ANU Crawford School of Public Policy Annual Lecture

Why Australia needs to turn compute into Sovereign AI I begin by acknowledging the traditional custodians of the land on which we meet today, the Ngunnawal people, and pay my respects to their elders, past and present. My sincere thanks to Professor Sharon Bessel and her team at the ANU for organising today’s event, in this distinguished forum. The Crawford School of Public Policy is one of the Australia’s flagship policy schools and is widely regarded as a leading centre for public policy, economics, development studies, and governance in the Asia-Pacific region.

Importing Intelligence In 1992, the American presidential candidate Ross Perot warned that globalisation would produce a "giant sucking sound" — the noise of factories leaving America for Mexico. Perot lost that election. And globalisation ultimately delivered enormous gains to the world economy.

But he understood something economists sometimes overlook. People care not only whether economic change creates wealth, but who captures it. Many Australians now hear another distant sucking sound.

This time it is not factories leaving our shores. It is artificial intelligence flowing in, and potential value from the economy flowing out. Australians have a long and largely successful national habit.

We dig up the ore and buy back the car. We grow the wool and buy back the suit. We ship the lithium and import the battery.

As an economy, Australia has been superb at the first stage of the value chain - but much of the value created after that has accrued elsewhere. In the resources era, we got away with it. More than got away with it.

Despite all our economic self-deprecation (We're a quarry! We're not complex enough! Foreigners own our economy!), our economy became one of the strongest in the world, and Australians became among the wealthiest people on earth by median national wealth.

Australia made the less fashionable end of the supply chain work for us. We were among the lowest-cost producers of key commodities on earth, and so we captured rents that lifted not just our GDP but our national income, and made the typical Australian one of the wealthiest people in the world. Now the global economy is being reshaped by artificial intelligence.

And the most urgent economic question before us is how Australia remains a first-rank nation in a world shaped by AI. The Prime Minister’s landmark speech three weeks ago laid out a vision for Australian Standards to shape the growth of AI in our national interest. In this speech I will go deeper into one part of that vision: how we ensure that Australia captures the economic benefits of AI and that we maintain our position as one of the world's strongest economies.

First, I want to show that Australia faces a large and growing bill for foreign AI. As AI lifts our productivity and national output, a significant part of that gain may flow back overseas to the foreign owners of the technology on which we increasingly depend. Second, I’ll argue that Australia is playing in the bottom of the stack - in energy and data centres - but most of the real economic value is higher up in the chips, models, software and intellectual property.

And third, I want to show that we can increase our economic dividend from AI by leveraging our strengths to play higher up the value chain. Unless we take timely action, Australia is on course to be a large and permanent importer of intelligence. This might mean that AI grows the Australian economy, but much of the value flows offshore.

Australia’s AI Bill Let's start with the cost of AI to Australians. There is no official statistical series for AI imports. But economically the concept is straightforward.

Every time an Australian household subscribes to ChatGPT, Claude or Gemini, or an Australian business pays an API bill, purchases coding agents or licenses AI-enabled enterprise software, it is importing a service. It does not arrive in a shipping container. It arrives down a fibre optic cable.

It is no less an import because it is invisible. Around fifteen to twenty per cent of Australians now appear to pay for an AI subscription. Business adoption is moving even faster.

Roughly one in three Australian firms now pays for some form of artificial intelligence, and for large organisations — banks, law firms, miners, retailers, insurers, and governments — those costs increasingly run into the millions or even tens of millions of dollars each year. Putting households and businesses together, a reasonable estimate is that Australians are already spending between five and eight billion dollars each year on artificial intelligence, with the overwhelming majority of those payments flowing offshore.

In effect, we are creating a new import bill. An import bill for intelligence. How large might that bill become?

Forecasting is challenging because two forces are moving in opposite directions. The cost of each individual token continues to fall rapidly, but the number of tokens consumed is growing even faster. So far, volume is winning decisively.

The available evidence suggests Australian AI spending is increasing by something like twenty to forty per cent every quarter — effectively doubling every year. This is corroborated by the observation that the revenue of the major AI companies is expanding at similar or faster rates. What was a novelty two years ago is becoming a utility.

For households it increasingly resembles another phone bill. For businesses it is becoming another monthly operating expense. If artificial intelligence becomes embedded across most knowledge work, Australian spending could plausibly reach twenty to forty billion dollars annually within a decade.

