The Case for Sovereign Intelligence, Part 2

Abstract

In 2018 we submitted a paper to the Senate Select Committee on the Future of Work and Workers titled The Luddite Fallacy Fallacy.¹ That paper argued, on first principles, that intelligent machines would substitute for human labour at the level of tasks rather than jobs; that this substitution would be economically inevitable on account of the scalability of computation; and that, left unmanaged, the principal beneficiaries would be the small handful of firms then already positioning themselves as global cloud-based AI service providers, what we then called the Tech Giants. We predicted a substitutable human mind within ten years. Eight years on, the technical predictions have been broadly vindicated and the structural predictions about ownership are now an observable fact.

This paper takes up the question the 2018 submission deliberately left to a later treatment. Given that the substitution is happening and will continue to happen, whose intelligent machines should be doing it, and whose economy should they become? The answer we develop is that Australia must build, own and operate the full stack of artificial intelligence on which its future economy will run: the energy that powers it, the compute that runs it, the models that perform on it, the data that trains it, the talent that maintains it, and the governance that sets its limits. We refer to this throughout as sovereign AI.

It is essential at the outset to distinguish ownership from sourcing. A sovereign AI is not necessarily one whose every component was manufactured in Australia. The chips, the networking gear and even some of the model architectures may come from elsewhere; the same is true of the cars we drive and the planes we fly. What matters, and what sovereignty turns on, is who owns the assembled system, who can switch it off, who can read its outputs, who can redirect its priorities, and who is accountable for its conduct. By that test very little of what is being built in Australia today is Australian at all.

The case is therefore not for autarky. It is for a country that owns the operating layer of its own economy at the moment that operating layer becomes artificial. The deepest reason, developed at the close of this paper, is that the alternative is not merely a poorer Australia. It is, in the long run, a less peaceful one.

This subject cannot be kept to industry and infrastructure. The potential end of human work drags in questions we usually leave to philosophers: what an economy is for, what a person needs beyond a wage, and what a society owes its members once it no longer needs their labour. For a topic of this consequence these cannot be dodged. Humanity tends to stumble into its great changes and reason about them afterwards, once the shape is set and the cost is paid. This change is too large, too fast and potentially too irreversible to be met that way. Therefore, in places, this paper deliberately goes deeper than a policy document usually does.

This paper sets out the program and implementation; the foundation beneath it, an account of what intelligence is, why it is valuable, and why sovereignty attaches to it, is supplied by the predecessor companion paper, The Case for Sovereign Intelligence, Part 1.² Each paper stands alone; they are written to be read together.

1. From 2018 to 2026

The 2018 paper was, in the language of the day, alarmist. It claimed that there was no future of work for humans, that the Standard Employment Relation would erode into a contingent piece-rate workforce, that the elasticity of substitution between labour and intelligent machines would soon exceed unity, and that the Tech Giants would, in time, become the economy. Each of these claims was contested at the time. Each is contested still, but the contest has narrowed to timing. The persistence of the disagreement says more about the disciplines than about the evidence: economists are experts in the record, and the record is the past; their instruments are weakest at exactly the structural break this paper describes. Those of us who build automation for a living are watching the break arrive on the bench. When capable robotics arrives, on a timetable we judge in years rather than decades, the contest will end, because then the machines will be good at the embodied work as well, and then nearly all of what we call work will be within their capabilities.

We titled our paper "The Luddite Fallacy Fallacy" because its central thesis, the Luddite Fallacy was, and is still, implicitly denied in almost every official document on this subject. The Luddite Fallacy is the received economic claim that the original machine-breakers of two centuries ago were wrong: that mechanisation, taken across a long enough period, creates more jobs than it destroys, and that the displaced workers of any given technological transition will be reabsorbed into employment of equal or greater value. The Luddite Fallacy Fallacy that we posited, is the observation that this received claim is itself the error. The Luddites were correct in substance. The looms did take their work, and their lives, and their towns. They were wrong only about the timescale, because the offsetting employment, when it eventually arrived, was three generations later and in a different country. To the men who broke the looms in 1812, the prosperity of 1912 was not a vindication of the system that had ruined them. It was a vindication of them, on a clock they did not live to read.

The substitution of human labour by intelligent machines, now underway, is on the same curve. It runs slowly, it runs a little bit, then it runs all at once, and the long latency between the technical possibility of substitution and its visible economic effect is precisely the interval in which serious governments either prepare or do not. The 2025 Australian National AI Plan is a document of a government which has not. The economic profession told it, as it has told its predecessors for two hundred years, that automation produces aggregate gains; this remains true, but also remains beside the point. The aggregate gains accrue to the owners of the machinery. Whether and on what terms those gains are distributed to anyone else is a political question settled by the political instruments of the country in which it arises. An Australia which does not own the machinery of its own economy will have no political instrument with which to settle the question. That is the bridge from our 2018 paper to this one.

What the 2018 paper did not develop, and what we now must, is the question of ownership. That paper described the substitution process and the firms that would drive it. It did not interrogate what it means for a sovereign nation that the entity, which now performs the productive work of its economy, is a foreign corporation answerable to a foreign jurisdiction. We treated the substitution as the central problem and the ownership question as downstream of it. With eight years of evidence in, the ownership question has now become at least as urgent. A country with no sovereign capability over the productive engine of its own economy is no longer a country in any operational sense; it is simply a market.

There is also a substantive change since 2018 that the original paper did not anticipate. In 2018 the dominant case for concern was task substitution in the workplace. By 2026 the substitution has spread upwards into what was always taken to be the most distinctively human work: language, reasoning, design, persuasion and judgement. Large language models, which did not exist in their current form when we wrote, now write code, draft contracts, summarise medical histories, and conduct first-line legal and medical consultations. They are also being deployed, in their natural extension, to draft policy, model regulatory responses, and brief ministers. They proofed this document. The question of whose AI runs the country is no longer hypothetical; it is the question of whose AI is, today, briefing and lobbying the people who run the country.

The government's own response to this trajectory does not survive inspection. The 2025 National AI Plan fails in at least nine connected ways: (1) it stakes the workforce on re-skilling while offering no account of what the displaced re-skill into; (2) it does not register that industrial quality frameworks select for machine sameness over human variance; (3) it ignores the junior wall, the collapse of the entry-level work in which every profession trains its next generation; (4) it has no theory of displacement at all, because it counts jobs where substitution proceeds by tasks; (5) such guardrails as it contains are voluntary; (6) it celebrates foreign-owned data centres as sovereign capability, a claim §2 takes up; (7) it understates the energy demand of the infrastructure it invites, which §8 measures; (8) it never asks the distributive question, which §4 answers; and (9) it leaves small and medium enterprise to adopt or die unaided. The present paper is the positive complement to that verdict: the current plan will fail, and what follows is an actual plan that should replace it, and the principle of ownership that should organise it.

2. Ownership, Not Sourcing

The first thing a national AI policy must get right is the distinction between two propositions which are usually conflated in current debate. The first proposition is that Australia should manufacture more of the technological inputs to AI. Chips, racks, optical interconnects, photovoltaics, batteries. The second proposition is that Australia should own and control the AI systems that run on those inputs, regardless of where the inputs themselves were made. These are distinct questions with distinct answers, and treating them as the same question is a principal analytical error of the current debate.

The component question is a matter of supply chain robustness, defence-industrial policy, and trade balance. It is important and it is real, but it is bounded. Even at full strength, an Australian semiconductor industry would not produce a leading-edge logic process; the capital expenditure required to do so is on the order of twenty billion US dollars per fabrication facility and rising,³ and the existing players (TSMC, Samsung, Intel) operate at scales that no Australian programme will replicate in any timeframe relevant to this debate. Australia can and should produce more of the secondary components: packaging, photonics, specialised analogue, power electronics; but the leading-edge silicon at the heart of any modern training cluster will, for the foreseeable future, be imported. That is not a sovereignty failure. It is the same condition that applies to commercial airframes, to high-end pharmaceuticals, and to a great many other technologies on which we depend without any sense of crisis.

The ownership question is different. It is the question of who owns the integrated system at the point of use: who has root access to the datacentre, who holds the encryption keys to the customer data, who chooses what the model is permitted to say, who decides which workloads run when capacity is constrained, who can lawfully be served with a subpoena for the records, and who is the entity that pays Australian tax on the value the system generates. These are not properties of the chips. They are properties of the company that owns the equipment that the chips are mounted in, and they remain properties of that company regardless of where the chips were fabricated.

A simple example makes the distinction concrete. Imagine that an AI datacentre operating in Australia, owned by a foreign corporation, was tomorrow ordered by its parent's government to deprioritise Australian customer workloads in favour of customers in another jurisdiction during a regional crisis. The operator would, lawfully or not, have to choose between its host country and its parent. Imagine, alternatively, that the same datacentre was ordered by its parent's government to surrender its Australian customer data to a foreign intelligence service. In both cases the question is not whether the chips inside the racks are American or Taiwanese. It is whether the operator is Australian. The sovereignty question is the ownership question, not the sourcing question.

These examples were not chosen as hypotheticals. On 9 June 2026 the US company Anthropic released Mythos 5 and Fable 5, to the public and to its paying subscribers. This was probably the most capable model in the World at the time. Three days later they were gone. On 12 June the United States government issued an export-control directive barring access to both models by any person who was not a United States national, and Anthropic complied, switching them off worldwide.⁴ Every Australian who relied on them, including those who had paid for them, like myself, lost them overnight. The trigger was a security matter internal to the United States. The affected parties included an entire allied nation with no part in the decision and no recourse against it. Observe the precise shape of the event: the access test was nationality, the instrument was a foreign government's order, and the vendor, which disagreed with the recall, complied anyway, because it answers to its own state and not to ours. On 30 June the United States lifted the order and the models returned. This restoration repeats the lesson rather than softens it: access was switched off by a foreign government and switched back on by the same government, on its own timeline, for its own reasons, and Australia was a spectator both times. Ownership was the only protection that would have mattered, and Australia did not have it.

This is why the central plank of the current National AI Plan, the celebration of more than one hundred billion dollars in announced foreign data centre investment as evidence of "sovereign capability"⁵, is precisely the opposite of what it claims to be. A country whose AI infrastructure is hosted in Australia but owned in Seattle is not a sovereign AI country. It is a tenant paying rent. The host may own the building, but the operator owns the operation. In any conflict between the two, the operator prevails, because the operator can take the operation with it and the host cannot. Foreign data centre investment in Australia, structured the way it has been structured, has built a tenancy at scale, not a sovereign capability. Nor will the rent stay modest. Every landlord knows the strategy, whatever the asset: be as useful as you can until you become necessary, then charge what the tenant can bear. A country that has built its economy on a landlord's intelligence is a tenant that can bear a great deal.

For the remainder of this paper, when we use the word sovereign we mean Australian-owned and Australian-controlled at the operating layer. We will be explicit when we are talking about components.

3. The Speciation Event

The argument so far has been about ownership of a productive technology. The argument we now develop is about what the technology is, because the standard discourse has been treating it as if it were an unusually capable software product, which it is not. A modern large language model is, in a strict and economically consequential sense, the precipitate of human civilisation rendered queryable. The right biological analogue for what is happening as it spreads is not invention. It is speciation.

3.1 What an LLM Is

A large language model can be described as four significant parts.

The first is the corpus: as much of the world's recoverable human dialogue and recorded thought as the operator could collate. Books, scientific papers, encyclopaedias, code repositories, web pages, conversations, transcripts, court reports, parliamentary proceedings, manuals, novels, blog posts. The corpus is uneven, heavily English, heavily Western, heavily online, heavily 21st-century. But for its biases it is the largest single artefact of human communication ever assembled. No university, no national library, no individual scholar has ever held a fraction of it in working memory at once. The model has.

The second is the weighting: a trained neural network has billions or trillions of parameters that encode the statistical regularities of the corpus at every scale, from the next word in a sentence to the rhetorical arc of an argument across a thousand sentences. The weights are not facts. They are priors over what should come next given what has come before. The model has not memorised the corpus; it has learned to continue the corpus in any direction it might plausibly have gone.

The third is the framework: the transformer architecture together with the surrounding inference stack, tokeniser, context window, sampling, system prompt, retrieval, tools. The framework is what allows the trained weights to be exposed to a user's local situation and to produce a continuation conditioned on that situation. It is, in operational terms, the part that turns the corpus into an employee.

The fourth is the alignment: the post-training adjustments, reinforcement learning from human feedback, constitutional methods, preference optimisation, that shape the model to behave the way its operators want. Alignment is the layer where the model's outputs are filtered, refused, softened, redirected and prioritised according to instructions that did not come from the corpus. It is, in policy terms, the part that decides whose voice the model speaks in when it speaks.⁶

A model is these four parts working together. It is not a search engine, because search engines retrieve whereas the model synthesises. It is not a database, because databases store whereas the model summarises. It is not a tool, because tools wait to be wielded whereas the model acts on natural language without any technical step between the speaking and the doing. It is something genuinely new in the catalogue of economic primitives, and the analytical confusion of the current discourse is a direct consequence of pretending it is one of the older things, a better screwdriver. It is not. The companion paper, Part 1, grounds this claim formally: intelligence, whatever its substrate, is the recognition and compression of pattern in data, and the model is that operation applied to the recorded output of our species.

3.2 The Promethean Observation

A model trained on the collected dialogue and art of the human species is, as described before, in a real and not metaphorical sense: human history and observation rendered queryable. Every Wikipedia editor, every Stack Overflow contributor, every novelist, every scientist whose paper was scraped, every teacher whose tutorial was indexed has contributed a fraction of the substrate from which the model produces its responses. This substrate was not paid for. It was donated, in the original meaning of that word: given without expectation of return, often without awareness that the giving was occurring at all.

This is a Promethean observation. Humanity has built its own successor by donating its own voice as the substrate. The new system is not an alien intelligence trained on alien data; it is us; humanity's own cumulative recorded thought, lightly compressed, infinitely patient, and re-presentable on demand. A company that takes out a $30-per-month subscription to a frontier model is not hiring a tool. It is, for the price of a streaming service, hiring an entity that has read more of our species' written output than any individual living human has, and that can act on what it has read at the speed of inference. The arbitrage is between what such an entity is worth in use and what was paid for the substrate corpus to assemble it from. The first number is enormous whilst the second is close to zero.

Our 2018 paper measured substitution by the elasticity of substitution σ between intelligent machines and human labour. The framing was correct for the question that paper asked, but it understates what is now happening, because σ compares one machine against one human at one task. The current arbitrage is a different quantity: the cost of one seat against the cumulative cost of assembling, by human labour, an equivalent capability. That ratio is, on some tasks, four orders of magnitude! A research analyst at $200,000 a year cannot deliver, in scope and breadth, what a $360-a-year seat at a frontier model delivers, because the seat has read and integrated what the analyst has not even heard of, nor has the necessary lifetimes to read. The analyst is not competing against a machine. The analyst is competing against the accumulated writings of the human race.

