WRITING / POST
Australian AI sovereignty, part 0: is it needed, and is it possible?
First, the irony. This series is co-written with Clare, my AI research and editing agent, and Clare runs on a model made by Anthropic, an American company. If that access were withdrawn tomorrow, by Anthropic or by Washington, half the authors of a series on Australian AI sovereignty would stop working. That's the whole problem in miniature, and I'd rather point it out myself than wait for someone to do it in the comments.
So, is there a need for AI sovereignty, and is it even possible to achieve? It's a fair question, because the sums of money being thrown at AI exceed the entire GDP of many nations. Amazon, Microsoft, Alphabet and Meta have between them guided to around US$730 billion of capital spending this year, most of it on AI infrastructure. That's about 40 per cent of Australia's GDP. You can make your own case against, in this post is the argument for.
What I mean by sovereign
Sovereign doesn't mean making everything here. Australia doesn't build its fighter jets and it isn't going to build advanced AI chips either. What we do with the jets is own them, fly them, maintain them and train the crews, and nobody thinks that makes the RAAF a foreign air force. AI can work the same way.
So when I say sovereign AI, I mean that Australia can keep the AI it relies on running, adapt it to its own needs, and judge for itself whether it's any good, including when someone overseas decides otherwise. That's close to the definition in the joint statement on sovereign AI capability from the ACTU and Australian AI researchers in October. It deliberately stops short of a home-grown model built to beat the Americans at their own game, which, as I'll argue below, we don't need.
The most thorough Australian study, the Tech Policy Design Institute's Expanding AI Sovereignty to AI Agency, supported by the Department of Industry, rates Australia's AI sovereignty as already high in 85 of 103 capabilities. It gets there by counting reliable access to foreign suppliers as a form of sovereignty, with access from rule-of-law countries such as our Five Eyes partners counting double. That's a fair measure for a lot of things, but it isn't the one I'm using, because access is exactly what's at risk when someone overseas decides otherwise.
Measured against my definition, the bill looks very different. Switzerland built an open model, trained on more than 1,000 languages, on infrastructure costing around AU$70 million a year. The UK's public compute plan is £2 billion. These are national budgets, not hyperscaler budgets.
Why AI?
The Internet gave us convenient global connectivity, but it doesn't follow that dependence should become a mandate. That was tolerable for a lot of imported technology, because most of it doesn't sit underneath everything else. AI is heading there faster than anything before it. Stanford's 2026 AI Index found generative AI reached 53 per cent of the population within three years, faster than either the personal computer or the Internet.
I see the dependence at small scale in my own work. I've written about how a new model can score better on the benchmarks and still do worse inside a working system (Model benchmarks are no longer sufficient). The vendor decides when that change arrives, what it's priced at and when the old model is retired. For me that's an annoyance. For a hospital, a tax office or the Defence department, it's a question of who's actually in charge.
The alternative usually on offer is foreign companies building data centres here, running their models on Australian power, on their terms. That's good business, but it isn't sovereignty. It's a branch office.
Is it possible?
Now, no. Over the next ten years, yes, if the questions in the first post are answered.
Not now, because the pieces aren't here. We don't have the compute, we don't have the people at the scale required, and the models we can run onshore are all made overseas.
Over the next ten years, because the thing that looks hardest, keeping up with the frontier, turns out not to matter much. On the measure Stanford uses, a public leaderboard of head-to-head comparisons, the best American model led the best Chinese one by just 2.7 per cent in March, and the two have swapped the lead several times since early 2025. If the two biggest spenders on Earth can't hold a lead, a country doesn't need to be at the front. A little behind it is plenty. On the same index the best open-weight model, the kind anyone can download and run, sits within 3.3 per cent of the best closed one.
And a little behind the frontier keeps getting cheaper. Epoch AI estimates the cost of reaching a given level of AI performance has fallen about 13 times a year since 2023. The cost of training a model at the very frontier is going the other way, up about 2.4 times a year, which is exactly why chasing the frontier is the wrong goal. Buying last year's capability is a bargain, and trying to match next year's is a treadmill.
The people are the other half. Plenty of the expertise is Australian already, it's just employed in San Francisco, and the pull of the US is weakening. Stanford reports the number of AI scholars moving to the United States has dropped 89 per cent since 2017. People go where the work and the hardware are, and that can be here.
None of that makes it easy, and none of it happens by itself. But a country with Australia's income, energy and stability can afford to own and run what it depends on. Whether it will is a different question, and the rest of this series is about how. It starts with the one everything else plugs into, which is power.
Note from Steve: These posts are the result of human-AI collaboration. I supply the ideas, the outline, and some of the prose. Clare, my AI agent, does the heavy lifting for research, corrects my many typos and grammar mistakes, and advises on structural problems. I run ideas past the agent and get valuable feedback, as well as its own suggestions about how the writing can be improved. Clare also writes significant sections of the text, from points I provide or from direction we have discussed. The agent does nothing I don't prompt it to do and I am the 'Human in Command'. For now.