At that point it would rank among Australia's largest imports. We could easily spend more importing artificial intelligence than we currently earn exporting wheat. That would represent one of the largest recurring outflows in the Australian economy.

Should Australia Try to Reduce This Dependence on Foreign AI? An economist might reasonably ask why Australia should care about this new import bill. We have always imported technologies — aircraft, machinery, medicines and software — because they make us more productive.

If an Australian business spends one dollar on AI and creates more than one dollar of additional value, Australia is better off. The benefits of adoption are real, and it’s why the National AI Centre is helping Australian businesses adopt AI quickly and widely. But adoption is only half the story.

The other half is ownership. Smart adoption helps our businesses keep up and is vital for competitiveness, but it does not necessarily create a lasting national advantage. As the technology spreads, competition passes much of the productivity gain on to customers through lower prices and better products.

The more durable returns flow to those who own the scarce assets behind AI: the models, chips, platforms, intellectual property and customer relationships. Because these assets are concentrated in a small number of firms, their owners can retain a significant share of the value AI creates. So adoption lifts productivity and improves competitiveness, but ownership determines who captures the profits and economic rents.

AI may lift production and GDP in Australia, while much of the resulting income flows offshore. The goal is not to choose between adoption and ownership. It is to adopt the best technology while retaining a meaningful share of the value it creates.

From a public policy perspective, it matters not just how much Australia spends on AI. It matters where that money ultimately ends up — and who captures the economic rents. There is no perfect accounting of an AI dollar, because the layers of the supply chain overlap and the companies disclose relatively little about their unit economics.

But if you start with a dollar paid by a customer to a frontier AI company, and trace that dollar down through the cost of providing the service, a reasonable rule of thumb is this: Around 5¢ buys electricity. Around 10¢ buys the physical infrastructure of the data centre — the buildings, cooling and power equipment. Around 35–45¢ buys the computing equipment — the GPUs, servers and networking that actually run the models.

The remaining 40–50¢ supports everything further up the stack — the engineers, software, training and development of new models, intellectual property and, ultimately, returns to capital. These are indicative numbers, not an accounting identity. Different models, customers and providers will have very different economics.

And today’s frontier AI companies are investing so heavily that the final category should not be mistaken for profit: in many cases they are spending more developing the next generation of models than they earn from the current one. But the broad picture is striking. Only a few cents of the AI dollar buys electricity.

Perhaps another ten cents buys the physical data centre. Most of the dollar flows further up the stack - to the chips and servers, the models and software, the research and intellectual property. For Australia, that matters.

We can generate the electricity. We can build and host the data centres. But many of the most valuable assets further up the stack are overwhelmingly owned overseas.

That is where the potential economic rents become important and the confronting point for Australia — Simply building data centres in Australia does not necessarily mean that Australia captures the economics of AI. We can supply the land. We can supply the electricity.

We can host the machines. And still find that much of the value — and many of the economic rents — flow offshore. This is not an entirely new problem.

We have seen it before in globalised supply chains, in other software we rely on, and now it’s playing out in a new domain. But AI raises the stakes. Because this time, what we risk importing is not simply a manufactured product.

We risk becoming dependent on imported intelligence itself. How Can Australia Capture More Value from AI? By leveraging data centres to move up the value chain.

That brings us to the obvious question. How can Australia capture more value in the AI value chain? What is the role for government industry policy in achieving that objective?

The answer draws on one of the oldest lessons in industrial policy: leverage your existing strengths in one part of the value chain to build your way into the new parts of the value chain. Australia has strengths, in areas like energy and data centres. We should leverage those advantages to build capabilities higher up the value chain – in models, software and applications.

Australia is already succeeding in attracting data centres on our terms. New investment in equipment and machinery is at the highest level in the history of the series. Industry analysts put the Australian pipeline of data centre investment at well over $150 billion dollars.

Australia is the second most attractive destination for data centre investment in the world, and we will shape this boom carefully to ensure it works for Australia, not the other way around The next step is using the data centre advantage and investment boom to grow our AI training ecosystem, and give us advantage in compute. Why compute? Because compute is becoming as essential to research and innovation as laboratories and libraries once were.

A research team training a diagnostic model on Australian medical imaging needs serious compute to do it. A start-up fine-tuning an open model for mining safety or crop disease needs compute to iterate, test and deploy. Access to compute is emerging as the binding constraint on what our researchers can attempt and what our founders can build.

