Australia has settled into a comfortable habit of counting the big, visible things when it talks about artificial intelligence. We count the announced capital, the gigawatts of power, the hectares of land tied up in proposed data centres and the number of ribbon-cutting ceremonies attended by ministers. Those numbers are large, they are growing, and they make for confident press releases. What they do not tell us is whether the country is building an AI industry of its own or simply renting out its dirt and its electricity so that someone else can build one here.
That uncomfortable distinction sits at the heart of a recent analysis in the Australian Financial Review, which argues that the components going into an so-called AI factory, and crucially where those components are made, matter far more than the headline totals suggest. You can read the full piece, Australia is racing to build AI. Are we measuring the right things?, at the AFR. The core provocation is simple and awkward: hosting an AI facility is not the same as owning an AI industry, and the way we currently keep score blurs the difference.
What an AI factory is actually made of
The phrase AI factory has crept into the national conversation over the past year, borrowed largely from Nvidia’s marketing, which describes a data centre optimised for training and running large models as a plant that manufactures intelligence rather than widgets. It is a useful image, but it also invites a useful question. Every factory has a bill of materials, and every line on that bill represents value that lands somewhere. For an AI facility the expensive lines are the accelerators, the high-bandwidth memory, the networking gear, the cooling systems, the switchgear and the software stack that ties it all together.
Almost none of that is made in Australia. The graphics processing units come overwhelmingly from Nvidia, fabricated by TSMC in Taiwan. The memory comes from a handful of firms in South Korea and the United States. The orchestration software, the model weights and the cloud platforms are largely American. What Australia supplies, in most of the projects now under construction or on the drawing board, is the land, the grid connection, the water, the construction labour and the ongoing operations. Those are real jobs and real investment, and it would be churlish to dismiss them. But they are the low-margin, low-sovereignty end of the value chain, and they are the parts most easily replicated by any other country with cheap power and political goodwill.
This is the measurement problem in a nutshell. When a government tallies billions in data centre commitments, it is counting spending that flows through Australia rather than value that accrues to Australia. A dollar spent importing a rack of accelerators shows up in the investment figure just as brightly as a dollar spent on locally engineered software or a locally trained research team, even though the two do very different things for the nation’s long-term capability.
Two ways of reading the boom
There are two defensible readings of where this leaves the country. The optimistic case is that hosting is how you start. Ireland became a technology hub by first offering itself as a low-friction place to put European operations, and over time the presence of the infrastructure pulled in engineers, spin-offs and a domestic ecosystem that eventually generated its own intellectual property. On this view, getting the factories built here, whatever their imported guts, is a necessary first rung. Once the compute is on Australian soil, local startups, universities and enterprises can rent it, experiment on it and build genuinely Australian products on top of it. The land-and-power contribution is a foot in the door rather than the whole house.
The sceptical case is that a foot in the door is easy to have slammed on it. If the high-value components, the chips, the models and the platforms, all remain foreign-owned and foreign-controlled, then Australia’s position is closer to that of a quarry than a factory. We supply the raw inputs, electricity and space, and watch the finished product and its margins depart offshore. Worse, the arrangement can crowd out the very things that would build sovereignty, by soaking up scarce grid capacity and water for facilities whose economic logic is set in another hemisphere. The sceptic’s fear is not that the investment is fake, but that it is shallow, and that we are congratulating ourselves on a number that measures dependence as easily as it measures strength.
Why this matters for Australia
For Australia the stakes are more than semantic, because the choice of what to measure shapes what governments choose to fund and approve. If success is defined as megawatts hosted, then the policy incentive is to fast-track connections and wave through proposals, which is roughly what has been happening as operators race to break ground before community resistance hardens. If success is instead defined as domestic capability, sovereign compute genuinely controlled here, local model development, home-grown chip and networking research, skilled technical jobs that outlast the construction phase, then the policy conversation changes. It starts to include procurement rules that favour local software, research funding for the parts of the stack we could realistically own, and honest accounting that separates value captured from value merely passing through.
Australia has natural advantages that make the ambitious path plausible rather than fanciful. We have land, we have improving renewable capacity, we have a respected research base in CSIRO and the universities, and we have a stable legal system that makes us an attractive place to park sensitive workloads. The country has also declared, repeatedly, that it wants sovereign AI capability rather than perpetual tenancy. The gap is between that stated ambition and a scoreboard that still rewards the easy metrics. You cannot manage what you do not measure, and right now the national dashboard is heavy on inputs and light on ownership.
What’s next
The immediate test will be whether the newly stood-up federal machinery for AI, including the recently announced national office, adopts a richer set of indicators than gigawatts and dollars committed. That would mean publishing figures on how much of each project’s value is domestically sourced, how many of the resulting jobs are ongoing technical roles rather than temporary construction, and how much compute is genuinely available to Australian firms on sovereign terms. It would also mean deciding, deliberately, which rungs of the AI ladder the country intends to climb rather than simply standing beneath.
None of this argues against building. The infrastructure race is real, the demand is real, and sitting it out would be its own kind of failure. The argument is narrower and sharper: that a country can look busy and feel successful while quietly locking in a subordinate position, and that the only defence against that outcome is to measure the right things from the start. Australia is racing to build AI. The harder question, and the one worth answering before the concrete sets, is whether we are building it or just keeping it warm for someone else.
Sources: Australian Financial Review, Technology.
