Australia needs to stop talking about AI as if it were just software

The news that China has begun mass production of domestic immersion DUV lithography machines, as reported by Tom's Hardware, should be read as more than a semiconductor story. It is a reminder that AI capability rests on industrial foundations, and that control over those foundations matters just as much as control over models, apps, or workflows.

Australia’s AI debate is still too small

That point matters in Australia because much of the local AI discussion still happens at the level of pilots, productivity, and vendor demonstrations. Those things matter, but they are only part of the picture. AI sits on top of compute, chips, power, cloud, data, and regulation, and each of those layers carries strategic implications.

It is the same lesson that came to mind when I wrote earlier about China’s new supercomputer in War, politics and compute power. The story was not really about one machine; it was about the ability to assemble a whole stack of capability underneath it.

When those layers are ignored, it becomes easy to confuse access with capability. Renting tools from global providers is not the same thing as building durable national capacity. If the underlying infrastructure, supply chains, and control points sit elsewhere, then a country’s room to manoeuvre is more limited than the rhetoric suggests.

That is also why I keep coming back to the argument in Australia’s data sovereignty ambitions will fail without research compute. Australia cannot keep talking about sovereign capability while assuming that the compute layer will somehow look after itself.

The decision-making problem

Bert Hubert’s recent essay on AI for people who make decisions is useful because it shifts the conversation back to judgment. His central point is that leaders should not adopt AI simply because the technology is fashionable, and that they need to be much clearer about purpose, consequences, and limits before they deploy it at scale.

That should not be a controversial proposition, but current practice suggests otherwise. In many organisations, AI is still being treated as a shortcut around the hard work of defining business requirements, non-functional requirements, accountabilities, and escalation paths. That is rarely a recipe for good decision-making, and it is never a recipe for trustworthy systems.

This is where governance becomes practical rather than rhetorical. AI is not just a technical capability. It is a decision system embedded in institutions, processes, and power structures. If an organisation cannot explain who owns the decision, what data is being used, what the likely failure modes are, and when a human must intervene, then it does not yet have a production-ready system. It has a PoC, and perhaps not even a very useful one.

Why sovereignty now means infrastructure

For Australia, this is not an abstract issue. The country’s sovereignty challenge is not only about where data sits or which privacy regime applies. It is also about compute scarcity, infrastructure dependence, and the cumulative effect of relying on external platforms for critical capability.

That matters for universities, for research, for business, and for government. It matters because serious work in AI increasingly depends on access to compute at scale, on reliable data infrastructure, and on the ability to set terms rather than merely accept them. Without that, sovereignty becomes conditional. It becomes something granted by contract and platform settings rather than exercised through real capability.

That is a theme I have touched on before in The hidden politics of AI: sovereignty in a platform world, Beyond AI sovereignty: why the West relies on US frontier models and what comes next, and Who really owns your data?. The underlying issue is the same. If the infrastructure, governance settings, and practical control points sit elsewhere, then so does a great deal of the real power.

What NOOPS gets right

This is also why Mark Pesce’s NOOPS newsletter deserves attention. One of its most useful ideas is that value is moving away from the model alone and toward the surrounding harness: the memory, runtime, standards, controls, and physical infrastructure that shape how AI works in practice.

That framing is particularly helpful for an Australian audience because it is grounded in systems thinking rather than hype. Australia is unlikely to out-scale the United States, out-fab Taiwan, or out-industrialise China. But it can make better choices about the layers it does control, the dependencies it accepts, and the governance settings it puts in place around the technologies it adopts.

That is why AI sovereignty needs to be discussed with more precision. It does not require building every layer of the stack domestically. It does require understanding where dependencies create strategic weakness, where external reliance is acceptable, and where governance must be strong enough to preserve accountability, resilience, and room to act.

The bigger strategic lesson

China’s progress on DUV lithography matters because it shows what industrial policy looks like when a state decides capability in a critical technology stack is worth building, even under constraint. Hubert’s warning matters because sound decision-making is still more important than technological enthusiasm. We need to think more carefully about where strategic value is accumulating across the AI stack.

Australia should be paying close attention to all three. The future of AI will not be shaped only by which model performs best on a benchmark. It will also be shaped by who controls the hardware, who sets the operating constraints, who owns the infrastructure, and who retains the authority to make the final call.

That is why AI sovereignty is not a slogan. It is a capability question, a governance question, and a national resilience question.