Without it, the contestable layer of the AI economy is not contestable by us. By growing our compute and AI training ecosystem, we will be able to retain, create and attract world-class research and entrepreneurial talent. That deepening talent and entrepreneurial base, without any hand of government, will go on to deliver economic growth, enhance sovereignty and deliver a return for Australia.

This is where the Australian economy needs to be: in the mix to genuinely capture value across 80 per cent of the AI stack rather than being confined to 20 per cent. And as I outlined - Australia’s approach to data centres and AI training is an integrated part of this broader strategy. That is the express thinking behind Expectation 5 of the government's National Data Centre Expectations – the vital connective tissue that transforms data centres from being a dry utility to a strategic catalyst that unlocks second and third round benefits for Australia.

Expectation 5 asks providers of large scale compute to enable access to compute for Australian start-ups, innovative small businesses, researchers and not-for-profits, on favourable terms. Australia is building a booming data centre industry. As the Prime Minister set out in his address at the University of Sydney last month, we hold a strong hand.

We have land. We have world-class renewable resources and the potential to build far more. We have political and legal stability, strong privacy protections, high-quality connectivity, a skilled workforce, and comparatively low construction costs.

We are a trusted partner, and our tier-one standing under US export controls is a genuine advantage in this region. I want to acknowledge the Prime Minister's leadership here. Establishing an Office of AI inside his own department and legislating a single set of Australian Standards for AI early next year bring questions that most countries are still handling piecemeal — energy, water, siting, national interest, workforce, innovation — into one national framework.

Australia will be among the first countries in the world to do it. It ends the case-by-case approach. It gives investors clarity and it gives communities confidence.

That is the right foundation, and it was built at the right moment: before the concrete is poured, not after. I also want to acknowledge Minister for Industry and Innovation Tim Ayres, who sees the AI opportunity as an essential part of the Future Made In Australia vision that he is building. But I want to be honest with you about what the analysis I have just walked through implies.

Data centres are important. In fact, data centres and enabling AI training are the strategic step ladder for Australia. But data centres alone will not be enough.

If we provide the land, the power and the buildings while foreign firms own the chips, the models and the customer relationship, we will capture some of the wealth while the largest rents accrue somewhere else. Which brings me to the harder question. Can a country like Australia compete in the intelligence layer itself?

Twelve months ago, most serious people would have said no. The frontier looked like a closed contest between a handful of American labs with capital budgets larger than the GDP of small nations, and the only rational strategy for a middle power was to buy well and adopt fast. That judgment is being overtaken by events, for three reasons.

First, the usage of open-weight models is now almost on par with that of closed models, and on their current trajectory, is expected to pass them within a matter of months – not years. Between January and mid-July this year, open models’ share of weekly volume rose by 15 per cent on OpenRouter, the largest AI routing platform, while closed models’ share of weekly volume fell by 16 per cent, leaving just a narrow 2-point gap.

Second, the cost of capability has collapsed. The strongest open models now sit within distance of the frontier at roughly a third of the price. Techniques that were research curiosities eighteen months ago — parameter-efficient fine-tuning, distillation into small task-specific models, retrieval, sophisticated post-training — mean that a mid-sized organisation could now take a strong open model and make it better than a frontier system at a narrow, valuable task, using its own proprietary data.

That is not a hypothetical. It is happening in Australian firms today. These market developments carry opportunities but also present safety and security risks, particularly to critical infrastructure – and these are matters which the Government is working through, and I’d encourage you to keep up with the advice from the Australian Cyber Security Centre on the security risks associated with AI-enabled systems.

Third, the market is becoming genuinely more competitive. We are seeing other players, such as Mistral from France. It seems likely the future will see routine tasks routed to cheaper, less capable models, with the most difficult of tasks reserved for the more capable – but expensive – ones.

This has strategic implications for Australia and the region – both in terms of opportunities and risks. The overall implication is that capabilities which seemed out of reach for a country like Australia just twelve months ago, whether strategic or commercial, now appear more feasible – and its likely this trend will continue. Furthermore, the layer where commercial models are specialised, evaluated, deployed and embedded into real industrial processes is increasingly contestable — and it is where a great deal of the durable value is going to be created.

It is where a Future Made in Australia can be built. We have real assets to work with. More than fifteen hundred AI-focused companies.

Genuine strength in medical imaging, in agriculture, in mining technology, in fintech, in quantum. World-class research institutions. And data of extraordinary quality in domains where we are already a global leader.