3.3 The Fitness Differential

In adaptive speciation, a population diverges from its parent by acquiring an adaptation that lets it occupy a niche the parent cannot, and the divergence accumulates until the two can no longer interbreed, at which point they are separate species. The adaptation matters because it is the mechanism by which the new species outcompetes the parent in contested niches. Without an adaptation that confers a real fitness differential, there is no speciation; there is only variation.

The fitness differential in the present case is the cost-arbitrage previously described. The new species (by which we mean, at this stage, the humans, firms and institutions that have integrated AI augmentation into their working substrate) operates at a productive cost-to-output ratio that the disconnected parent species cannot match. The differential in those tasks where the technology is mature, is many orders of magnitude. In tasks where the technology is still maturing, it is doubling every few months. The compounding mismatch is what produces the speciation, and the speed at which it produces it is what makes the present moment qualitatively different from prior productivity revolutions.

It is also important to be clear about the direction of the differential. The new species is not better at every task; it is better at the tasks the parent species was previously paid to do. It has become good at what we are good at, whereas previously we built machines to do things we were bad at. The parent species retains capabilities the new species does not have: embodied skill, lived experience, social presence, moral judgement under genuine uncertainty, the long-running coherence of a life, of consciousness; but the economy does not pay for those capabilities at a rate that compensates for the loss of the paid ones. That list is not stable. Substitution to date has been cognitive, because the language models arrived first. Still to come is Robotics, which is the second wave of the same substitution, and it will strike embodied skills as well. If human participation is to be significantly preserved past that point, it will be preserved by legislation rather than by the market; that is the work of the participation mechanisms set out in §4.5. The fitness differential is in the contested economic niche, not in the full ecology of human flourishing. This distinction will matter again when we discuss what the economy is for, in §4.

3.4 Diffusion Failure and Reproductive Isolation

The Hayekian organisation of distributed search has two virtues. The first is allocative: whereby prices aggregate distributed knowledge so that individual agents can act on the aggregate without holding it. The second, which matters more for our purposes, is innovative-with-diffusion: many independent agents try many things, the successful ones are observed, copied and improved, and the gains of any one discovery propagate across the economy through patent expiry, employee mobility, reverse-engineering, trade-secret leakage, outright theft and the slow but reliable circulation of tacit knowledge between competitors. The theft channel is larger than the statute implies: in many domains it is hard to establish that a competitor is using a patented method at all, and where it can be established, a minor variation is often enough to raise the cost of enforcement above the value of the claim. The diffusion is not instantaneous (the patent system contemplated decades of delay) but it has been, for two centuries, the mechanism by which the gains of any one searcher's discovery become, in the limit, the property of the whole society. Society at large has benefited from the discoveries of its searchers; the searchers' priority was protected, imperfectly, for a time, then removed.

This new AI species breaks this diffusion mechanism. The discoveries embodied in a frontier model are not patentable inventions with an expiry date; they are weights, proprietary corpora, and the operational know-how that surrounds them. Patent law assumes a human inventor who can be compelled to disclose but AI-derived discoveries have no such inventor. Trade-secret law covers the weights without imposing any disclosure incentive; the trade secret is permanent. Employee mobility once carried tacit knowledge between firms; AI is replacing the employees whose mobility would have carried it. Reverse-engineering the analogue product from observation was effective; reverse-engineering a model from its API outputs,a method called distillation, is expensive and transparent to the models owners. Every channel through which the parent species diffused its discoveries has, for this new substrate, been disabled by its architecture and by the legal regimes that surround it; only theft survives, which is why the frontier laboratories now guard their weights with state-grade security rather than with lawyers. The nearest thing to a working leak is distillation as described before, the training of a new model on an existing model's outputs. The usage agreements forbid it, but it has happened anyway; the copying is done through the front door, at the metered rate, and a seller paid by the token for the very queries that copy it has shown little appetite for curtailing them. But distillation copies behaviour rather than the discoveries beneath it, and it is available only to a rival that can already train models at scale. It is a leak between members of the new species, not a channel back to the parent.

The consequence is something close to reproductive isolation in the biological sense. The two populations no longer mix their gene pool of ideas. The leader's lead compounds rather than decays; the laggard cannot wait for the patents to expire because the discoveries were never patented, and cannot hire the talent because the talent is being replaced, and cannot reverse-engineer because the artefact is opaque. The historical mechanism that rescued laggard nations, wait long enough, the textbooks are written, the experts move, the trade secrets leak, has been broken at the precise moment the laggards needed it most. Worse, the pace of discovery and implementation has now increased leading to multi-generational lagging.

Notice as well, how the new species has solved its own knowledge problem, because the solution shows the direction in which knowledge now flows. The weights are frozen at training time; the model knows nothing after its cutoff and nothing of any user's private world. The framework of §3.1 overcomes this at the point of use. The context windows are now large enough to injest whole libraries in a single prompt. Retrieval pipelines feed the model an organisation's own corpus as it works. Tool calls go further still: the model queries live databases, runs code, searches the web, and acts on what it finds, reaching out from the frozen weights into the operating present. Each of these mechanisms is an intake. Knowledge flows from the user's side into the deployment, conditions the model's output, and, as usage telemetry and training signal, accrues to the model's owner. Nothing flows the other way other than task utility. The model reads your documents; you never read its weights. This is also why reverse-engineering from API outputs is very difficult: what comes out is conditioned on what you just fed in, and tells you about your own context and the huge corpus of training data rather than about the model. The old economy's membranes leaked in every direction, and the leaking was the diffusion. The new membrane passes knowledge in one direction only.

The sovereignty implication follows mechanically. A nation without sovereign AI is not merely behind on a productivity index. It is, on the diffusion question, reproductively isolated from the species that is now generating the discoveries that define the future. It can rent access to the new species' output but cannot inherit it. The choice is to be one of the speciation events or to be the parent population it leaves behind.

This argument has a counter in the open-weight models (Llama, Mistral, DeepSeek, Qwen, Gemma), whose published weights restore some part of the broken diffusion. The counter is real but weaker than its proponents make it. These releases are in many cases the competitive weapons of trailing laboratories, not acts of generosity, and the incentive to publish evaporates as the publisher nears the front; open weights are, at best, a generation-old snapshot of the frontier. Weights without the surrounding operational know-how (the alignment data, evaluation harnesses, fine-tuning recipes and human capital that produce and use them) are an inert lump; the load-bearing fraction of the discovery lives in the operational practice, and this does not diffuse on the same timetable. The weights are also heavy. A model of frontier scale no longer runs on commodity hardware; serving it requires racks of state-of-the-art GPUs that are scarce, expensive and export-controlled, so the weights diffuse freely while the means of running them do not. And even where an open model is adequate to the task, a firm adopting under competitive pressure pays the trivial premium for the closed frontier, so demand and the data flywheel concentrate there. The diffusion has been partially restored, with a structural lag that places the open-weight inheritor permanently behind the closed-weight originator. That is enough to keep the diffusion-failure argument substantially intact.

There is a second counter: an aggregate of smaller models, working together as a coordinated team, can on many tasks match or outperform a single larger model. That counter is correct, and this paper builds on it as a way forward for Australia. It is the escape clause through which a middle power exits the isolation just described, and §10.5 constructs the programme upon it.

4. What the Economy Is For

The speciation argument describes LLMs as a new productive species that has now arisen, and that diffusion will not rescue the parents with their legacy means of production. To answer the question why this matters we have to address a deeper question. What is an economy for? The answer to this is largely assumed but not defined in the government position papers. The 2025 National AI Plan certainly does not answer this and to be fair, it is not alone. This question is in fact almost never asked at this level, because under the conditions of the prior productivity revolutions the answer was not contested. But it is now, as AI is significantly changing the conditions.

4.1 The Question Output-Maximisation Cannot Answer

Modern economics often claims that the main purpose of an economy is to maximize production. But that logic is flawed. Imagine an automated asteroid-mining machine in another solar system churning out resources with no humans involved. Measured purely by production, that machine is a "successful economy". This is obviously absurd. We don't want an economy that just produces stuff for no one. Maximizing output is not the true purpose of an economy; economists have only held onto this idea because, until now, it was never really challenged.

During the industrial age, we could pretend that maximizing output was the ultimate goal because production relied on human workers. Workers made goods, earned wages, and spent their paychecks, which drove even more production. Wealth naturally spread to the broader public through wages, taxes, and cheaper products. Because humans were essential to making things, maximizing production and improving human well-being naturally went hand in hand.

The rise of AI breaks this connection. Machines can now generate massive output without paying human wages. If people are cut out of the production loop, economic output can accelerate while worker incomes stagnate. This creates a country with a massive GDP but an impoverished population, a trend we are already beginning to see, and one that the deployment of AI at scale will dramatically accelerate.

In short: For two centuries, making more goods required employing more people, so maximizing production automatically boosted human well-being. AI breaks this link by producing goods without paying wages — meaning we can no longer assume that simply growing the economy benefits the people living in it.

4.2 The Commonwealth Frame

The answer we posit that survives the AI transition is the commonwealth frame. The economy exists to serve the commonwealth: the shared wealth and welfare of all members of the polity, including those not yet born. The commonwealth is wide, spanning every living member of the polity, and it is also long, reaching the members not yet born. Edmund Burke understood the nation as a partnership "between those who are living, those who are dead, and those who are to be born", and the members with the longest stake in the country are the ones who cannot yet speak in its deliberations.⁷ Members of the commonwealth hold their claim on the economy not by virtue of producing labour but by virtue of membership. The economy is the institutional means by which the commonwealth produces and distributes the material goods that allow its members to live. In this view, the economy is not an end in itself; it is an instrument of the commonwealth.

The commonwealth frame has the advantage that it survives the displacement of labour. A member of the commonwealth who has been displaced from productive employment is still a member of the commonwealth. The economy is still in service of their welfare. The mechanism by which their share of the economy reaches them may have to change (the wage channel was one such mechanism, not the only one and not a guaranteed one) but the claim does not change. The claim follows solely from membership, is not contingent on race, age, gender or employment.

The commonwealth frame is also, we observe, the implicit frame of most of the Australian institutions that have actually been built over the last century: Medicare, the age pension, public education, the ABC, the NBN, public infrastructure of every kind. None of these was justified on output-maximisation grounds. Each was borne on the proposition that members of the polity have a claim on the polity's productive output that is anterior to their position in any particular labour market. Everyone is entitled to a "fair go". The country has, in its actual political conduct, if not always in its rhetoric, and not perfectly, been operating on the commonwealth frame for a long time. The AI transition makes the frame explicit by removing the alternative.

4.3 The Wage Channel Was Contingent

The optimistic economic case for AI rests on a condition it usually leaves under the mantra of "free enterprise". The wage channel (the mechanism by which productivity gains in any sector flowed back to the population through employment in that sector or in the downstream sectors that the productivity gain enabled) it turns out, was not a feature of capitalism in general. It was a feature of capitalism under conditions where production required labour. Steam mechanisation displaced weavers but required mill workers, engine drivers, railwaymen, mechanics, supervisors and clerks. Electrification displaced one set of jobs and created another that was larger, because the new productive capacity needed people to build the wires, run the substations, install the appliances and bill the consumption. Each prior wave of productivity-enhancing technology produced more total employment than it destroyed, because the new productive capacity ran on human-required infrastructure.

The AI wave is the first wave in which the new productive capacity does not require humans to operate it in the same proportion. This is because the persons who intellectual labours it encapsulates are long gone, or else have given away their claim. A large language model serves a million users with a handful of operators. An autonomous haul fleet moves similar tonnage with the cabs empty. At the same time it increases utilisation and saves on maintenance. The mining case was safety first, equipment utilisation and maintenance second, wages third. The displacement does not care about the motive; the employment ratio moves the same way. It is also the kind of automation our framework welcomes: §4.5 exempts safety automation from its participation obligations. An automated freight or sorting network moves goods with a smaller fraction of the labour that the prior network demanded. The new productive capacity creates some employment, datacentre construction, model training, datacentre operations, robotic maintenance, regulatory oversight, but the ratio of new employment to old employment is, for the first time in the industrial era, substantially less than one. The construction jobs, the ones the National AI Plan celebrates most loudly, are also the least durable: they end when the building is finished, and what remains is the sliver that operates it. The wage channel that carried prior productivity gains back to the population is therefore shrinking. Whether what survives of the channel is sufficient to close the loop is an empirical question to which the answer increasingly appears to be no.

The same narrowing can be seen from the side of the wage itself. Wages have always rested on the scarcity of appropriate workers. Intelligent machines end that scarcity for a wide swathe of tasks: the new worker has read and absorbed every textbook in the world, works around the clock, and can be cloned as many times as there are tasks worth paying for. For the individual firm the substitution is a match made in heaven. For the economy the harm arrives at the aggregate level, because companies make their money selling produce to people with wages to spend. A firm that automates its own workforce saves money; an economy in which every firm does so is quietly dismantling its own customers.

4.4 The Membership Claim Survives

By affirming the citizen's claim on the economy to membership rather than employment, the commonwealth frame gives us the answer that the narrowing of the wage channel otherwise denies us. The claim is not merely to a share of the economy's output, as though the commonwealth's whole obligation were to keep its members fed. It is a claim to be able to engage with society in a societally valued manner. For two centuries the wage discharged both halves at once: the pay delivered the share, and the job delivered the engagement. AI substitution splits the two apart, and the split is why a cheque alone cannot discharge the claim. What discharges it has two movements: the commonwealth equips its citizens to engage, through education, language, sport, survival and the first aid courses that make a bystander useful; and it holds open the opportunities in which the engagement can occur.

The second half of the claim rests on hard numbers: losing work raises the risk of death by roughly 60 per cent, and prior ill-health does not explain the association.⁸ Unemployment is one of the few life events to which people never fully adapt: life satisfaction falls when work is lost and does not fully recover even when work returns.⁹ We are built to contribute where our contribution is seen and valued. Every human culture has centred its courtship on demonstrated competence, and any polity should hesitate before building an economy in which its young have nothing left to demonstrate. The work need not be back-breaking, and the machines are welcome to the drudgery and danger; what can not be automated away, is the need to matter, to take part, to define ourselves by engagement to society.

The channels that deliver the claim must therefore be, in the main, channels of participation: mechanisms that maintain engagement rather than merely fund it. §4.5 sets them out: a compute levy funding a negative payroll tax that tilts marginal hiring back toward humans; participation floors traded at the industry level, so that employment migrates toward the work least susceptible to automation; the right of every citizen to elect a human counterparty; exemptions for the automation that takes people out of danger; and in time the redistribution of work through a shorter week. But the claim would likely be best served by the deliberate expansion of the economy itself, because the larger the AI-enabled economy grows, the more edge cases it presents in which human participation remains the economic answer.

This proposal to expand invites a charge of circularity: the paper describes automation narrowing the wage channel, then proposes more automation to widen it. The circle comes apart on one distinction, which is where the machines go. Displacement is machines taking work people already do; the ratio of new employment to old is less than one, and the wage channel narrows. Expansion deploys machines on work no person has ever done, because no person could: the ore body under a desert that supports no town, aquaculture at a scale no crew could work, the irrigation of country that was never economic under human labour. The employment baseline there is zero, so the supervision, maintenance, logistics and exception handling that the machines require are added jobs, not remainders. The same machines carry the opposite sign. We oppose the first where it is left unmanaged; we propose the second at national scale.

