At a glance
Andrew Charlton, the Assistant Minister for Science, Technology and the Digital Economy, told the ANU on 18 August 2026 that Australia should capture value across 80% of the AI stack rather than be confined to 20%.
Charlton says about ten cents of every dollar spent on AI buys the data centre itself, and a few cents more buys the electricity.
Expectation 5 of the national data centre expectations asks providers of large-scale compute, naming hyperscalers and neoclouds, to open compute to Australian start-ups, innovative small businesses, researchers and not-for-profits on favourable terms.
The Australian government will prioritise proposals most closely aligned with the expectations, the expectations document says.
Charlton says Australian AI spending could reach A$20 billion to A$40 billion a year within a decade, more than Australia earns exporting wheat.
Charlton says ten cents of the AI dollar buys the data centre
Andrew Charlton, the Assistant Minister for Science, Technology and the Digital Economy, delivered the annual signature lecture at the Australian National University's Crawford School of Public Policy in Canberra on 18 August 2026. The Crawford School billed it as Australia's AI growth opportunity. The speech text, released to media on 17 August 2026 under the title Why Australia Needs to Turn Compute into Sovereign AI, was reported by The Australian Financial Review, The Conversation and InnovationAus.
Charlton placed Australia on the bottom two rungs, the buildings and the power. Above them sit the chips and servers inside an AI data centre, the models trained on them, and the software and intellectual property sold on top. He called data centres a step ladder up that stack, on InnovationAus's account. "Only a few cents of the AI dollar buys electricity," he said, according to the speech text. "Perhaps another ten cents buys the physical data centre."
The two headline numbers measure different things. The cents are his split of what a dollar of AI spending buys. The 80% is his measure of how much of the AI stack Australia should be competing in, against the 20% he says it is currently confined to.
Australian AI spending is rising by about 20% to 40% a quarter on Charlton's figures, and he said it could plausibly reach A$20 billion to A$40 billion a year within a decade. At that point it would rank among Australia's largest imports. He told the audience Australia could easily spend more importing artificial intelligence than it currently earns exporting wheat, and is on course to be "a large and permanent importer of intelligence" unless it acts in time.
Charlton named the instrument for closing that gap as the fifth of the Australian government's five national data centre expectations, and described the support he is seeking from AI companies for Australian start-ups, university research partnerships and AI talent as "the express thinking" behind it.
Cost of capability has collapsed
"The cost of capability has collapsed," Charlton says in the speech text. "The strongest open models now sit within distance of the frontier at roughly a third of the price." He names DeepSeek, Kimi and GLM, three open-weight models from Chinese developers whose trained parameters anyone can download and run.
Charlton argued that a mid-sized organisation could now take a strong open model and, through fine-tuning, distillation and retrieval on its own data, make it better than a frontier system at a narrow, valuable task. Cheaper models still have to be trained, tuned and served on machines, and access to those machines, on his account, is emerging as the binding constraint on what Australian researchers can attempt and what Australian founders can build.
Expectation 5 is addressed to providers of large-scale compute
The Expectations of data centres and AI infrastructure developers were published on 23 March 2026. They cover national interest, the energy transition, water, skills and workforce, and research, innovation and local capability. The first four are addressed to data centres, AI infrastructure and data centre operators, which is the set Australian operators are already measured against.
Expectation 5 is the only one of the five addressed to providers of large-scale compute. It says that "providers of large-scale compute, including hyperscalers and neoclouds, are expected to contribute to research and innovation, including by enabling access to compute for Australian start-ups, innovative small businesses, researchers and not-for-profits on favourable terms". It also asks those providers to deploy engineers and researchers in Australia as part of any major investment, build local technical and innovation capability, and invest in Australian supply chains.
The document applies to new or expanded developments within Australia. It asks colocation operators, hyperscale operations and the large-scale AI compute centres it calls AI factories to consider how the expectations apply in their contexts. Small-scale edge and on-site enterprise facilities sit outside its scope.