To give these Australian ideas the best chance of scaling globally we need Australian customers — business and government — to give Australian AI companies what they most lack: reference customers and a pathway to scale. I have made this argument bluntly before and I will make it again. In a noisy global marketplace, the question of whether there is an Australian option, and whether it deserves serious consideration, is too often drowned out.

That is why I convened a Buy Australian AI roundtable earlier this year, and why this needs to become a habit rather than an initiative. And the Government will have further announcements with the National AI Centre to make in coming weeks on this front. The Finance Minister and I have been doing further work and plan to address Commonwealth CIOs on the matter later this year.

Second, we can turn our advantages in compute into advantages in models. Our Expectations of data centres and AI infrastructure developers already ask hyperscalers to make compute available to Australian start-ups and to partner with our innovation ecosystem. Our goal is to convert physical investment into national AI capability.

We can build the ecosystem around the infrastructure, deliberately, by making it a condition of the welcome. Minister Ayres and I are working on how we can make that expectation a requirement of hyperscalers, to support the Prime Minister’s National AI Standards. Third, we should retain the value from scarce national assets.

We need to convert Australian data into Australian value. Our agricultural data, our resources data and our creative outputs are valuable intellectual property. Ownership, licensing and procurement determine who captures that value.

It is why the Prime Minister was right to rule out weakening the rights of Australian creators as a price of investment. Our data and our creative work are national assets, and assets are licensed, not surrendered. And fourth, we need to leverage the ingenuity of Australian firms across the economy to unlock value through how they apply AI – to augment and enhance the creativity of our workers in supporting growth, productivity and creating new opportunities.

We are not starting from nothing. The National AI Plan sets out the framework. There is a billion dollars committed for critical technologies through the National Reconstruction Fund.

There is an AI Accelerator round of the Cooperative Research Centres program to move ideas from Australian labs into Australian companies. There is the National AI Centre, the AI Adopt centres for small business, and an AI Safety Institute. And there is now an Office of AI to bring action across government together.

That is a serious foundation. My argument today is that it needs to be pointed, deliberately and relentlessly, at the question of value capture. Australia has, for most of our history, built our economy on foreign capital and foreign technology.

For decades, people wrung their hands about that model. We weren't sophisticated enough. We were just a quarry.

But the truth is we made it work for us. We made it work because we understood the iron law of national accounting: GDP and GNI are not the same thing. When foreign money or foreign technology comes in, it lifts productivity and GDP as Australian workers and firms produce more.

But Australian national income, the income that actually accrues to Australians, often rises by less than that. Some of the return flows back to the foreign owner as profits and dividends. GDP measures what is produced here.

National income measures what we keep. We worked this distinction in the resources boom. Yes, we imported foreign capital and we imported foreign technology.

And yes, foreign shareholders earned substantial returns. But we owned the choke point in the value chain. We owned the ore.

So Australians earned good wages in jobs that couldn't be shifted abroad. Local Australian companies built world-leading mining services businesses. Governments collected taxes and royalties.

So productivity rose, GDP rose, and a large part of the resulting income remained Australian. That is why the resources boom lifted not just our output but our living standards. AI does not automatically offer us that deal.

So Australia faces two tasks in the age of AI. First, use it. Adopt AI everywhere we can, to lift the productivity of Australian workers and Australian businesses.

That sets the size of the prize. Second, build around it. Create the companies, the capabilities, the intellectual property and the complementary assets that ensure Australians capture a meaningful share of the value AI creates.

That sets our share of the prize. We should not choose between productivity and value capture. We need both.

Because GDP tells us how much Australia produces. But our prosperity depends on how much of the resulting income Australians receive. Expectation 5 sits within a broader effort, the National AI Plan, and it works alongside investments in critical technologies, adoption support for small business, an AI Safety Institute, and a new Office of AI to bring the effort together.

But I have dwelt on it today because it is the clearest example of the principle at the heart of this speech: taking our single greatest point of leverage, the world's appetite to build data centres in Australia, and pointing it deliberately at the question that actually determines our prosperity. Value capture. That is what Australia got right in the resources era.

We welcomed the world's capital and the world's technology, but we brought something valuable of our own to the bargain, and used that build our step ladders. Our task in the age of intelligence is to do it again. Subscribe and stay up to date Connect with us PM&C acknowledges the traditional owners and custodians of country throughout Australia and acknowledges their continuing connection to land, water and community.

We pay our respects to the people, the cultures and the elders past, present and emerging.