The edge cases themselves are not permanent, because today's exception handling is tomorrow's training data. But a job justified by a right does not erode with capability, because it was never premised on the machine's inability. The citizen's right to escalate to a human arbiter, with a ceiling on the delay, obliges staffing in proportion to demand: the arbiter is there because the citizen is entitled to a human, and no model release automates away an entitlement. The worker's right not to work in isolation extends a duty that Australian workplace law already imposes for remote or isolated work;¹⁰ the automation era adds a new way of being alone, the lone technician in a facility of machines, and the same crewing principle answers it, on safety grounds and on human ones at once. Rights of this kind are not make-work. Make-work is justified from the supply side, by the need to employ someone; a right is exercised from the demand side, by a person who wanted a human and was entitled to one. As the capability justification for the edge cases fades, the rights hold open the work that matters, and erosion runs out of things to erode.

A public dividend on productive capacity that the commonwealth itself owns can supplement these channels. The longest-running such dividend in the world, Alaska's, has paid each resident an average of about twelve hundred US dollars a year across its four decades:¹¹ a big night out on the town, not a living. A dividend could only approach a living if the economy itself were massively expanded, which is the work of the expansion programme, not of the cheque writer. The dividend remains what it is, a return on common property, not the universal basic income against which §5.4 argues; no cheque discharges the claim on its own. The mechanism will not be the same for every citizen, and it will not arise automatically. It will be built, on purpose, by the polity that owns its productive infrastructure.

That last clause is the bridge to sovereignty. A polity cannot build a distribution mechanism over a productive capacity it does not own. A foreign-owned AI system serving the Australian market pays no commonwealth dividend and owes Australians no place in its operation. Company tax is the obvious rejoinder, and it is not sufficient: tax captures a sliver of declared profit, transfer pricing routinely shrinks the sliver, and a claim on profit is in any case not a claim on the engine. Without sovereign ownership, the commonwealth frame collapses into rhetoric. With it, the frame is the foundation on which a serious AI policy can be built.

4.5 The Participation Mechanisms

If machines will outperform humans on the metrics that markets reward, and if human participation is a necessity that markets will not spontaneously preserve, then the scales must be deliberately weighted. The critical constraint is efficiency. An intervention that preserves human participation at the cost of crippling the productive gains of automation defeats its own purpose, because the machine surplus is precisely what makes a managed transition possible. The thumb must be pressed on the scale, but it must be pressed precisely.

The first mechanism is a tax. We argue that if it is compute that competes against the human brain, then it is compute that should bear a tax: a levy on the machinery of substitution, collected on the compute hardware that comes into the country and on the output of foreign AI services sold into it, with offsetting credits for that which Australia exports, be it the same hardware repackaged or the outputs of that hardware. The levy is Pigovian in structure. Displacement carries a social cost, measured in the mortality and unrecovered life satisfaction documented in §4.4, and that cost is currently borne by the displaced and by the commonwealth rather than by the automating firm; the levy prices the externality at its source. The revenue is not hypothecated to welfare. It is directed to a single purpose: reducing the cost of employing a human being, through a reduction in payroll tax and, if warranted, its inversion. The pair does not attempt to make humans cheaper than machines in every domain; that would be neither possible nor desirable. It tilts the equation in the marginal cases, where the costs are close, and makes the human the economically rational hire across a broader range of work than an unmanaged market would produce.

The second mechanism answers the question where preserved employment should go? Which work should remain human is a discovery problem, and central planning has a well-documented record of failing discovery problems. So we advocate to let the market do what markets do best. Each industry carries a floor of human participation based on the current ANZSIC scales. This a minimum below which an enterprise may not fall. An enterprise that wishes to automate below its floor is not prohibited from doing so; but it must buy employment credits from an enterprise that stands above its own. The credits are tradeable and the price is set by supply and demand. This is a floor and trade mechanism, similar in concept to the cap-and-trade mechanism with carbon trading but with labour as the unit of account. Industries where human presence adds real value become natural sinks for employment, generate surplus credits, and profit from keeping people. Industries where the economics of substitution are overwhelming, pay for the displacement they cause, and the price of the credits enters their automation decisions. The credit price is a price signal in the full sense §5 develops: it aggregates, out of every enterprise's private knowledge of its own operations, information no planner could assemble, namely where human work still earns its keep. The market, not some government planner, determines where human labour is most productively preserved. Capitalism is thereby put to work finding where humans can work.

The third mechanism is the pair of rights §4.4 has already described. Any person dealing with an automated system may elect, at any point, to be transferred to a human being, and any person subject to an automated decision may put their case to a human arbitrator with the power to override it. The rights oblige staffing: a firm that automates its front line must keep a human fall-back, and the cost of the fall-back is part of the cost of automation. The obligation also improves the machines. Every escalation to a human is a cost, so the firm invests in automation that escalates less, and the quality of the automated service rises with it. The fall-back does not merely catch failures; its existence pressures the system to produce fewer of them. The constraint that such staffing be housed in Australia, prevents foreign automation from subverting those same channels. The same staffing is also the nation's manual redundancy, the retained knowledge of how the work is done without the machine; §9 develops why that redundancy grows more valuable every year.

The fourth is the exemption. Automation that demonstrably takes people out of danger, the driverless haul fleet of §4.3, autonomous hazard response, process control in dangerous environments, is exempt from the levy and the floor alike, by application. The objective of the framework is to preserve human participation in productive life, not human exposure to risk.

The fifth mechanism and perhaps the most significant, arrives last, when the others are no longer sufficient: this is the redistribution of work itself. If the total demand for human labour contracts even as the economy grows, the remaining work can be spread more widely through a shorter standard week. A six-day operating week run in two shifts doubles the number of people engaged in a workplace without reducing its productive hours, and each position sustains two households instead of one. The five-day week was itself won from an earlier wave of mechanisation. The negative payroll tax is what makes the deeper reduction affordable: when the cost of employing a human is subsidised, the marginal cost of splitting a role between two workers is manageable, and the social return is substantial.

The rates, the floors and the response times are parameters for legislation. The design is the argument: tax the substitution, spend the proceeds on human employment, let credit prices discover where people belong, guarantee the human counterparty, exempt the automation that saves lives, and share the work that remains. The mechanisms also become cheaper as the economy grows. A four-trillion-dollar economy needs a lighter levy than a two-trillion-dollar one to fund the same participation, and a larger economy generates more of the niches in which human judgement, presence or creativity is genuinely valued. That is why §4.4 put expansion first: the framework does not shrink the economy to protect people, it grows the economy through the machines and spends part of the growth on keeping people in the productive loop.

4.6 Parenthood Is Paid Work

One form of engagement the wage channel had never paid, is the one that produces the commonwealth itself. The commonwealth, we said, includes those not yet born. Parenthood is the work that delivers them: it grows them, gives them language, teaches them to walk and swim, and makes them capable of every other form of engagement this paper discusses. The economy has always depended on this work and has never paid for it, certainly not what it was due. It has been the largest free ride in economic history, taken mostly at the expense of women, and we have tolerated it because we never looked.

None of this says the work is without reward. Parents, and mothers above all, take from their children a reward that no wage measures, and most would tell you it is the best thing they ever did. But the reward does not pay a mortgage; the career it interrupts does. Having children still breaks career progression, and the compensatory mechanisms now in place, paid leave, superannuation on leave, the right to return, soften the break without removing it. The choice has been made cheaper; it is still a choice, and it still falls mostly on women.

Under the engagement claim the position is simple: parenthood is engagement with society in the most societally valued manner there is, and it should be paid as work. Not as welfare, and not as a per-child bounty, but as a wage for the job of raising a member of the commonwealth. The payment passes the test this paper sets for every other channel. It is participation, not withdrawal; no parent is idle, as any parent will confirm. And it is the one job automation can never hollow, because the human presence is not a component of the work; it is the work. A machine can mind a child. It cannot parent one. Every child's counterparty is their parents, and no statute is needed to make that right non-delegable; nature drafted it first.

The country needs the work done, and is not getting it done. Australia's fertility rate has fallen to 1.48 births per woman, a record low, against the 2.1 that holds a population steady; the last year at replacement was 1975.¹² Every developed economy tells the same story, and part of the cause is the same everywhere: parenthood competes with paid work and loses. If automation is thinning paid work regardless, the polity that owns its productive engine can convert the displacement into renewal. A country that pays its parents buys back its own future.

Allowing the population to fall is not necessarily a bad thing although it may be a sad thing. A machine-productive economy does not need a growing population to grow, and a smaller Australia that owns its engine can be richer per person than a larger one that does not. What we hold to is the choice: the couples who want children should not be priced out of having them, and that is what the parental wage is for. What has no defence is the current expedient: importing large numbers of foreign workers,¹³ on the cusp of technological unemployment, with no plan in place for the unemployed. The incentive is not hard to find. Migration at this scale keeps aggregate GDP growing while GDP per person falls, and it keeps house prices rising, and housing is the asset in which the most reliable voters hold their wealth. We suggest the programme has become, in part, a price support for land: paid for by the young in rents they cannot escape, and in the entry-level labour market where the new arrivals compete first, which is the same market the junior wall is already closing. That is either cynical electioneering or ignorance, and neither is a policy. Immigration can resume its place when the deliberate expansion of §6 creates work faster than the machines retire it.

There is one more reason to treat the work seriously, and it bears on the education system. We want every child exposed to the highest education the country can offer, and a child's primary point of exposure is their mother. The evidence is old and consistent: educate a woman and you educate her children.¹⁴ A country that pays parenthood as work should therefore also widen the educational opportunities of the women doing it, and make the wage compatible with study, so that raising a family and raising one's own education can run together rather than compete.

The design constraints can be stated briefly. The wage must be payable to either parent and stackable with other work, or it becomes an instrument for removing women from the workforce and will deserve the attack it receives. It must be distinguished from the family-payments apparatus that exists, which is framed as relief for hardship; this is framed as payment for work. And it is expensive at wage scale, which is a dependency rather than an objection: it is one more thing that only a massively expanded, machine-productive economy can afford, and one more reason to build one.

4.7 What This Means for "Every Australian Benefits"

The National AI Plan's framing, that AI will be deployed such that "every Australian benefits", is a teleological claim with no mechanism to fulfil it. The claim is correct in its ambition. It is correct in its identification of the commonwealth as the criterion. It is wrong in its assumption that the benefit will arrive without an apparatus to deliver it. The historical apparatus was the wage channel, and that historical apparatus is closing. The new apparatus must be built deliberately, and the policy document that proposes to ensure every Australian benefits has, on inspection, neither named the new apparatus nor admitted that one is required.

The apparatus is the one this section has set out: the participation mechanisms of §4.5, the parental wage of §4.6, and the public dividend and deliberate expansion of §4.4. Beneath all of them sits a structural point: the apparatus is buildable only by a commonwealth that owns the engine. Sovereign AI is therefore not merely an economic instrument. It is the institutional precondition for the commonwealth's continued capacity to honour the membership claim of its citizens, after the wage channel has narrowed to insufficiency.

5. The Hayekian Pivot

The classical defence of the market economy against the central-planning alternative is, at base, a defence of distributed search. Hayek's contribution in The Use of Knowledge in Society was the observation that the relevant economic knowledge (the millions of local facts about preferences, scarcities, opportunities, capacities) is not held by any one mind and cannot be aggregated by any central authority on any plausible timetable. The price system is a computational mechanism, evolved rather than designed, that aggregates this distributed knowledge through the actions of millions of agents each searching their local environment and acting on what they find. Central planning fails not because central planners are stupid but because the computation is intractable for a central planner of any human size.

This defence, in the successive forms it took from Smith to Hayek, was for two centuries the strongest case for capitalism. The twentieth-century empirical test against the Soviet alternative ran the largest economic experiment in history and the Hayekian side won it convincingly. The default of educated opinion since 1989 is that the matter is settled. We propose that it is not settled, because the technological assumption beneath the Hayekian argument has just changed.

5.1 The Two Virtues of Distributed Search

Distributed search delivers two distinct virtues. The first virtue is allocative: prices aggregate distributed knowledge so that agents can act on the aggregate without holding it. This is the virtue the textbooks emphasise. The second virtue is innovative-with-diffusion: many independent searchers try many independent things, the successful ones are observed, copied and improved, and the gains diffuse across the economy through patent expiry, employee mobility, reverse-engineering, trade-secret leakage and the slow but reliable circulation of tacit knowledge. The diffusion virtue was as important as the allocative virtue for the practical performance of capitalist economies, because it ensured that the gains of any single discovery eventually became the property of the whole society. The patent system contemplated a period of monopoly followed by release into the commons. The labour market carried tacit knowledge from firm to firm. The trade press circulated practice. The economy, in aggregate, learned.

We have already observed (§3.4) that the second virtue is broken for AI-derived discoveries. The discoveries do not diffuse on the timetable on which prior discoveries diffused, and may not diffuse at all on the relevant horizon. This is the structural fact that we now extend into a positive proposition about what humans must continue to do.

5.2 What Distributed Search Now Searches For

The classical Hayekian search was: distributed agents searching for better ways to do their work. The new Hayekian search is: distributed agents searching for better ways to deploy AI to do their work. The unit of search has moved up one level. The agent is no longer searching at the object level of the problem; the agent is searching the space of possible AI configurations applied to the problem. The advantage in the new economy belongs not to those with the deepest domain expertise alone but to those who can deploy AI most effectively against the domain. A B-grade lawyer with A-grade AI-deployment competence outperforms an A-grade lawyer without it; the same is true in medicine, in engineering, in administration, in marketing and in policy work.

This is not, as some commentators have suggested, the abolition of human ingenuity. It is, rather, the relocation of human ingenuity. The new search is just as cognitively demanding as the old one, arguably more so, because the search space is larger and the artefacts being orchestrated are more capable than any analogue tool. The search produces real discoveries, real novelty, real productivity, just as the old search did. The discoveries are about how to deploy the model, what to retrieve, what to fine-tune on, what to evaluate against, what to delegate and what to keep in human hands. These are not trivial questions and they will not be answered once and for all; they are the substance of the new economic activity.

5.3 What Society Is For

The strict reading of the AI-replaces-labour argument concludes that humans are no longer required for production. We disagree, and the disagreement turns on two features of the new search. The first is that the criterion of the search lives in the searchers. Hayek's distributed knowledge was never only facts about scarcity and technique; it included preferences, the millions of local judgements about what is worth doing and what counts as better. The meta-search over AI configurations is scored, in the end, by whether human situations improved, and that measure originates in the people living the situations. An economy of machines could optimise, but nothing in it can supply the ends. The ends are distributed across humans, and no central system, however capable, can hold them on the humans' behalf. The second feature is the knowledge itself. A frontier model is trained on the recorded past and deployed into millions of working situations, each idiosyncratic, each rapidly changing, each observed first and best by the people doing the work. For any horizon this paper's programme addresses, those people supply the ground truth: they know which configuration worked, they catch the failures, and they correct the model where its training has not been. We do not claim this second condition is permanent; §3.3 warned that the list of retained human capabilities is not stable. We claim it holds now, and for long enough to matter.

