Australia’s AI Strategy Is Australia

Canberra is betting that land, power, law and geopolitical trust can turn a country that does not own the frontier into one that sets the terms.

Anthony Albanese opened his speech on artificial intelligence with Medicare.

From there he moved through the eight-hour day, universal superannuation and the social media ban for children, a sequence that had nothing to do with chips and everything to do with a theory of the state. In this telling, Australia is a country that meets a large economic or technological force and applies a social settlement to it: writes conditions into law, and occasionally produces a standard the rest of the world copies. Across the speech and the television interview that followed, a second Australia kept appearing, in a different list. Space to build. Sun to generate power. Minerals, universities, Five Eyes membership, a reliable legal system, proximity to the fastest growing region on earth.

The first Australia writes rules. The second is a site. Wednesday’s promise of a single mandatory national framework for AI, and for the data centres that underpin it, is where the two meet.

The clearest statement of what the government thinks it is doing came that evening, on 7.30 (evening news show), and it came under pressure. Sarah Ferguson put the obvious question: Canada is spending two billion dollars building sovereign computing capacity, the United Kingdom is doing the same, and Albanese had spent the morning citing the NBN as proof Australia can build sovereign capability when it chooses. Why not do the same with AI? The Prime Minister answered that the investment here would be overwhelmingly private, and then said the thing the whole policy rests on. “This will be all about sovereignty,” but the way to it would be Australian standards and Australian rules. Your investment is welcome. These are the conditions.

Call that sovereignty by jurisdiction rather than sovereignty by ownership. It is a strategy, not an evasion, and it deserves to be examined as one.

It helps to notice that four different things travel under the one word. Sovereignty over a model means holding the weights and deciding who may use them. Compute sovereignty means assured access to the machines that train and run advanced systems. Data sovereignty means deciding how information held here is stored, moved and used. Regulatory sovereignty means the power to set conditions on any technology operating inside your borders. Australia is putting its weight on the last two, betting that hosting the physical infrastructure will improve the second, and leaving the first, beyond smaller local systems and specialised applications, to other countries. For a nation of 27 million that may be the right distribution of effort. It is a choice, though, and using one word for all four lets control in one layer sound like control of the whole stack.

The copyright declaration belongs to the same theory. No company, Albanese said, should train on Australian books, music, art or news without the artist’s control, and “Anything less is theft.” That evening he told Ferguson that intellectual property would be mandated alongside the data centre standards, control and payment both, and when she asked whether he was prepared to watch Anthropic walk away from its reported A$21.6 billion Australian capacity search rather than bend, he answered that the advantages Australia offers would keep the investment coming. Confidence in the jurisdiction, again, standing in for machinery that does not yet exist. The government ruled out a text-and-data-mining exception last October. Nine months on there is still no settled mechanism for licensing or disclosing AI training, and no low-cost enforcement path an individual writer could use; whether the Copyright Act itself must change is still, on the Prime Minister’s own account, under examination.

Even the new Office of AI reads differently once you see the theory. It sits inside the Department of the Prime Minister and Cabinet from Wednesday, and the coordination it exists to do is mostly domestic: copyright, safety, energy, planning and the states. Albanese was explicit that national standards are needed partly to stop the states bidding against each other for projects the way they bid for sporting events. An office of artificial intelligence, then, but also an office of Australian federalism, whose first task is to fuse three tiers of government into one bargaining position before the negotiations that count.

Which brings us to what is being bargained over, because the strategy only makes sense once you are literal about it. A generative model presents itself as weightless, an answer arriving from nowhere. A data centre is the correction: a building that turns electricity, water, land and fibre into computation, a power station running in reverse. Nor is a data centre simply an AI machine. The same buildings carry cloud services, bank and government systems, health records, telecommunications, streaming and scientific computing, and some contain no frontier accelerators at all. Generative AI is the demand shock accelerating the build-out rather than the definition of the asset.

The scale is already an energy story. Australia has 162 data centres in operation, mostly in Sydney and Melbourne, drawing about 2 per cent of grid-supplied electricity. The market operator expects that share to reach roughly 6 per cent by 2030 and around 12 per cent by 2050, and at the end of March it had eleven large projects, 5.4 gigawatts of maximum demand, queued in the connection process. Numbers like that stop being forecasts about one industry and become assumptions about the future design of the electricity system.

The physics cuts both ways, too. In July 2024 a single network fault in Virginia dropped about 1,500 megawatts of data centre load off the grid at once, and the system around it shuddered. A large data centre approval is an energy decision, a water decision, a planning decision, a foreign investment decision and a national security decision, all wearing one development application. Some large data storage and processing assets already sit inside Australia’s critical infrastructure regime, with the registration and reporting duties that follow.

The National AI Plan goes further still and imagines the country as a “trusted exporter of AI computing power”, a computation platform for the Indo-Pacific: allied, stable, increasingly renewable, close to the demand. That is a larger ambition than fast following. It is the site learning to think of itself as a port. Five Eyes membership is doing economic work in this argument. Australia is selling more than land and power: it is selling a stable allied jurisdiction, connected by cable to Asian demand. The country itself is part of the product.

The government reached for the resources analogy before any critic could. Andrew Charlton, the assistant minister carrying much of this agenda, has been calling Australia the lucky country of the new data centre boom, and frames the national expectations as what companies owe in return for access to the grid, the land and the market. His account of gas is an unusually sharp self-criticism from inside government: “We let the boom set the terms, instead of setting the terms of the boom.” Households and factories ended up paying more for gas drawn from beneath their own feet, and the repair, a domestic reservation policy setting aside Australian gas for Australians, arrived a decade after the damage.

