judge it by the domestic capabilities, operational control and freedom of choice it gains in return.
“We must do a deal with the United States.”
That was Shadow Foreign Minister Ted O’Brien’s answer to Australia’s AI challenge at an open dialogue hosted by the Australia–China Relations Institute at the University of Technology Sydney on October 7.
O’Brien says neither Beijing nor Washington should define Australian interests. This message was welcomed by the audience. But his proposed AI partnership deserves closer scrutiny.
His case implies that a US deal would demonstrate more agency than regulatory or multilateral initiatives. His criticism of Anthony Albanese for returning from New York without an AI framework makes securing a deal the test of that agency.
O’Brien would exchange Australia’s energy, land, infrastructure and strategic trust for preferential access to American technology, then build sovereign capability, following his stated approach: “think big, start small, scale fast.” Shadow Defence Minister James Paterson has echoed the argument, suggesting that attracting frontier-model training investment would effectively guarantee access. Their statements do not establish settled Coalition policy. But their public case gives strategic alignment priority while leaving commercial terms and alternatives largely unspecified.
The test of any deal is whether it builds sovereign capability. For Australia, AI sovereignty means having the authority and domestic capability to choose, evaluate, adapt and govern AI systems in its interests. It includes control over sensitive data, and the ability to provide essential services when overseas access is disrupted. It also includes Australian firms’ ability to choose and adopt the models that best meet their business needs and enhance productivity.
This reasoning seems to assume that a deal with the US would give Australia the sovereign AI capabilities it needs to enhance both AI safety and productivity. Yet several obstacles stand in the way of “scaling fast”.
First, a government-to-government framework proposed by O’Brien could improve access. The models are, however, mostly owned by companies, which set the terms of access.
Second, access to proprietary frontier models doesn’t necessarily include the right to run them independently or modify them. Developers can retain the model weights, intellectual property and control over deployment decisions.
Third, Australia would supply the electricity and infrastructure. Yet hosting investment should be judged by the capabilities it creates, rather than its location.
To make such a deal work, Canberra would need enforceable terms covering costs, data use, adaptation rights and continuity if access were withdrawn. Washington’s assurances would also need to withstand changes in American policy. The Australian public deserves to know what their resources are buying and whether Australia would retain a share of the knowledge and profit generated by such a deal.
O’Brien argued that an AI “shield” secured through the deal could protect Australia against AI-enabled cyberattacks. However, reliance on foreign providers can introduce vulnerabilities of its own.
Australia risks a widening AI remediation gap as the US–China AI race accelerates: offensive capability increasingly operates at machine speed, while much of the world still responds at the slower pace of human decision-making and institutional coordination across multiple stakeholders.
A bilateral deal could help narrow this gap. But it cannot cover every source of cross-border exposure. Australia needs international arrangements for early warnings, confidential technical exchanges and remediation assistance, including cooperation with the United States and China.
As frontier-models pre-training grows more expensive, building sovereign AI capabilities need not entail training a frontier model from scratch. Most firms can start with an existing model. Post-training adapts pre-trained models to particular tasks. For most middle-power countries, including Australia, adaptation and evaluation offer a more accessible route to capability.
Open-weight models offer one route to such adaptation. Yet O’Brien didn’t mention them, which left many in the audience puzzled. Open-weight models carry risks, but also offer greater autonomy: their downloadable parameters permit local operation and adaptation, subject to licensing and expertise.
They are a strategic choice for many Chinese firms and allow smaller firms to develop AI capabilities quickly. Ali Ghodsi, Databricks’ chief executive, argues that firms can take an existing open-weight model and use reinforcement learning to improve performance on their business tasks. Once businesses use AI at scale, the emphasis shifts from encouraging adoption and maximising token consumption to improving returns per dollar. As Ghodsi suggests, businesses need a dynamic mix of models: the most capable models for difficult tasks with high returns, and less expensive models for the vast volume of routine work.
This is not an argument for either adopting or banning Chinese open-weight models. But a deal with the US could cut off this option for Australian businesses if it restricted which models they may use. Defence, critical infrastructure and government functions require stringent security assessments, and strategic alignment may constrain the choice of providers. For non-critical businesses, clear rules on data use and security should preserve opportunities to learn from open-weight models, regardless of origin.
Securing a deal would only begin the work. Its terms would determine whether Australia strengthens its sovereignty or deepens its dependence.
Dr Marina Yue Zhang is an Associate Professor at the Australia–China Relations Institute, University of Technology Sydney. Her research focuses on technology geopolitics, China’s innovation system, AI, semiconductors and strategic supply chains. She is co-author of Demystifying China’s Innovation Machine: Chaotic Order (Oxford University Press, 2022) and writes regularly on the intersection of technological change, economic security and US-China strategic competition.
This article is published under a Creative Commons Licence and may be republished with attribution.
Photo: “P20251020DT-0362” by The White House, United States Government Work (CC0)