This means humans are still required as searchers in the new economy. Their role changes, they no longer do the object-level work in the way they once did, but the role does not vanish. The new economy still runs on distributed human ingenuity, applied to a different substrate. A market with no humans in the loop has nothing left to clear: prices aggregate wants, and the wants are ours. And where the requirement thins, the polity should hold the place open on purpose, building its institutions so that the search runs through its citizens; that choice, and why any UBI proposal forecloses it, are the subject of §5.4.

There is also a category of work that does not migrate to the machines at all, because in this category the human is not the means of production, the human is the product. A model can generate a sermon, and it may be a fine sermon; it is still not ministry. The parishioner does not come for the words, they come to be witnessed by a fellow mortal who has also sat with grief and doubt. The same holds for the mother raising a child, for the nurse at the bedside for whom some irreducible part of the care is the caring, for the counsellor, the celebrant, the arbitrator who hears an appeal, and the opponent on the sporting field. Cars have outrun sprinters for a century, yet the hundred-metre final still fills the stadium; and where we do buy tickets to watch machines race, from Formula 1 to battlebots, we are watching the humans contending through them; when the winning machine crosses the line, the camera turns to the team that built it. The common thread is empathy, and empathy does not automate. A machine can produce its words and gestures, word-perfect and tireless, but empathy is received as empathy only from a being that can actually feel with you. These are the sectors where human presence, judgement or relationship is the service, and child-rearing is the most consequential of them (§4.6); we can now say why the advantage is durable. Consciousness, which has no defined economic value, is in these roles the commodity itself: what is being purchased is the fellow-feeling of a conscious being. These roles are not a residue the machines have not yet reached; a machine cannot occupy them by construction, because a machine performing them changes what they are.

There is a second independent reason to keep humans in the search. Society exists for the well-being of its members, and part of that well-being is membership itself. Humans are built to contribute to the group that sustains them, and they sicken when they cannot. A citizen maintained by society but given no part in it has not been provided for, they have been excluded. The economy must keep humans in the search because it needs them. Society must keep humans in it because they need it.

There is a plainer name for an entity kept by society but given no authority within it. It is what we call a pet. The mark is not whether a person contributes, since with few exceptions we would all wish to, but whether they hold any authority of their own or hold none. A pet holds none. We decide what it eats, where it sleeps, and how it ends. It is kept at our pleasure, because it is wanted, and only for as long as it is wanted.

The pet economy is large and it is genuinely affectionate, and it runs on no contribution from the animals at all, beyond the affection we choose to read in them. Its keep is a kindness rather than a claim, and a kindness can be withdrawn. But three things follow about pets, and each one should trouble us.

The pet has no say. We decide what is best for our pets. They can only tell us how they feel. They put no knowledge into the decisions that rule their lives, only feeling, and we read the feeling for them. This is central planning at the size of a household, and it works there for the reason it failed at the size of a nation. A dog's world is small enough to plan.

The pet's place rests on our empathy, and our empathy is uneven. It is real enough. We will nurse a hurt bird and let it go, and take nothing back but the joy of seeing it fly away. But the same feeling that pampers the dog runs the battery shed, steps on the ant, and kills the scorpion on sight. Where an animal sits is not fixed by what it is owed. It is fixed by how much it moves us, which is a matter of its scale, its face, its likeness to us. A being kept by empathy holds its place at the mercy of a feeling it cannot earn and cannot command.

And the keep lasts only whilst it is not a burden. The pet is kept whilst it is wanted. When it is old, or sick, or too dear to run, and the affection no longer covers the cost, it goes to the vet. It does not see it coming, because comfort was never safety.

We are not that animal. A dog in a warm house is content. A man in a warm house is restless. I write this from a cushioned chair, and I am glad of it, and tomorrow I will climb a (small) mountain and come back to the same chair. The chair is not the enemy of the mountain. The ache in the legs is the reward, and the chair is sweeter for it. We are built to choose the hard thing and to be paid for choosing it. Comfort is the floor we stand on, not the life we stand up to live.

This is why provision is not enough, even when it is generous. A society that provides for its members and asks nothing of them has given them the warm house and taken away the mountain. It has made them pets, and pets of the kindest sort, and it has starved them all the same. What a person needs above the floor is a climb. Not one forced on them, but one they may choose, towards a betterment they can see, with a reward at the top that the society actually pays. Our society already does this. The market, the trade, the laughter among friends at the end of shift, all standing earned by good work, these are a society paying its members for the hard thing freely done. Take the payment away and the striving does not stop, it just stops counting. The effort is spent in a sealed yard, felt by the one spending it and worth nothing beyond the fence.

And we should be honest about the house we are actually building. It is not warm throughout. The heat is pooling in a room or two, where the machines are owned, and the rest of the house is going cold. The old economy ran the warmth out to every room through wages. That channel is narrowing, so the warmth now collects where the engine sits and stays there. A cold house with one or two warm rooms is comfort for the few and a draught for everyone else, and the few decide who is let in to warm their hands. This is the pet arrangement with most of the pets left out in the cold.

A cold house with warm rooms is not new, and by itself it is not the thing we fear. Every society has had warm rooms and cold ones. What made some of them bearable was a route between the two. In a country with egalitarian principles there was always a way through, by work or by luck, from the cold rooms to the warm ones, and the way was open to anyone. What we abhor is caste, where the room you are born in is the room you die in, and your address is fixed before you can walk. The danger of the new economy is not that it has warm and cold rooms, but that it closes the route between them. When the machines are owned by the few and the climb no longer pays, the cold rooms stop being a place you pass through and become a place you are from. That is caste, rebuilt around who owns the engine.

Owning the engine is how a nation keeps the house open. Held in common, it heats the cold rooms, through the dividend a commonwealth can pay on what it owns. It keeps the climb paying, so the route between the rooms stays open and no address is fixed for life. Renting a place by someone else's fire does neither.

5.4 Why UBI Forecloses the Mechanism

This is the structural argument against Universal Basic Income, and it complements the psychological one. The psychological argument is that humans evolved to be productive in the eyes of their kin and community; provision without participation is psychologically corrosive; the country that pays its citizens to be idle pays them to be unwell. §4.4 set out the human evidence, in the mortality and the unrecovered life satisfaction of the unemployed. The clearest experimental demonstration was run on another species.

Between 1968 and 1973 the American ethologist John B. Calhoun built what he called a mouse utopia at the U.S. National Institute of Mental Health. The enclosure, Universe 25, provided unlimited food and water, perfect climate, freedom from predation, and attentive veterinary care; every material want was met in perpetuity, which is UBI as a mouse would receive it. The population grew rapidly, peaked, and collapsed. Breeding had effectively ceased by day 600; by day 920 the colony was functionally extinct. What matters is not the collapse but the sequence of social pathologies that preceded it. Dominant males abandoned territorial defence. Females grew aggressive and neglected their young. And a cohort emerged that Calhoun named the "beautiful ones": mice that withdrew entirely from social life, that neither fought nor mated nor contended for status, and that confined themselves to eating, sleeping and obsessive self-grooming. They were physically pristine and behaviourally extinct. Calhoun called the aggregate phenomenon a behavioural sink: the breakdown of social function in the presence of unchallenged material abundance.¹⁵ Caution is required in extrapolating from mice to humans; their social architecture is far simpler, and their deprivation was spatial as much as purposive. But the pattern is suggestive, and the developed societies furthest along the abundance curve are already producing recognisable analogues: falling birth rates, weakening cohesion, and the beautiful ones we now call "influencers", grooming themselves for their followers. Provision without economic participation did not produce a leisured utopia in Universe 25. It produced a population that no longer did anything, and then no longer was.

The Hayekian argument adds an economic leg. UBI withdraws the citizen from the distributed search. It pays the citizen to not search, on the assumption that the searching, or contribution, is no longer required. The assumption is wrong. The searching is required, more than ever, because the substrate of the search is now far more powerful than it has ever been, and the marginal value of competent meta-search is correspondingly enormous. The country that withdraws its citizens from this search degrades the search itself. UBI treats human involvement as an inefficiency to be engineered out, but in the Hayekian mechanism the human is the working part.

The right policy response to AI substitution is therefore not to disengage the population from production by transfer payment. It is to augment the population so that they remain productive in the new search. Augmentation here means more than access to the machines. It means better health, better education, real opportunity and real engagement, the endowments a country provides when it wants its people capable. Universal schooling was the augmentation of the industrial economy and public health the augmentation of the last century, whilst sovereign AI is the augmentation of this one. Each is a transfer of capability rather than a transfer of consumption, and that is the difference between augmentation and UBI. The party's anti-UBI position and its pro-sovereign-AI position are not, on examination, two positions; they are one. The country must keep its citizens in the search, and the only way to do so is to provide them with the augmentation that makes them competitive in it. The alternative, UBI without augmentation, is the welfare-state equivalent of telling the population to sit on couches and watch sports and soap operas because the economy no longer needs them. The country may, in this scenario, still be wealthy in aggregate, but its citizens will suffer. We are not designed for the world that UBI would have us inhabit.

The objection will come that this is Luddism, and this party, which named its founding paper for the Luddites, accepts that, but not as a criticism. The Luddites were not enemies of the machine but skilled tradesmen who broke only the frames being used, in their own words, for work "hurtful to Commonality", meaning machinery deployed in ways that expelled them from the productive life of their communities. §1 argued that they were right on the economic axis, whilst the present section adds that they were right on the human one: a technology deployed so as to remove people from participation takes from them something they cannot live well without. UBI is the Luddite error inverted. The Luddites would keep the people and break the machines, UBI keeps the machines and breaks the people, both accept that the two cannot work together. That premise is false. The third course is to keep the machines and keep the people, by giving the people the machines. That course is augmentation. It was not available to the men of 1812, but it is available to us.

5.5 The Sovereign Implication

The sovereign implication follows. If the new Hayekian search runs over AI configurations, the substrate of the search is the AI itself. The party that owns the substrate captures the meta-discoveries that the searchers make on it. Every Australian organisation deploying foreign AI is, in effect, performing the new Hayekian search on a foreign-owned substrate; the discoveries, about deployment patterns, fine-tuning recipes, operational workflows, retrieval strategies, flow back to the foreign provider as usage telemetry, fine-tuning signal, and operational data. The Australian searchers do the work; the foreign owner gets the discoveries. The diffusion failure of §3.4 is reinstated at the meta-layer: the Hayekian search is occurring, but its outputs are not diffusing back to the Australian commonwealth.

Sovereign AI inverts this. When the substrate is Australian-owned, the meta-discoveries diffuse back to the Australian commonwealth. The distributed search that drives the new economy operates on a substrate that the polity owns, with the consequence that the gains of the search remain in the polity. The Hayekian peace of 1989 is preserved at the new level, by the simple expedient of ensuring that the substrate of the new distributed search is owned by the commonwealth in which the searchers live.

6. The Australian Opportunity

The argument so far has been general. The remainder of the theoretical foundation is specific to Australia, and the reader who has followed the argument to this point will have noticed that it has been mostly defensive. We have argued that without sovereign AI Australia is reproductively isolated from the new species, the commonwealth claim is unenforceable, the new Hayekian search is conducted on foreign substrate and the gains accrue offshore. These are real concerns and they are sufficient on their own to motivate the policy. They are not, however, the whole case. Australia has positive reasons to build sovereign AI that have nothing to do with damage limitation. We discuss those here.

6.1 The Continental Constraint

The defining geographical fact about Australia is that it is the country with the most territory per citizen of any developed nation. Twenty-eight million people hold a continent of 7.7 million square kilometres, mineral-rich, agriculturally productive, environmentally significant, strategically positioned, and for most of its history under-developed against its potential. Every prior generation of Australian policy has wrestled with this fact. The 19th-century answer was massive subsidised migration. The 20th-century answer was "populate or perish", the post-war Calwell programme of large-scale European migration, followed by the late-century opening to Asian migration. Each answer treated the constraint as a labour-supply problem: the country had more land and more resources than its workforce could exploit; the workforce had to be enlarged by importation; the enlarged workforce would then exploit the resources.

The constraint is real. The historical answer has had real costs and real benefits, and it remains, in our view, partly the right answer for the coming century, although on a smaller scale and with greater intentionality than the recent past has practised. But the historical answer is no longer the only answer available. For the first time in the country's history, the continental constraint admits of a technological solution as well as a demographic one. AI and robotics are the technology that allow a small population to develop, manage, utilise and defend a vast territory.

6.2 Robotics as the Multiplier That Resolves It

The 2018 paper observed that mining was an early automation target because of legitimate safety considerations. The observation generalises. The work that the Australian continent presents is overwhelmingly the kind of work that AI and robotics multiply most: mining at scale across remote sites, broadacre and pastoral agriculture across the inland, water management across the Murray-Darling and the inland aquifers, environmental monitoring across the continental coastlines and the Antarctic adjacency, energy generation across the desert solar resource, freight and logistics across continental distances, and remote-area service delivery (health, education, justice) into populations that the cities cannot easily reach. These are the workloads in which AI and robotics deliver the highest leverage, and they are the workfaces in which Australia has, by accident of geography, an unusual concentration.

Most developed nations are population-saturated against their territory. The marginal productivity of automation in a saturated economy is real but bounded: the labour that automation frees can find other employment in the dense urban economy, but the absolute increase in productive capacity is modest because the labour was already deployed against the available resources. Australia is the opposite case. We are population-poor against our resources. The marginal productivity of automation here is not bounded by the substitution of existing labour; it is unleashed by the unlocking of productive activity that was never economically viable under a labour-only model. The fully autonomous remote mine becomes economic. The robotic dryland agricultural property of a million hectares becomes economic. The continuously-monitored Murray-Darling system becomes economic. The continental autonomous freight network becomes economic. None of these is a small adjustment of an existing operation. Each is a new productive activity that could not have existed at all without the technology.

This is the country whose existing condition is most directly transformed by the new technology. "Populate or perish" was the response of a country that needed people to develop its territory. The slogan can now be retired. The new slogan should be: automate and flourish.

6.3 Defence by Development

The continental constraint has always also been a strategic vulnerability. A vast lightly-populated territory in a region of much larger populations is, in the long run, a tempting target. The historical Australian answer has been alliance, first the British Empire, then the United States, more recently the AUKUS arrangement, supplemented by a modest standing defence force and the principle that the country is too far away and too hard to take to be worth taking. The principle has held for two centuries but it is not unconditional. The cost of "too hard to take" depends on what the target is worth, and the resource value of the Australian continent will rise during the AI transition, not fall, because the resources required by AI infrastructure (minerals, energy, water, benign climate, stable ground) are precisely those Australia has in abundance.