He points to Dublin, where data centres passed a tenth of national electricity demand and the grid effectively stopped taking new connections, and to the United States, where the scale came first and the backlash after. Getting the rules in before the concrete is the whole lesson, and on the physical side Australia is proposing to apply it.

The analogy carries a warning as well as a plan. Australia is practised at hosting booms and less practised at keeping the layers of value they create. The boom appears as investment in the national accounts and, as the Bureau of Statistics has pointed out, much of the equipment filling the buildings appears again as imports. The building is fixed here. Control of almost every layer above it can sit elsewhere: the chips, the model, the cloud contract, the engineering team, the intellectual property, the profit. A country can host a great deal of computation without acquiring much computational capability.

The government’s own planning documents say as much. They describe a dual-track ecosystem, an established AI-taker and a developing AI-maker: one track adopting systems built elsewhere, the other building specialised products where Australia already has depth, in health, agriculture, mining and advanced manufacturing. The Productivity Commission’s work supports the same case: for a country like this, most of the near-term value comes from diffusion, ordinary firms and public services adopting the technology well, rather than from winning a race Australia was never entered in. Put the whole design together: imported frontier models at the base, hyperscale compute physically here, Australian firms building applications and smaller models on local data and local strengths, national standards governing use and Australian facilities selling trusted computation into the region. The architecture is coherent. Its weakness is in the joins, because hosting does not connect those layers by itself. Racks do not teach. A hyperscale campus down the road is not a research allocation, and the construction jobs end when the construction does. Everything depends on a conversion mechanism, the means by which buildings full of other people’s computers become access, skills, firms and public capability.

The government deserves credit on a point its critics mostly missed: it has already named that mechanism. The expectations published in March say that providers of large-scale compute are expected to open access to Australian start-ups, researchers, small businesses and not-for-profits, and to do it “on favourable terms”. They say major investors should deploy engineers and researchers in Australia, build local technical capability, spend into Australian supply chains. Read as a list of demands, it is exactly right. Read as law, it does not exist.

Ferguson asked the Prime Minister directly whether the new standards would be more than expectations, and his answer was unambiguous: mandatory. But the list he then gave was energy additional to the grid, transmission costs, water, location. The capability half of the bargain, the compute access, the engineers, the skills and the supply chains, stays in the language of should.

The costs data centres impose on Australia are heading for the statute book. The benefits that would make Australia more capable remain requests.

The industry has noticed which is which. In its response to the March expectations, the sector’s peak body singled out the favourable-terms clause, asking what role government should properly play in shaping commercial arrangements between providers and researchers. The July speech, meanwhile, attached no new public money. The A$460 million the National AI Plan points to is existing and committed funding gathered under one heading, useful programmes, none of them remotely a conversion mechanism for tens of billions of dollars of infrastructure.

So Australia stands, for now, as a demanding landlord of the buildings and a tenant of much of what runs inside them, with the tenancy softened by polite requests. I notice I am inside this sentence rather than above it. Universities are among the intended beneficiaries of favourable compute access, and I run a university library: the kind of institution that in time, deploys, governs and increasingly depends on these systems. The soft half of the bargain is meant for institutions like mine. I would like the verbs upgraded.

There is an Australian instrument for exactly this, and Charlton’s own gas story points to it. When the gas market failed households, the eventual repair was reservation: a share of the resource set aside, by law, for the country it came from, ten years too late. If data centres are the next resources boom, take part of the royalty in the commodity the boom produces. For every project above a defined scale, require an Australian compute contribution from whoever controls the capacity: reserved access where the proponent owns or allocates the accelerators, a levy into a pooled national compute-credit scheme where it does not, with access and pricing published so that favourable terms can be measured rather than advertised.

Attach auditable obligations covering engineering and research teams based here, graduate places, apprenticeships and partnerships that create Australian-owned or Australian-licensed intellectual property and lasting technical capability, tested by how much difficult technical work remains after the ribbon is cut rather than by how many people poured the concrete. Give government and other critical users continuity and exit rights, the contractual ability to move workloads, keep their data, substitute models and stay operational if a provider changes its terms.

And write the physical rules with precision. Additional power should mean generation and firming that did not exist before, matched to a facility’s continuous load, and a staged campus should not slip the standards because its first stage was approved before they existed. That question went live within a day. On Thursday morning, asked whether the rules would reach data centres already under construction, the Prime Minister said they would apply to new proposals: “You can’t retrofit.” The night before, the argument for the standards had been to get in front of the boom instead of repairing after it. What counts as new, for a staged campus or a later expansion, will decide how much of the coming build-out the rules ever reach. Copyright needs the same upgrade from principle to machinery, licensing that works and remedies a single writer could use. None of this requires a taxpayer-built rival to OpenAI. All of it could be made part of the mandatory national framework the Prime Minister wants in place early next year.

The window for writing those terms is the window in which the investment still wants to come, and it is open now. Charlton says investors want to be here. The connection queue says the same, eleven projects and 5.4 gigawatts deep. Bargaining power of this kind peaks before approval and decays afterwards, which is why the gas reservation arrived ten years late and why a compute reservation should not. Albanese spent Wednesday telling the country that Australia will set the terms, and on the evidence he means it. The test is whether the terms leave Australia more capable than the buildings it approves. The concrete is already being poured.