Turning AI Infrastructure into Benefits for Australia

Background

On 17 September 2026, the Office of AI, in the Department of the Prime Minister and Cabinet, released the AI Standards Consultation Paper, Getting it right: Building AI infrastructure that works for Australia, to seek views on the conditions Australia should set for AI infrastructure. Part 2 of the paper covers standards for large data centres, including power, water and community impacts. Part 3 covers conditions for frontier AI companies that train AI in Australia. Submissions closed on 9 October 2026.

Our submission

The submission, Turning AI Infrastructure into Benefits for Australia, argues that Australia should host frontier AI training with strings attached. Few countries will host training, and Australia has energy, space and stability that few others can match. But Australia shouldn't overplay a strong hand. Frontier AI companies have genuine red lines and will weigh the overall Australian operating picture against their best alternatives. Government should aim to gain the most value it can without pushing those companies to train elsewhere.

The consultation paper gets the foundations right. Resolving the questions in Part 2 is essential to securing social licence and ensuring AI data centres are consistent with Australian wellbeing. Our submission to the Senate Inquiry into AI and Data Centres recommended mandatory standards tackling genuine issues around power, water and community impacts. This submission responds primarily to Part 3.

The submission recommends Government prioritise commitments that cannot be purchased on the free market, and pursue fees and taxes to fund other interventions. It discusses four areas:

1. Access to frontier models. Ongoing access to frontier AI is not assured, and may be the most important term Australia negotiates. Frontier AI companies training here should guarantee Australia the same model access as the US at three levels: the public, critical infrastructure operators, and national security agencies. Access needs to be secured through both frontier AI companies and the US Government. Linking Australian model access to permission to train in Australia gives US companies a reason to encourage the US Government to preserve Australian access.

2. Australian skills and compute. Hosting frontier AI should build Australian skills and compute capacity, with obligations that scale with the amount of compute a company hosts here. This could include:

• Requiring frontier AI companies training here to keep a research and engineering team in Australia, in proportion to their share of training compute in Australia.
• Including proportionate skills commitments in mandatory standards for data centres during construction and operation, such as apprenticeship targets and TAFE partnerships.
• Requiring frontier AI companies to fund 'AI apprenticeships' that place early-career Australians in not-for-profits, public research organisations and government agencies.
• Requiring large compute operators to offer a share of their Australian capacity to Australian buyers first, at market rates.
• Requiring large compute operators to donate a portion of their Australian capacity to not-for-profits, researchers, start-ups and Australian AI Safety Institute evaluations.

3. Transparency and incident reporting. Australian regulators only know what is public. Australia should impose robust transparency and incident reporting obligations on AI companies that carry on business in Australia, whether or not they train here.

• "Frontier systems" should be defined through a compute threshold, capabilities, and an explicit designation power.
• Incident reporting should set lower thresholds than equivalent laws in the EU and California, and cover internal use, near misses, misaligned behaviours, and theft or modification of model weights.
• Transparency reporting should cover internal models, detailed AI research and development automation metrics, and full biosecurity evaluations.
• Frontier AI companies should be required to share reports they give to overseas regulators.

4. Risk management. Hosting training brings frontier AI companies within reach of Australia's crisis management and critical infrastructure laws, which need updating to ensure Australia benefits.

• The Australian Government Crisis Management Framework should treat AI incidents as an explicit hazard, with a designated lead agency and a dedicated AI crisis plan.
• The Security of Critical Infrastructure Act should capture entities that train AI, not only data centre operators, and reach AI providers that other critical infrastructure depends on.
• Frontier AI companies training here should work with the Australian AI Safety Institute to build compute verification tools and, in the long term, be required to use them in their Australian facilities.
• Australia should require embedded third-party safety evaluators in frontier AI companies, a measure Anthropic and OpenAI have already called for.

The submission concludes that Government should give the most weight to conditions Australia cannot obtain any other way. Frontier model access is the most important. As the most capable models become restricted, linking access to permission to train may be the only reliable way for Australia to secure it.