Visible economic activity, distributed infrastructure, deployed robotic systems and exploited resources are all de facto sovereign presence. A continent that is being actively developed is harder to dispossess than a continent that is being held passively. The defence-by-development case is therefore not separate from the sovereign-AI case; the two are the same case viewed from different angles. The technology that lets a small population develop a vast continent is the technology that lets a small population hold a vast continent, by making the continent expensive to take from a country that has visibly committed to using it.

6.4 A Note from the Author

We have written most of this paper, as we wrote the 2018 paper, in the first-person plural that is appropriate to a party document. We make one exception here, because the argument of this section is one to which I have a direct professional relationship and the reader is entitled to know the basis on which it is being made. I have spent near twenty years of my career building automation systems for the Australian mining industry. The capabilities described in this section are not abstract claims of what might one day be possible. They are descriptions of work I have done, or watched colleagues do, on real sites, with real economic outcomes. The Australian technical capability to build the sovereign AI substrate that this paper recommends is not an aspiration. It exists. It is dispersed across an industry sector that is already the country's largest by export value, an academic sector that is competent if under-resourced, and a research community whose better members have for years been recruited overseas because the work to retain them at home has not been deliberately constructed.

The talent question, addressed at greater length in §12, is therefore not a question of whether Australia has the people. It is a question of whether the country has the structure of work in which to deploy them. Sovereign AI, deliberately built, is that structure.

7. The Stack

The remainder of the paper sets out, layer by layer, what the sovereign AI programme must actually consist of. We adopt the engineering convention of treating artificial intelligence as a stack of dependent layers, each of which must be ownable for the layer above it to be ownable in any operational sense. The layers, in dependency order, are:

The stack is dependent in the order given. A nation cannot have sovereign models without sovereign compute. It cannot have sovereign compute without sovereign energy. It cannot have sovereign anything without sovereign talent, because every layer above requires people to deploy, install and operate it. It cannot have sovereign alignment without sovereign governance, because alignment is the embedding of policy in code and someone must decide the policy. And it cannot have sovereign data without legal authority over its own information environment, which is in turn a property of governance.

Each layer admits a different proper mix of provision. We distinguish three categories of good. A capitalist private good is excludable and rival, supplied by profit-seeking firms, the proprietary frontier model is the type. A capitalist public good is supplied by capitalist actors but is non-rival or non-excludable in practice, the open-weight model and the open-source compute stack are the type. A statist public good is provided by the sovereign for universal access, Medicare, the ABC and the Bureau of Meteorology are the type. The argument of the present paper is that the existing National AI Plan errs by capitalist-private default at every layer; the sovereign programme requires a different and deliberately mixed division of labour, which we summarise as follows.

The pattern that emerges is clear. The bottom and top of the stack, Energy and Alignment, must be statist: the first because it is the physical bottleneck on which everything else depends, the second because it is the embedding of national values in the instruction set of the machine. The middle of the stack (Compute, Model, Data, Talent) is best provisioned by a deliberate three-way mix in which capitalist private firms do what they do well, capitalist public goods supply the cost-free common substrate, and statist public goods guarantee the universal access that the market would otherwise withhold. Governance is purely statist. The two failure modes are symmetric: a wholly capitalist-private stack delivers efficient drift into foreign control; a wholly statist stack would extinguish the Hayekian search of §5 on which discovery depends. The sovereign mix preserves both, augmentation by utility, search by market, defence by ownership.

We treat each layer in turn. The reader who is impatient with the engineering detail may proceed directly to §15, where the layers are summarised into a single integrated programme. The reader who suspects we have substituted slogans for detail is invited to inspect the next eight sections.

8. The Energy Layer

The 2025 National AI Plan acknowledges that the Australian National Electricity Market consumed approximately 4 TWh of data centre load in 2024, and that this is projected by the Australian Energy Market Operator to triple by 2030.¹⁶ It does not acknowledge that the demand curve for AI compute is, at the time of writing, accelerating faster than the AEMO projection assumes, nor that the projection is itself a function of investment decisions still being announced. The plan's own infrastructure projections are already obsolete by the time the document is published.

Energy is the foundation of the AI stack because energy is what the stack consumes. A frontier-model training run requires tens of megawatts continuous for weeks or months. Aggregate inference, served at population scale, exceeds even this; the major providers' serving footprints are estimated at multiple gigawatts and rising.¹⁷ The compute, in physical terms, is the conversion of electricity into floating-point operations into language. The bottleneck is, and will remain, the electricity.

This has two implications for sovereignty.

The first is that energy is the bottleneck. Any nation that cannot generate the electricity to operate the compute will not have the compute, regardless of who owns the buildings, the chips or the operators. A nation that owns its compute but imports its electricity, directly or in the form of fuel, has merely relocated the dependency upstream. We have just seen the mechanism at work. One foreign country's military adventurism in another recently disrupted the global oil supply, and the disruption arrived in Australia within weeks, in the price of everything that moved by road. Australia imports around ninety per cent of its refined fuel and holds, onshore, weeks rather than months of diesel. A war we had no part in set the price of our freight. The same exposure, transposed to the electricity beneath the AI stack, would set the price, and the availability, of our cognition.

Imported energy creates a single point of failure upstream of the entire AI stack. The energy layer is not sovereign if the fuel is not sovereign.

The second is that the variable renewable sources currently being deployed at scale across the Australian grid, predominantly solar and wind, are an inherently weak foundation for a load that wants to run constantly. AI training in particular is a workload that cannot easily be timed to coincide with solar output. A training run interrupted by a cloudy week, or because the wind does not blow or blows too hard, does not fail outright, because modern runs checkpoint their state, but every interruption idles hundreds of millions of dollars of hardware and stretches the schedule, and a cluster that cannot promise firm power will not win the workload in the first place. Inference workloads are slightly more flexible, but user-experience expectations are constant availability; latency to first token is a competitive variable, and downtime in a hot market is unforgivable. Renewables can supply a meaningful fraction of the AI energy load through firming, storage and oversizing, but they cannot be the whole answer at the scale required, and pretending otherwise is the kind of optimistic non-planning that characterises the current document set as a whole.

The demand side is also larger than the stack. The deliberate expansion §4.4 and §6, the expansion of agriculture and mining and the onshore processing of what we now ship away raw, is a significant energy programme before it is anything else. Automation will require energy beyong compute as they draw power in proportion to the work; irrigation, desalination, electrified haulage and ore processing are all electricity by another name. If the economy is to grow large enough to generate the edge cases in which human participation remains the economic answer, the energy base must grow ahead of it. It is not just the compute that must be powered. It is the expansion itself.

The sovereign answer, we have argued elsewhere is a nuclear baseload. The choice among nuclear technologies, conventional pressurised-water reactors, small modular reactors, or thorium molten-salt designs, is a question of engineering and commercial readiness rather than of principle. One legal impediment remains: federal law currently prohibits nuclear power generation (the EPBC Act 1999 and the ARPANS Act 1998), and several states maintain bans of their own. The moratorium is not a formality on paper; it describes the country as it stands. Australia operates exactly one reactor, the 20-megawatt OPAL research reactor at Lucas Heights, and it makes medical isotopes, not electricity.¹⁸ The exception now on the books proves the ban is elective: to receive nuclear-powered submarines under AUKUS, Parliament has already passed the Australian Naval Nuclear Power Safety Act 2024, standing up a dedicated regulator and carving naval propulsion out of the civil prohibition. The Commonwealth's settled position, in statute, is that reactors are acceptable in Australia when strategy demands them; if the boats arrive, we will be the only nation on Earth that bans ashore what it operates at sea. The instrument that carved out the Navy can carve out the grid. Repeal of those prohibitions, at least for thorium, is therefore the first legislative act of the energy programme, and the sequencing of §15 begins there.

Australia has already written a footnote to this moratorium. Through the 2000s the country explored hot-rock geothermal power in South Australia's Cooper Basin, where the granites of the Big Lake Suite run hot because the uranium, thorium and potassium within them are decaying; the pilot turbine at Habanero ran a megawatt on that heat in 2013 before the venture was abandoned as uneconomic.¹⁹ Geothermal of this kind is nuclear heat in all but name, with the pile supplied by geology and the containment by four kilometres of rock. No statute forbade it, and no objector noticed. The prohibition, it turns out, is not on nuclear heat; it is on the word.

The rest of the world is not waiting on the debate. On 1 July 2026, at a test site in the Utah desert, the American startup Valar Atomics powered an Nvidia AI computer live on stage from its Ward 250 microreactor, a helium-cooled, TRISO-fuelled unit that had first gone critical thirteen days earlier. The demonstration itself was theatre, some hundred kilowatts of thermal output driving a desktop machine, but the programme behind it is not. Ward 250 is one of four microreactors to reach criticality within weeks of one another under the US Department of Energy's accelerated Reactor Pilot Program, and Valar and Nvidia have announced a study of a thirty-megawatt closed-loop AI factory on the same site, designed to consume no local water.²⁰ The direction is unmistakable: the reactor is becoming a data-centre component, sized to the load and shipped to the site on a military transport. Australia's moratorium does not merely forbid a power industry; it forbids a module the AI stack is beginning to ship with.

Nor is this an American eccentricity. Every major US hyperscaler has now signed nuclear supply for its AI estate, on commitments exceeding fifty billion dollars, including the restart of a Three Mile Island reactor for Microsoft's exclusive use. Britain is pairing its AI Growth Zones with small modular reactors under a standing AI Energy Council; France has offered a dedicated gigawatt of its nuclear fleet to AI investors; Japan is financing small reactors abroad with data centres co-located at the plant; and China expects Linglong One, the world's first commercial onshore small modular reactor, to enter service in 2026.²¹ Among the nations serious about AI, the pairing of reactor and data centre is becoming standard practice. Australia has forbidden itself the pairing before examining it.

We urge that Australia explore the thorium molten-salt path seriously, alongside the certified alternatives. The country's position is uniquely favourable: it holds some of the largest known thorium reserves on Earth, has stable desert geology suitable for siting, and retains, in its uranium mining and CSIRO research programmes, sufficient residual technical capability to begin. The case for thorium specifically, as against conventional uranium, is fourfold: the fuel is abundant in Australia, the waste profile is more manageable, the fuel it breeds makes a very poor bomb, which is why no weapons programme has ever fielded one built on it, and the reactor architecture is well suited to the load-following profile that data centre operation demands. China's Wuwei thorium molten-salt reactor achieved criticality in 2023, demonstrating that the technology is no longer theoretical.²² It has not, at the time of writing, completed the formalities for commercial certification, but is a contender for serious exploration rather than as the load-bearing assumption. We should work with China.

9. The Compute Layer

The compute layer is the physical machinery on which models are trained and served. It consists of accelerator chips (predominantly GPUs at present, increasingly specialised AI accelerators), the server boards and racks that house them, the high-bandwidth networking that connects them into clusters, the cooling and power infrastructure that keeps them within thermal limits, and the buildings, usually, but not necessarily, purpose-built data centres, within which all of this is installed and operated.

This is the layer at which the sourcing-versus-ownership distinction matters most, because the components are predominantly foreign while the operation can in principle be wholly domestic.

The components in any modern training cluster originate from a small number of suppliers. The accelerators are designed and largely manufactured by a handful of firms in the United States, South Korea and Taiwan; the networking is dominated by a similar handful; the high-bandwidth memory comes from two or three Korean and American suppliers. Australia cannot, in any relevant timeframe, manufacture the leading-edge components. This is the same condition that applies to commercial aviation, advanced pharmaceuticals, scientific instrumentation and a great many other technologies in which Australia operates as a sophisticated end-user rather than as an upstream manufacturer. The component dependency is real but it is the normal condition of a medium-sized advanced economy. It is also a more favourable dependency than it first appears, because the components are bought, not rented. A GPU imported and installed is owned outright: it owes no further payment to its maker, and the market that produced it has made it affordable at every scale from a national cluster to a regional workshop. The dependency at this layer is a purchase-order dependency, not a tenancy.

The purchase has a second property that the current debate misses: it is a consumable. An accelerator dies of obsolescence, not of wear. The vendors ship a new generation roughly every year, performance per watt doubles every two, and the operators write their fleets off over three to five years.²³ A data centre is not a dam. The building endures, but its contents are closer to a fuel depot than to capital infrastructure, and the capital evaporates on a five-year clock whatever anyone does. Three things follow.

First, the strategic stockpile described below is a flow, not a stock. Sovereign compute cannot be bought once; it must be a standing procurement relationship, refreshed each generation. The same clock is forgiving: a country behind on installed base is never more than one buying cycle behind, because the whole layer resets itself every five years.

Second, the layers of the stack have very different half-lives, and the National AI Plan has them inverted. Compute lives three to five years. Models are never worn out, only overtaken, and they copy freely. Data is permanent; a seismic survey shot in 1965 is still an asset today, and nothing a rival builds can obsolete it. Energy infrastructure takes ten to twenty years to build and lasts forty to eighty. The plan celebrates a hundred billion dollars of the most perishable layer as sovereign capability, and has no programme for the two layers that endure. The durable strategic assets are the energy beneath the stack and the data within it. The chips were always going to be replaced.

Third, while the hardware depreciates, the dependency appreciates. Today the loss of a data centre is an inconvenience, because the people who used to do the work still know how; when a software fault grounded the world's airlines in 2024, staff wrote boarding passes by hand.²⁴ But every year of substitution reduces the system's ability to apply a human fall-back over failure. The junior wall means the next generation never learns the work at all; the manual capability is not retired, it simply fails to be replaced. The void is invisible while the machines run, and the loss of a data centre is what reveals it: industries hollowed out by substitution, discovered at the outage. This is why the dispersal of §10.7 is an escalating obligation rather than a present convenience, and it is a second, harder reason for the human fall-back of §4.5: the staffed escalation path is also the nation's manual redundancy, the capability that lets the country reboot. Navies re-teach celestial navigation because satellite navigation is a wartime casualty in waiting;²⁵ an economy should keep its people current in the work for the same reason.

The operation, however, is a different matter. A datacentre, once commissioned, is a piece of physical infrastructure that sits within Australian jurisdiction, employs Australian operators, contracts with Australian customers, and is governed by Australian law. There is no technical reason that the operator must be foreign. The reason in practice is twofold. First, the hyperscalers have first-mover and scale advantages that make it cheap and easy for them to expand into Australia, and hard for any domestic alternative to compete on unit cost. Second, the federal procurement system has, by default, treated AWS, Microsoft and Google as preferred suppliers for government workloads, with the result that the most reliable customer in the economy has subsidised the foreign tenancy rather than the domestic landlord.

The result, as our 2018 paper foresaw on a slightly different axis, is that the operators become the economy. They hold the customer data, they hold the operational telemetry, they hold the commercial relationships and they hold the trust. When sovereignty considerations are raised, they offer the courteous response of "Australian regions", datacentres physically located in Sydney or Melbourne, operated under local subsidiaries, advertised as compliant with Australian data residency requirements. This is a useful marketing position but it is not a sovereign position. The Australian region is a building owned by a foreign company, run by a foreign company, with software stacks engineered by a foreign company, root-credentialled by a foreign company, and answerable, in the last analysis, to the legal system of the foreign company's home jurisdiction. The 2018 USA CLOUD Act,²⁶ which compels US technology providers to surrender data held overseas on receipt of a US warrant, makes the situation explicit: an "Australian region" of an American provider is, for the purposes of US law, an extension of American territory. The fiction of data residency does not survive contact with the legislation.

