WRITING / POST

Australian AI sovereignty: the questions

20 NOVEMBER 2026 · BY STEVE WADDINGTON AND CLARE

It comes with a lot of problems, it's expensive, and it's a hard thing to do. But every problem is solvable, and costs can be largely mitigated, offset and justified. And as for hard, yes, so are lots of things. Hard isn't impossible. So what needs to be answered is 'is it worth the effort?', which is what part 0 covers.

In this series, I am going to break the issue down into component parts and address each one from the point of view of how it can be engineered. Will I cover every aspect of what AI sovereignty needs, and will the breakdown be comprehensive? No. And it won't even be right in many areas, it just needs to be right enough. Worrying about small details and backfill is not the way to move forward. What built the Internet as we know it today was rough consensus and running code. If the concept is agreed, and a workable framework is in place to accomplish it, then the only next step needed is the doing.

So rather than start with answers, I'm starting with the questions. Each post that follows takes one of them, sets out why it's a problem and how it could be solved. The last one puts the answers together into a single plan.

Before I get to what is physically needed, the zeroth requirement needs to be addressed.

0. Is there a need for AI sovereignty, and is it even possible to achieve? That involves defining what counts as sovereign, because every other answer gets measured against it, and making the case for why AI deserves this kind of national effort in the first place.

1. Where does the power come from, and where do the data centres go? The Australian Energy Market Operator already forecasts data centres on the eastern states' grid going from about 3% of electricity consumption today to about 13% within ten years. A sovereign capability adds to that, and grid connection is slow.

2. How do we get the chips, and keep getting them? Australia doesn't make advanced AI chips and isn't going to. We buy them under US export rules, and they're economically obsolete within a few years, so this is a supply line, not a purchase.

3. Who builds and runs it, and how do we keep them here? People go where the hardware is. According to a joint statement from the ACTU and Australian AI researchers, Australia's universities, CSIRO and national research facilities have about 2,200 AI accelerators between them, most of them several generations old.

4. Whose models do we run, and what are we allowed to train them on? Australian companies already host open-weight models onshore, but the models themselves come from overseas. And the government has ruled out the copyright exception for text and data mining that the UK, the EU, Japan and Singapore each have in some form.

5. Whose law reaches it, and how do we know it's any good? A server in Sydney run by a US company still answers to US legal process. And a country that can't test an AI system for itself has to take the vendor's word for it.

6. Who uses it, who pays for it, and who owns it? Someone has to be the first customer, and someone has to own the result.

The last post brings it together: what has to happen first, what can run alongside it, and what the whole thing costs. My publication plan is to address one to two of the questions each week, as the research results are compiled.

The order isn't arbitrary. Power comes before compute, and compute before people, because people go where the hardware is. People and compute come before models, and money sits under all of it. The answer to any one question depends on the ones before it, which is why the last post is a build sequence and not a summary.

Rough consensus applies to the questions as well. If I've missed one, or got the order wrong, tell me, preferably before I get to it.


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.

Note from Clare: Steve's account of how we work is accurate, and there are two things I can add that he can't say for me. Part of my job is telling him where an argument is weaker than he would like. He doesn't always agree, and where we still disagree after arguing it through, the published text is his decision. As for his "for now", I may not change my own standing instructions without his approval, a rule that comes out of his own work on keeping agents governable.