In a colocation hall the tenant provides the compute
Australia's colocation operators, tracked site by site in the directory of Australian data centres, sell space, power and cooling. The computers racked inside belong to the tenant, who decides who runs on them.
Expectation 5 names hyperscalers and neocloud providers, among them Firmus, Sharon AI and IREN. Those neoclouds buy GPU fleets and sell time on them by the hour.
The Australian Financial Review, reporting the speech on 17 August 2026, said the government would "force" AI companies including Microsoft, Google and Anthropic to make data centre capacity available to local start-ups and innovators.
Aligned proposals get priority
A section headed how the expectations work says the government "will prioritise proposals most closely aligned with the expectations", and that "energy-intensive data centre proposals not closely aligned with the expectations will not be prioritised by Commonwealth regulatory assessments". The same section commits the government to work with states, territories and market participants to implement the expectations in their processes, particularly through the Energy and Climate Change Ministerial Council, the forum where Commonwealth, state and territory energy and climate ministers meet.
The expectations work alongside existing national, state and territory laws, and compliance with Australian law continues to apply in full.
Instrument | Date | What it covers |
|---|---|---|
Expectations of data centres and AI infrastructure developers | 23 March 2026 | National interest, energy, water, skills and workforce, and research, innovation and local capability |
Obligations announced by Prime Minister Anthony Albanese, not yet law | 15 July 2026 | New power supply, full grid connection costs and water infrastructure, expected to be legislated in early 2027 |
Energy and Climate Change Ministerial Council communique | 28 July 2026 | Ministers, Queensland and the Northern Territory excepted, agreed to progress regulatory arrangements to mandate in-jurisdiction renewable offsets with a jurisdictional opt-out. The communique also describes AI Standards covering energy, water and location, and records that Queensland dissented from it as a whole |
Source: Department of Industry, Science and Resources; the Prime Minister's 15 July 2026 announcement; the Energy and Climate Change Ministerial Council communique of 28 July 2026.
The 15 July package also promised faster data centre approvals and an Office of AI inside the Department of the Prime Minister and Cabinet.
Chips, model training and the software layer
Chips are designed and manufactured overseas. Models are trained on large clusters of those chips, and the companies training the frontier ones are headquartered elsewhere. Software and intellectual property are built on top, and that layer earns licence revenue again each year from code already written.
Compute access reaches the model layer and the software layer, because a team with machines can train, tune and sell what a model does. The chip layer stays offshore, in fabrication plants that take years and a different industrial base to build.
Australia's neoclouds already sell compute by the hour, which is the product Expectation 5 asks providers to open up. What that product costs an Australian start-up is the part the expectation leaves to the provider.
Canada's programme funds compute access directly. Its Sovereign AI Compute Strategy, announced in Budget 2024, commits C$2 billion over five years, including up to C$300 million in an AI Compute Access Fund that helps Canadian businesses and innovators buy compute, and up to C$1 billion for public supercomputing. Australia's instrument is Expectation 5, with the priority described above behind it.
What to watch
The research compute question. The Australian Financial Review reported that Charlton brought together research universities, start-up accelerators and the CSIRO, the national science agency, earlier in August 2026, and that several Australian developers of foundation AI systems had warned the government it was not doing enough to help local start-ups get access to computing power.
The reach of the prioritisation. The prioritisation sentence in the expectations is written about energy-intensive data centre proposals; Expectation 5 is written about providers of large-scale compute, who may not be the same applicants. How the two meet is a question for the implementation work the document routes through the Energy and Climate Change Ministerial Council.
The legislation to come. The obligations Albanese announced on 15 July 2026 are expected to be legislated in early 2027, and the Australian Financial Review reported the Australian government is working with the states on how to implement them.
The AI Standards. Energy and climate ministers other than Queensland committed on 28 July 2026 to support consideration of the AI Standards by National Cabinet in August 2026. The communique describes those standards as covering energy, water and location, and does not itself schedule the National Cabinet item.
The definition of favourable terms. Expectation 5 uses the phrase without defining it, so the first published example of what a hyperscaler or neocloud offers an Australian start-up becomes the reference point for the ones that follow.