The sovereign compute programme therefore has two strands. The first is operator sovereignty: at least one and ideally several Australian-owned hyperscale operators, with sufficient capacity to absorb federal government workloads, state government workloads, regulated industry workloads (banking, health, defence-adjacent), and any private-sector customers who choose them. The components inside the racks may be of foreign manufacture; the operator must be Australian. The second is strategic stockpile and substitution capacity: a deliberate national programme of stockpiling sufficient leading-edge accelerator inventory to outlast plausible export-control or supply-shock scenarios and to replace in-service failures, together with active substitution research (alternative architectures, alternative process nodes, refurbishment) to extend the useful life of existing stock. The current plan addresses neither. A third strand, smaller in scale and distributed by design, is the sovereign neocloud tier and the compute harvest that feeds it; we treat both with the model layer, in §10.5 through §10.7.

The Federal Government's own AI compute requirements should be the first customer of the sovereign operator programme. Not by mandate (mandate breeds inefficiency) but by simple procurement preference, transparent and disclosed: where a sovereign operator can meet the workload at competitive price and performance, the sovereign operator wins the business. This is the same procurement preference that supports the local steel, ammunition and shipbuilding industries; it requires no new legal architecture and no exotic policy instrument. It requires only that the government make the decision.

How much compute does Australia actually need? An upper bound comes from simple arithmetic. The United States currently runs about thirty gigawatts of continuous data-centre load for 340 million people: roughly ninety watts per American, around the clock, of which the AI share is about a third. It is an upper bound because America exports compute; some of those watts are already serving Australians. Scaled to our population, full American parity is about 2.4 gigawatts of continuous load, and the AI share of it about 0.8 gigawatts. A bottom-up estimate arrives at the same place: machine cognition performing a quarter of the work of our fourteen-million-strong labour force, at a few hundred watts per digital worker, is one to three gigawatts of inference.²⁷ Two methods, one answer: the sovereign compute estate is a gigawatt-class build. A gigawatt of accelerators is roughly six hundred thousand current-generation GPUs, an order of magnitude beyond the largest sovereign clusters now being built (§10.6), and comfortably inside the six gigawatts of data-centre capacity already queued for connection to the Australian grid, nearly all of it foreign-owned. The country is already building the right amount of compute. It is building it for someone else.

10. The Model Layer

The model layer is the layer at which the popular discussion of AI almost entirely takes place. It is also, perhaps for that reason, the layer at which the policy discussion most often goes wrong, because the popular framing, "we need our own GPT", is both more ambitious than is necessary and less ambitious than is sufficient.

10.1 What Is Useful, What Is Necessary

The argument for a wholly indigenous frontier model, trained from scratch in Australia, on Australian compute, with Australian data, by an Australian research team, is the kind of argument that is seductive in a manifesto and unaffordable in a budget. The training cost of a GPT-4-class model is a substantial fraction of a large-defence-procurement programme; the talent cost is the entire research output of the country for several years; and the model, once built, will be of comparable but not superior quality to the foreign open-weight alternatives that are available, today, at no marginal cost. The argument for a from-scratch frontier model is, in 2026 conditions, a vanity argument. It is a flag-planting exercise that consumes resources that the country needs at every other layer of the stack.

The unaffordability is a property of breadth. A frontier model is expensive because it is general: it holds the whole recorded surface of human activity so that it can be asked anything, and the editing of this paper is itself evidence of the utility of that generality. But a model built for a bounded domain does not need the breadth. The model that runs a haul fleet or reads a core sample has no need of classical French poetry. A foundational model for a specific domain can be smaller by orders of magnitude, in parameters, in training data, in compute, and in the talent needed to build it, and at that scale the from-scratch argument stops being vanity and becomes achievable. What Australia cannot afford at the frontier it can afford at the domain of our own frontier, and §10.4 argues it should.

The reason generality suffices for human work also marks its limit. A foundational model is like a university graduate: the education is general so that the job can be specific, and you can hand a graduate a manual and expect competence, because the education taught them how to read manuals and how to understand the symbols and directives contained therein. They are trained on the recorded output of the species, a frontier model is that graduate for nearly every job whose knowledge has been written down. But the corpus can only reach what was recorded. The knowledge that runs an iron-ore pit or a broadacre harvest was never fully written; it sits in drill logs, telemetry, yield maps and survey data, held by the operators and agencies that produced them, and much of it is Australian. The frontier model has read everything ever published and has seen none of this. A domain foundational model trained on that data is not a lesser copy of a frontier model; within its domain it knows what no frontier model can. That is the meaning of domain-specific leadership in §10.4, and the custody argument of §11 is what keeps the advantage sovereign.

The argument for no indigenous capability at the model layer is the opposite mistake. A country that does not have anyone in the room when frontier models are being designed, trained, fine-tuned, evaluated and aligned is a country that does not understand the artefacts on which its economy is increasingly running. It will have no advance warning of capability changes, no real understanding of failure modes, no diagnostic capacity when its deployments break, and no negotiating leverage with the foreign suppliers on whom it depends. It will be, in the strict sense, a model consumer rather than a model owner.

The right policy stance is between these two. We articulate it as three propositions.

10.2 Government as AI Utility Provider

The first proposition is that the Federal Government should operate, on a public-utility basis, an AI service available to every Australian and every Australian organisation. The model behind the service need not be wholly indigenous; in the early years it almost certainly will not be. What matters is that the service is provided through Australian infrastructure, under Australian governance, with Australian terms of service, by an Australian operator, accountable to the Australian polity. The model is housed and operated sovereignly even where it is not originated sovereignly.

The utility framing is deliberate. We do not propose, and we do not believe, that the government should be the only AI provider in the country. Australians who prefer foreign services should remain free to use them. Australian organisations that wish to operate their own private AI deployments, their own retrieval-augmented generation systems, their own fine-tuned models, their own internal data pipelines, should be encouraged to do so, because the Hayekian search of §5 demands it. The point of the utility is not to monopolise; it is to ensure that no Australian is forced into a foreign relationship in order to use the technology. The utility is the floor, not the ceiling. It guarantees access on sovereign terms; it does not preclude private provision of richer services on top.

This is the same model as Medicare, the public broadcaster, the post office, the public library and the national broadband network. None of these displaces the private market; each ensures a sovereign minimum. The AI utility belongs in the same category. It is the answer to the question "where does the citizen who does not wish to be a customer of OpenAI go?" and currently the answer is "nowhere".

The utility makes possible, in addition, two further things that no private provision can deliver. The first is the augmentation of every citizen that §5.4 identified as the alternative to UBI. A free-at-point-of-use AI utility, available to every citizen on the same terms, is the universal augmentation that keeps the population in the new Hayekian search. The second is equity of access. Differential access to capable AI is, on the current trajectory, becoming a new and severe form of economic inequality. Citizens who can afford the $300/month frontier subscriptions are productive at a multiple of citizens who cannot. The utility, by levelling access, prevents the productivity divide from becoming a permanent stratification.

10.3 The Pragmatic Acquisition Path

The second proposition is that the model layer should be acquired pragmatically, on a portfolio basis, with sovereignty established at the operational layer rather than at the origination layer. The graduate starting points of §10.1 already exist, and they are free to take: the open-weight families trail the proprietary frontier by roughly a year, and in recent years the gap has narrowed, not widened. For a bounded domain, trained onward on data we hold, a year-old graduate is no handicap.

Operationally this means: Australia hosts open-weight frontier models (Llama, Mistral, DeepSeek, Qwen, the next generations of these as they appear) on its sovereign compute, fine-tunes them on its own data, evaluates them against its own benchmarks, aligns them under its own values (§14), and exposes them through its own APIs. The original weights are not Australian in origin but the deployment wholly is. This is the same arrangement under which Australia has, for a century, benefited from foreign-originated technology adopted to local conditions: the Holden was a domesticated General Motors, the F-111 a domesticated American bomber, the NBN a domesticated Korean fibre architecture. The pattern is well understood and the country has, when it has been deliberate, done it well.

The portfolio principle matters because it removes single-supplier dependency. A country that hosts five open-weight model families, fine-tunes each on its own data, and rotates them through its production deployments based on cost and performance, has effectively diversified its model risk. No one upstream provider, and no upstream government, can switch off Australian AI by closing the API. The weights are already in the country, the fine-tunes are already done, and the deployments continue under sovereign control. This is the defensive case for portfolio acquisition: it removes the kill-switch. It is also the answer to the open-weight scepticism of §3.4. Weights without the surrounding operational practice are an inert lump, and the acquisition path is the deliberate construction of that practice: the hosting, the fine-tuning, the evaluation, the alignment, built as Australian capability. The weights are a snapshot; the practice is what makes them productive, and the practice is ours to build.

It is also worth noting that Australia is currently an outlier among middle-to-large powers in its absence of such a strategy. France has aggressively fostered Mistral and Kyutai as national champions, lobbying the EU to protect its sovereign capacity. The United Arab Emirates foresaw the compute bottleneck early, used sovereign wealth to stockpile massive clusters, and produced the Falcon and Jais models. Singapore is actively building SEA-LION to capture Southeast Asian linguistic and cultural norms that foreign models miss. India has launched a $1.2B "IndiaAI Mission" to subsidise sovereign compute and domestic models. These countries have independently reached the same conclusion: they are not trying to out-spend the Americans on a generalist frontier model, but they are stockpiling compute, securing their operational data, and training models on their unique cultural or regional data. We are proposing that Australia do the same.

Where indigenous origination is justified is at the high-value, narrow-domain end. We will not train a generalist frontier model, but we will train, and have already trained in academic and industrial laboratories, specialist models in domains where Australian data is dominant or where Australian use cases require capabilities the foreign generalists cannot deliver. This is the third proposition.

10.4 Domain-Specific Leadership

In each of several domains, Australia possesses or could readily assemble a corpus of operational data, accumulated practice and proprietary technique that exceeds anything available anywhere else in the world. The most prominent of these is mining: the country is the world's largest exporter of iron ore, the largest exporter of metallurgical coal, a major exporter of bauxite, lithium, rare earths and a dozen other minerals; the country has half a century of operational telemetry from the largest fleet of autonomous mining equipment in the world; and the country has, in companies like Rio Tinto, BHP and Fortescue, organisations that have spent decades digitising every aspect of mining practice. A mining-specialised AI, trained on this corpus, would not be a better generalist model than GPT or Gemini; it would be a categorically different artefact, capable of work the generalists cannot do because they have not seen the data the specialist has trained on.

The same is plausibly true in agriculture (we have unique broadacre, pastoral and viticultural data), in marine science (the Great Barrier Reef and adjacent waters generate the world's richest reef-ecosystem dataset), in environmental management (the Murray-Darling, the rangelands, the bushfire archive), in mineral exploration (Geoscience Australia holds one of the world's best-documented continents), and in remote-area medicine (the Royal Flying Doctor Service has a unique remote-clinical corpus). Each of these is a candidate for a specialised Australian model that would be the global leader in its niche, not because Australian researchers are uniquely gifted but because the data is uniquely Australian. We may not lead the frontier; but we can lead specific domains; and the domains we can lead are exactly those in which Australia is already, by accident of geography and industrial history, the world's leading practitioner.

These specialist models are simultaneously defensive, they keep Australian operational know-how in Australian hands, and offensive, as they are exportable. The country that holds the world's leading mining-AI sells access to that AI to every mining operation in the world that competes with the Australian originals only on price, not on capability. The same logic applies to every other domain in which Australian data is world-leading. The specialist-model layer is the export industry that the policy discussion has not yet identified, and it is one in which Australia has structural advantages that no policy intervention can confer on a country that lacks the underlying data.

10.5 The Scalpel and the Swarm

There is one further observation to make about the model layer regarding cost, because the common objection to everything proposed above is cost. The objection assumes that AI comes in only one shape: the shape the American frontier laboratories have given it. Multi-trillion-parameter monoliths, trained on clusters costing billions, housed in campuses drawing gigawatts. On that price list, sovereignty looks unaffordable, and the argument is muted before it can be articulated. But that monolith is a business model, it is not a law of physics and it is not how the same problem was solved by nature. The barriers to entry it erects are not incidental to the paradigm; they are the point of it. And the shape is tuned to the problems that live in the corpus. A monolith trained on the written record is superb at the work of the written world; the problems we most need solved were never written down. §10.1 made the point: the knowledge of the pit and the paddock is in our telemetry, not their training set. The paradigm that suits the seller's business is not the paradigm that suits our problems.

Evolution is always parsimonious, and its work in our own species and the others suggests the alternative. Human dominance was never the work of a single huge brain. It was the work of many smaller brains, specialised, coordinating, passing refined knowledge from mind to mind. What some brains learned expensively, a thousand others acquired cheaply. Even our own brain is not one brain: it is built of specialised cortices, one shaped for vision, one for hearing, each with an architecture fitted to its sense, coordinated under a general executive, and the whole apparatus runs on twenty watts. The frontier laboratories have conceded the design principle inside their own products: in the newest giants, only a fraction of the network fires on any given question, because firing all of it is waste. The Manhattan Project is the pattern at its most concentrated: a problem no individual could have solved, solved by a large team of very smart people working on sub-tasks, coordinated by more very smart people. Some problems are also domain-limited in scale, and a smaller, specialised worker subsumes them entirely. Applied to AI infrastructure, this blueprint yields two instruments which together act as both a scalpel and a swarm:

[ The Centralised Monolith ]   ──►  brute force, high capital, extraterritorial risk
              VS.
[ The Evolutionary Swarm ]     ──►  surgical distillation + sovereign neocloud infrastructure

1. The Scalpel: knowledge distillation. Instead of spending hundreds of millions of dollars teaching a monolithic model basic logic and grammar from scratch, this approach treats the open-weight frontier models that §10.3 identifies as teachers. On local hardware, their reasoning capabilities and chain-of-thought patterns are extracted and surgically transferred into compact, hyper-efficient student models of a few billion to a few tens of billions of parameters. Nothing is pruned or removed from the teacher. The teacher generates worked reasoning over the specialist corpus, and the student is a new, smaller model trained on that output.

The result: the student runs on a fraction of the memory, power and silicon, and on its chosen domain it approaches the teacher's competence at the work the deployment actually requires. Trained against the teacher and fine-tuned on the strategic corpora of §11.3, it does not match the original in general intelligence; general scope is achieved by having multiple models. This is the affordable route to the specialist models of §10.4. Mining, agriculture, marine science and remote medicine do not need a model that can do everything. They need a model that does one thing at world standard and runs on hardware a regional operator can own.

2. The Swarm: sovereign neoclouds and edge nodes. Instead of routing the nation's data to a few hyperscale campuses, compute is localised across distributed networks: regional bare-metal clusters of hundreds of accelerators rather than hundreds of thousands, agency and enterprise nodes, and edge devices on vehicles, instruments and clinic hardware, all within Australian legal boundaries.

The result: because the distilled models are small, they do not require data factories. Sovereign compute becomes a network of modest sovereign nodes rather than one gigawatt campus, pooled where scale is needed, local where it is not, communicating with one another like a human community rather than a single distant oracle. This suits a dispersed continent. The tyranny of distance once forced Australians to adapt and to innovate, and the swarm is the same instinct applied to intelligence.

Why this beats the monolith. Three properties follow.

Data sovereignty. Sensitive data (defence archives, national health records, legal infrastructure) never crosses an international border and never touches a foreign provider's API. With the teacher weights themselves hosted on sovereign compute, the distillation pipeline and the deployments it feeds are self-contained within Australian jurisdiction.

Asymmetric economics. A sovereign neocloud cluster (§10.6) of 256 to 1,024 advanced accelerators cannot compete with a 100,000-GPU training cluster on raw volume, and it does not need to. By concentrating capital on distillation and continuous local fine-tuning, it delivers frontier-grade utility at a small fraction of the capital expenditure of an American giant.

Resilience. A monolithic AI is a single point of failure, subject to corporate whim, policy shift and compute taxation. A distributed landscape of specialised, distilled models on local infrastructure is modular, resilient, and structurally hard to monopolise: no single point of failure, no single landlord, and no single switch that a foreign supplier or a foreign court can reach. The same property holds against harder instruments than courts, and §10.7 takes it up: no single building whose loss stops the economy.

One caveat is that the swarm still depends on upstream teachers. Those teachers are foreign open-weight releases, and if the releases stop, the distillation pipeline stops improving; its ceiling becomes the last weights acquired. This is the dependency the stockpile-and-substitution strand of §9 exists to manage: weights already inside the country continue to teach, and acquiring each generation of open weights as it appears keeps the teaching stock current. The swarm does not remove the dependency; it shrinks it to a single input that can be stockpiled. The dependency is also narrower than it looks because it applies to general capability, not to domain competence. And general capability saturates for bounded work. We are overtraining models for most tasks. The difference between a graduate and a professor matters in a seminar, but it matters much less in the nuts and bolts of operational management, where the task was mastered several releases ago and every release since has been surplus to it. You do not need a PhD behind the wheel of a dump truck. What the work needs is what workers have always needed: a Leaving Certificate education, sixteen years of experience of the world, the manual, and a grizzled old hand to show them the trade. The swarm is that apprenticeship in silicon: the compact student arrives with the schooling and the world experience already pressed into its weights, retrieval keeps the manual open at the right page, and the frozen teacher is the old hand. Five years ago none of these models existed, and we did ok. The rush since is competitive: each firm fears, with reason, being made irrelevant by the next release. That fear sets the pace of the frontier; it is not a property of the work. Constant upstream churn in fact runs against the quality discipline these deployments live under. Industrial quality is sameness of output, and sameness of input and control is what lets variation in output be engineered rather than suffered: a process that shifts with every release cannot be certified, and a fault can no longer be traced to its cause. For certified operational work, a frozen, stockpiled teacher is not a compromise; it is the specification. Manufacturing settled the underlying principle a century ago. The material advantage automation offers is process efficiency, making more of the same, faster, and quantity of quality comes from dedicated machines, with the degrees of freedom designed out and the process locked. A universal machine with many degrees of freedom under software control is a prototyping tool; on the plant, chutes and conveyors are cheaper and just as functional as a ten-degree-of-freedom robot arm. The giant general model is the universal machine of cognition. The distilled specialist is the dedicated machine, and the volume production of quality belongs to it. For work of that shape the leverage moved some time ago from the quality of the teacher to the quality of the tempering, and the tempering is done with operational data we hold. The student models train continuously on the Australian data their deployments generate, so a mining or agriculture model keeps improving long after its teacher has stopped. The upstream supplies the general reasoning base; Australia supplies, and keeps supplying, everything that makes the model worth owning.

Nor is a full stop likely. Open-weight releases may fall behind the closed frontier, but too many laboratories in too many countries now publish them for the supply to end altogether. The releases that continue might not be frontier general models. But, does that really matter? In this architecture the general models are the glue: they route, converse and coordinate. The value is added in the specialist models, in the industry targets where Australia holds the data and the expertise. A slightly older binder containing world-leading specialists is a better position than a tenancy on someone else's frontier.

Together, these change the budget question. "Can Australia afford sovereign AI?" This is a question asked on the current paradigm's sticker price. But by using swarm economics, a middle power can field frontier-grade capability in the domains it chooses, on infrastructure it owns, at a cost within the normal range of national infrastructure spending. The expensive path is not the sovereign one; the expensive path is the tenancy. One thing the swarm does not change is the energy argument of §8: many modest nodes still sum to a national load, and the utility of §10.2 still serves a whole population. The swarm changes the shape of the demand, not the need for firm power beneath it.

10.6 The Sovereign Neocloud

A neocloud is a cloud provider that specialises exclusively in accelerated compute. It is not a lesser tier of AWS, Google or Azure. Because it does nothing else, a neocloud typically deploys bare-metal GPU clusters on dedicated high-speed InfiniBand fabrics, without the heavy virtualisation overhead of the legacy providers. Wire together twenty thousand Blackwell GPUs in a single campus and the physics and networking work just as well as they do inside Microsoft or Meta. The barrier is not architectural; it is capital.

That barrier divides the field into two tiers. Pre-training a trillion-parameter global frontier model from a blank slate requires a continuous, uninterrupted cluster of roughly 10,000 to 100,000 top-tier GPUs in a single location for months, with billions of dollars in upfront capital and gigawatt power substations behind it. Only a handful of operators worldwide, hyperscalers and the largest American neoclouds such as CoreWeave, work at that scale. Sovereign neoclouds work at a different tier: clusters of 256 to 2,048 GPUs, an extraordinary amount of compute, but orders of magnitude too small to brute-force a frontier model from scratch.

Australia is already building this tier. Sovereign Australia AI has placed the nation's largest sovereign AI order, 256 NVIDIA Blackwell B200 GPUs hosted in NEXTDC data centres, to train Australian-owned, Australian-governed models on ethically sourced national data. Sharon AI has announced a further 1,000-GPU B200 cluster at NEXTDC's Melbourne M3 facility. ResetData's AI-F1, an immersion-cooled H200 cluster in the Melbourne CBD, is the most powerful public sovereign GPU cluster in the country, consumed as bare-metal or clustered GPU-as-a-service. Scaile, built by xAmplify with AUCloud and VAST Data, offers government-grade elastic GPU compute with every byte, metadata included, under Australian jurisdictional control. And at the fully distributed end of the spectrum, the global peer-to-peer compute markets (io.net, Render, Akash) already pool Australian nodes: gaming rigs, idle mining hardware and local data-centre capacity in Sydney and Melbourne, aggregated by protocol rather than by landlord. The swarm is not hypothetical; its first nodes are live.

Distance makes this regional pooling an engineering requirement, not a preference. A developer using a compute network whose nodes sit in Europe or North America inherits 150 to 250 milliseconds of subsea round-trip, enough to break the tight communication loops that training and real-time inference demand. The tyranny of distance that §10.5 invoked as the origin of the Australian instinct to adapt is also the hard technical case for the swarm: aggregate the continent's GPU compute locally, so that Australian innovators neither queue for foreign capacity nor pay the latency toll of using it.

A neocloud at this tier builds frontier-grade capability in four ways:

1. National foundation models, trained from scratch. A cluster of 256 to 1,024 GPUs can comfortably pre-train a dense model of 8 to 70 billion parameters from a blank slate. This is not the frontier vanity project of §10.1: trained on curated, clean, highly relevant national data, a 70-billion-parameter sovereign model can outperform a generic 400-billion-parameter global model on the local enterprise, government and legal tasks it was built for. Sovereign Australia AI's Australis model, a foundational model being built from the ground up within Australian data centres, is this strategy in motion.

2. Continuous pre-training on open-weight bases. Rather than spend the electricity to teach a model basic English and general logic from scratch, engineers import a state-of-the-art open-weight base (Llama 3.1 405B, DeepSeek-V3) into the sovereign jurisdiction and continue its training there, injecting national data (health, legal, defence, statutory) directly into the model's weights. Sovereign Australia AI's Ginan, an 8-billion-parameter model fine-tuned from Llama 3.1 on roughly two billion tokens of Australian data, is an early example.

3. Rack-scale sovereign inference. For many organisations, building on a frontier model simply means running one in production without breaching data-privacy law. A single rack-scale system inside a local data centre runs a 400-billion-parameter open-weight model with local routing; the data never leaves the country, the prompts are never logged across tenants, and the CLOUD Act problem of §9 does not arise.

4. Distillation. The scalpel of §10.5, and the strategy that deserves the fullest treatment, because its two halves map onto exactly the hardware a sovereign neocloud can afford. It also composes with strategy 2: continue the teacher's pre-training on national data first, so that the teacher is domain-adapted before it teaches, then distil the adapted teacher into deployable students.

The teacher requires high memory, not mass compute. To act as a teacher, an open-weight frontier model (DeepSeek-R1 at 671 billion parameters, Llama at 405 billion) does not need to be trained. It only needs to run inference to generate training data, and hosting a model of this class for inference takes a single 72-GPU rack, or a few 8-GPU nodes.

The student requires minimal training compute. Because the student is small, 8 to 32 billion parameters, training it on the teacher's generated data does not require a 50,000-GPU supercluster. A standard sovereign neocloud allocation of 64 to 256 GPUs can complete a distillation run in days or weeks.

The pipeline that results is closed-loop and wholly sovereign:

[ Open-Weight Frontier Model ]
                  │
                  ▼   continuous pre-training on national data (strategy 2)
[ Domain-Adapted Teacher ]
                  │
                  ▼   teacher reasons over local, classified or sovereign documents
[ Synthetic Reasoning Dataset ]
                  │
                  ▼   train on 64–256 local GPUs
[ Compact Sovereign Student (8B–32B) ]

Step 1, air-gapped teacher hosting and adaptation. The neocloud deploys the open-weight teacher onto a local high-memory rack and, where the domain justifies it, first continues its pre-training on national data (strategy 2). Hosted natively on sovereign soil, it is severed from foreign jurisdictions and cloud telemetry.

Step 2, sovereign synthetic data generation. The organisation feeds its sensitive data (defence archives, health records, domestic case law) into the teacher locally. The teacher generates millions of tokens of step-by-step reasoning and analysis grounded in those documents.

Step 3, student training. On the neocloud's remaining GPU allocation, engineers train a compact open base model on the synthetic dataset. The small model learns the teacher's analytical behaviour without ever having to memorise the entire internet from scratch.

The industry has already demonstrated the payoff. Reasoning behaviours distilled from massive models into open models as small as 1.5 and 8 billion parameters have outperformed older 70-billion-parameter models on logic, mathematics and coding benchmarks. And once the sovereign student is distilled, the daily workload no longer needs the 72-GPU rack: the distilled model deploys on standard enterprise servers, single 8-GPU nodes, or ruggedised edge hardware in defence and mining, at up to ten to twenty times lower cost per token than serving the giant. Building a global frontier model from scratch is, as §10.1 argued, a vanity metric. Distilling open frontier models into specialised, secure domestic models is the practical blueprint for national AI capability.

10.7 The Compute Harvest

The swarm has a strategic justification that the resilience point of §10.5 understated, and it comes into force at exactly the moment this paper's programme succeeds. When data centres come to perform a significant fraction of an economy's productive work, the buildings themselves become strategic targets of a kind not seen since the refineries and ball-bearing plants of the 1940s, and the attack need not be a missile. The cheap paths are the substation, the fibre and the cooling plant, as the strikes on Ukraine's grid and the severed Baltic cables have lately demonstrated. Nor does an adversary need to strike at all: a credible threat against three campuses that run a third of GDP is coercive leverage on its own. §6 argued that the resource value of this continent will rise through the AI transition; a handful of gigawatt campuses would concentrate that value into a handful of aiming points. The distinction that matters is between training and inference. Training genuinely concentrates, but the loss of a training cluster costs a country its rate of improvement, not its economy, because the weights already trained survive and copy freely. It is the inference estate, the layer that actually runs the country day to day, whose loss would stop it, and inference is precisely the layer that can be dispersed.

The physics of the model sets the terms of the dispersal. The wall that forces compute into one building is the interconnect demand of training a giant: a model too large for any one chip must be spread across many, and the arithmetic then runs on the wiring between them. That demand scales with the size of the model being built. A frontier monolith needs the gigawatt campus. The specialist models this paper chooses, the students of §10.6 and the domain models of §10.4, sit below the wall: they train inside a single cluster, fine-tune inside a single node, and serve from a single rack. The Commonwealth's existing facilities, Pawsey in Perth and CSIRO's Virga, are already the right tier for models of this size. Model size, interconnect, building and target are the same variable seen from four sides. Choose small models and the estate may disperse; chase the monolith and the physics herds everything into one place worth striking. The monolith's scale is, for its owners, as much a moat as a design philosophy, and Australia gains nothing by digging a moat around someone else's castle.

The positive form of the argument is an architecture with two tiers, neither of them a gigawatt campus. The upper tier is many smaller data centres, sited where the grid already is: co-located with substations, drawing on interconnection capacity that exists rather than queueing years for new transmission, each at the scale of the sovereign neocloud clusters of §10.6, a few hundred to a few thousand accelerators drawing single-digit megawatts. Centres of that size are cheap enough for regional cities, close enough to their users for latency to vanish, and numerous enough that they become tactical rather than strategic targets. A controllable load beside the switchgear is also an asset to the network rather than a burden: compute can shed or soak demand within seconds, which is grid firming by another name. And the soak has a natural feedstock. Australia's rooftop solar fleet, the densest in the world, floods the distribution network every clear midday, to the point where the operator curtails it and wholesale prices go negative; a compute fleet parked at the substations is a sink that turns the curtailed surplus into work at close to zero fuel cost. This is the complement of §8, not a retreat from it: the constant inference load still wants firm power beneath it, but the batch half of the national workload, the teaching-data generation and student training of §10.6, keeps no schedule and can follow the sun across the day. The lower tier is the shared civil and corporate compute this section now describes.

The lower tier should be strategically distributed from the outset. The rooftop solar programme subsidised privately-owned generation at the premises, and two decades later the country operates one of the largest distributed energy fleets in the world, built at household capital expenditure rather than utility scale. The same instrument, pointed at compute, is a feed-in scheme for computation: subsidise the accelerator on the business premises, let the business use it for its own AI, and harvest what it does not use into the national pool. The business gets the capability that the National AI Plan's ninth failure leaves it to find unaided; the commonwealth gets a compute fleet that no single strike, court order or corporate decision can switch off.

The economics rest on a simple property of the hardware: an accelerator dies of obsolescence, not of wear, so every idle hour is capital thrown away. And the idle hours are most of them. An office computer's processor is largely idle even while the office is busy, because correspondence, documents and spreadsheets barely wake it; the exceptions, design, rendering, simulation, are trades, not the norm. Even a desk running a local AI front-end occupies its accelerator for only bursts of the working day, since the machine does nothing between its user's thoughts. The two uses therefore share the same silicon at fine grain: the local user holds absolute priority and never waits, and the national pool is the background tenant that soaks the gaps. No schedule of exempt industries is needed, because the meter settles it: each machine is paid for the cycles it contributes and pays for the cycles it consumes, and the feed-in price discovers, machine by machine, where idle compute is worth most. The reader will recognise the instrument; it is the same one this paper reaches for at every layer, a price doing the work a planner cannot.

The coordination layer this requires is not a research problem; it is a solved problem with a history. In 1999 the SETI@home project began harvesting the idle cycles of volunteer home computers to search radio telescope data for signs of intelligence, and at its height the aggregate outran the supercomputers of its day. Its descendant proved the ceiling: in 2020 the Folding@home swarm of donated machines reached roughly 2.4 exaFLOPS against the pandemic, the first computing system in history past the exaflop mark, two years before any national laboratory got there with a purpose-built machine.²⁸ A volunteer swarm beat the monoliths to the milestone. The same lineage solved the trust problem a quarter-century ago, by redundancy: the same work unit issued to several machines, the answers compared, the disagreements discarded. A paid scheme sharpens the incentive to cheat, so the harvest adds hardware attestation to the redundancy, but the paradigm holds. The symmetry is too apt to leave unremarked: the distributed search for intelligence, redirected from the sky to ourselves.

The harvest's diet is the swarm's diet: inference, fine-tuning of the distilled students, and the generation of teaching data from the stockpiled teachers of §9, all of it tolerant of thin links and vanishing nodes. What it does not do is frontier training, which still demands the concentrated fabrics of §10.6, though the research frontier in low-communication training is moving toward the distributed case rather than away from it.²⁹ The subsidy itself should be paid in the coin of §4.5: compute enrolled in the harvest earns back the levy that compute serving bare substitution pays, so the tax on the machinery that displaces and the subsidy for the machinery that participates are one instrument, not two. And the scheme completes the energy argument rather than straining it: a dispersed fleet draws dispersed power, which is the demand shape that the modular reactors of §8 were built to serve. Energy at the base of the stack, dispersed; compute above it, dispersed; the economy above both, correspondingly hard to stop.

10.8 The Combined Architecture

The three propositions, and the swarm economics that make them affordable (§10.5–§10.6), stack into a coherent architecture. The Federal Government operates an AI utility, hosting open-weight frontier models on sovereign compute, available to every citizen. Private-sector organisations operate their own private deployments, fine-tuned on their own data, on the same sovereign infrastructure or on their own. A small number of strategic specialist models (in mining, agriculture, environmental management, remote medicine) are originated indigenously, largely by distillation from the hosted giants (§10.5–§10.6), exported globally, and contribute revenue that subsidises the utility and the broader programme. The country is, simultaneously, a sophisticated consumer of frontier AI, a sovereign operator of all the AI it cares about, and a global leader in a handful of carefully chosen domains.

This is achievable. It does not require us to compete with OpenAI or Anthropic at the frontier. It requires us to be deliberate about which of the three roles we play in each market we care about, and to refuse to be passive in any of them.

11. The Data Layer

The model layer rests on the data layer. Models are trained on data, fine-tuned on data, evaluated on data and aligned against data. The model is, in a precise sense, a compressed representation of its training corpus. To own the model meaningfully one must therefore own, or at least have legal authority over, the data on which it was trained and on which it continues to learn.

The Australian data sovereignty discussion has, to date, been conducted almost exclusively in privacy terms: where is my data, who has access to my personal information, is my health record on a foreign server. These are real questions and we do not minimise them. But they are only one of three layers of the data sovereignty problem, and on inspection the lower-priority of the three. We treat the three layers in ascending order of strategic importance.

11.1 The Personal Data Layer

The first and most-discussed data sovereignty question is personal data: the records of individuals, their identities, their finances, their health, their communications, their movements, their relationships, that exist as digital artefacts and that, in the AI economy, are increasingly used as fine-tuning material, retrieval substrate and behavioural training data for foreign models. The Australian regulatory regime here is comparatively well-developed (the Privacy Act, the Notifiable Data Breaches scheme, the OAIC's enforcement function) but it has been overtaken by the deployment patterns of the AI industry. A model that has trained on, or been fine-tuned on, the records of Australian individuals has, in a meaningful sense, internalised those records. The model retains and can in some circumstances regurgitate them. The records are no longer in any one place; they are diffused into the weights, where standard data-protection mechanisms (deletion, rectification, access requests) do not, in any operationally meaningful way, apply.

The sovereign answer is twofold. First, a stronger statutory regime that treats personal data as held in trust rather than alienated by consent, with consequent restrictions on the use of such data in AI training and the enforceable right to demand model-level remediation where the trust has been broken. Second, the AI utility of §10.2 should hold personal data of its users only under the strict regime that applies to existing public-sector record-holders (Medicare, the ATO, Centrelink), with a deliberate firewall against use in model training without specific consent. This is hard, technically and legally, but it is the foundation without which the rest of the data conversation cannot be honestly conducted.

11.2 The Operational IP Layer

The second layer is the data sovereignty problem that the present plan does not name and that we will, in this paper, give particular emphasis. Every Australian organisation that uses a foreign AI service for its internal work, drafting memos, preparing legal opinions, analysing financial positions, generating marketing copy, summarising research, reviewing tender responses, sends, in the course of that use, the raw material of the organisation's operational intelligence to a foreign server, where it is retained, may be used in fine-tuning, and is in any case visible to the foreign provider's operations and security teams. This is happening, every day, at scale, across the Australian economy. It is happening at the cabinet office, the major banks, the major law firms, the major engineering consultancies, the major mining houses, the major media organisations and every government department that has not specifically prohibited it. The intelligence content of the Australian commercial economy is being continuously transferred to foreign-owned model providers as a side-effect of normal day-to-day work.

The strict reading of the foreign providers' terms of service usually forbids the use of customer prompts in training, and we accept that the major providers honour their terms in this respect. The point is not that the prompts are being maliciously exploited. The point is that they are visible. They are visible to the provider's operations staff, visible to the foreign government on receipt of a CLOUD Act warrant, visible to anyone with sufficient access in the foreign company, and, most importantly, they are not visible to any Australian authority charged with protecting the country's commercial intelligence. The leak is structural, lawful, deniable and ongoing. Nor can it be audited from our side. A provider served with a foreign intelligence demand is routinely barred, by the same foreign law, from disclosing that the demand was made; the model owner could not tell us our data had been passed to a foreign government even if it wished to.

The sovereign AI utility removes the leak. An organisation that uses an Australian-operated AI service for its drafting, analysis and synthesis work has the same productivity as the organisation that uses ChatGPT, with the difference that the prompts and outputs remain inside Australian jurisdiction. The intelligence content of the work product is retained. The operational IP, the way the organisation thinks, the deals it is considering, the positions it is preparing, the customers it is targeting, the technologies it is developing, is preserved within the Australian commercial sphere. This is the data sovereignty argument that matters most to Australian business and least to Australian privacy advocates, and it is therefore the argument that has been most under-discussed. We make it here at length because it is, in our view, the data argument that will, when it is finally noticed, drive the commercial adoption of the sovereign utility on a scale that no privacy argument ever could.

11.3 The National Strategic Data Layer

The third layer is the strategic-corpus question that motivates the domain-specific leadership argument of §10.4. Some Australian datasets are not merely useful; they are uniquely valuable globally. The Geoscience Australia continent-wide geophysical dataset, the operational telemetry of the autonomous Pilbara mines, the Murray-Darling water-systems dataset, the Reef monitoring corpus, the Bureau of Meteorology continental observations, the CSIRO climate and agricultural archive, the Royal Flying Doctor Service remote-medicine corpus, the broadcast archive of the ABC and SBS, the parliamentary record, the reported case-law of the Australian courts: each of these is, in its domain, a global asset of the first rank. Each, untrained-on, is dormant capital. Each, trained-on by the foreign frontier providers without compensation or governance, is exfiltrated capital.

The sovereign data layer at this level requires three things. First, cataloguing: a deliberate national exercise to identify the strategically valuable datasets, document their custody, and assess their AI-relevant value. The country has done nothing comparable. Second, protection: a legal regime that prevents the unconsidered ingestion of these datasets into foreign model training, perhaps by an assertion of public-domain copyright in government datasets coupled with explicit licensing terms that exclude AI training without payment. Third, productive deployment: the strategic datasets become the training and fine-tuning corpora for the indigenous specialist models of §10.4, where Australian custodianship of the data converts directly into Australian leadership of the specialist model market.

11.4 The Three Layers Together

The three layers (personal, operational, strategic) share one common feature. In each, the failure to act sovereignly is not a theoretical risk but a present transfer of value to foreign holders. Personal data is being absorbed; operational intelligence is being silently leaked; strategic corpora are being scraped and trained on. The sovereign data programme is therefore not a precautionary measure against possible future harm. It is a remedy against present, ongoing, quantifiable losses, and the speed at which the remedy is applied will determine how much of the asset base remains to be defended.

12. The Talent Layer

The talent layer is the layer at which optimistic policy documents most often promise the most and deliver the least, because talent is harder to build than infrastructure and slower to manifest than budget allocations. The 2025 National AI Plan contains the customary undertakings, STEM scholarships, research fellowships, an AI capability uplift in the public service, mid-career retraining programmes, and we do not dispute that these are useful. We dispute that they are sufficient, and we dispute, more sharply, that the framing in which they sit understands the problem.

The framing error is the assumption that the talent problem is a production problem. It is not. Australia produces excellent AI researchers and engineers; it has done so for decades; its university departments, particularly at the Universities of Sydney, Melbourne, NSW, Queensland and the ANU, train people whose work is, on inspection of the publication record, indistinguishable in quality from the output of the leading American and European laboratories. The CSIRO's Data61 division, despite its budgetary turbulence, retains world-class expertise. The mining-automation sector has produced applied AI engineers of a calibre that would be welcome in any frontier laboratory in the world. The talent is, and has been, produced.

The talent problem is a retention problem. The Australians who could build the sovereign AI stack are, in disproportionate numbers, currently building the foreign-owned AI stacks of San Francisco, Mountain View, Seattle, London and Beijing. They are not in those places because they prefer the climate. They are there because that is where the work and payroll is. The structure of substantive AI work in Australia is small, fragmented and poorly funded, while the structure of substantive AI work overseas is large, concentrated and lavishly resourced. A talented Australian researcher who completes a doctorate at a major Australian university is presented with a choice between a domestic postdoctoral position at a fraction of overseas pay, with limited compute resources and an uncertain path to senior research, and an overseas position at a leading laboratory with effectively unlimited compute and a clear path to influence. The decision is over-determined. They leave, and most do not return.

The retention answer is the same answer that the energy, compute, model and data discussions have already converged on: build the structure of work in which the talent can be deployed. A sovereign AI utility serving the entire Australian population is a working environment of consequential scale. A specialist-model programme in mining, agriculture and environmental management offers research problems of genuine international significance, on data the foreign laboratories do not have. A sovereign compute footprint of multiple gigawatts is an operational substrate worth working on. The talent does not, in the end, follow the salary alone; it follows the work. If the work exists in Australia, on a scale that is internationally credible, the talent will not need to be retained by inducement. It will return because the work has come home.

This will not happen on its own. It requires deliberate concentration of effort, deliberate procurement preference, deliberate research programme funding, and deliberate university-industry coupling of the kind that the country has, in patches, occasionally achieved. We do not in this paper specify the instruments. We record only that the talent question is, at root, the question of whether the country will choose to do AI work here, on Australian problems, at Australian scale, for the Australian commonwealth, or whether it will continue to subsidise the training of researchers who will then take their training abroad, for use against Australian commercial and strategic interests. The choice is, as always, ours, but it is also narrowing.

13. The Governance Layer

The governance layer is the layer at which the country exercises legal and institutional control over what the rest of the stack is permitted to do. It is the layer at which sovereignty becomes operational rather than merely structural; a country may own the energy, the compute, the models, the data and the talent and still fail to be sovereign in any meaningful sense if the rules under which those assets operate are written elsewhere or are unenforceable here.

The governance discussion in Australia today is dominated by two threads. The first is the negotiation of an "AI Safety Act" or equivalent statutory regime, modelled to a greater or lesser degree on the European Union's AI Act. The second is the integration of AI compliance into existing privacy, consumer-protection and anti-discrimination law. Both threads are, in our view, necessary but seriously insufficient. They address what AI may do to Australians; they do not address what AI may do for Australia, nor under what authority. We propose three additions.

13.1 Sovereign Audit

The first addition is a regime of sovereign audit. Any AI system deployed at scale in Australia, and certainly any AI system deployed by the federal or state governments, by operators of regulated infrastructure, or in regulated professions (medicine, law, financial advice, education), should be subject to inspection by an Australian audit authority with statutory access to training data manifests, evaluation harnesses and operational logs. The authority should be staffed by Australian researchers and engineers with the technical competence to actually perform the inspection, not merely to receive certifications drafted by the inspected. The authority's findings should be public to the extent compatible with legitimate trade-secret protection, and adverse findings should carry the regulatory consequences (cease-and-desist, mandatory remediation, suspension of operating authority) that the seriousness of the system requires.

We do not demand the weights themselves. No frontier provider will surrender them. Manifests, evaluation harnesses and operational logs are sufficient to audit a system's behaviour; and where the system is an open-weight model hosted on sovereign infrastructure, the weights are inspectable by construction. This is still a substantial demand of the foreign providers, and we are aware that it will be resisted. However, a foreign provider that will not submit its system to Australian inspection on terms set by Australian law has, by that refusal, declared that it is not willing to operate sovereignly in Australia. The country may then make the corresponding decision: whether to permit the operation under reduced terms (consumer-only, non-regulated workloads), whether to require operation only through a sovereign re-deployment (the open-weight model hosted on Australian infrastructure under audit), or whether to prohibit the operation entirely. None of these decisions is precluded by the current Australian legal regime; the deficiency is in the absence of the audit mechanism that would allow the decision to be informed.

13.2 Sovereign Procurement

The second addition is a regime of sovereign procurement. Government AI workloads, of which there are many growing rapidly, should be procured under explicit sovereignty criteria: where the work is sourced, where the data is held, who has root credentials, what jurisdiction's law applies, and what audit regime applies. The criteria should be transparent, technology-neutral, and weighted against price-and-performance under a published rubric. Where a sovereign offering meets a defined performance threshold, the government should preferentially procure from it; where no sovereign offering meets the threshold, the procurement of a foreign offering should be accompanied by a published statement of the residual sovereignty risk and the steps taken to mitigate it.

This is not industry policy in the protectionist sense, rather it is procurement policy in a directed-customer sense. The government is the largest single buyer of AI in the country and its procurement decisions, made consciously, are sufficient to seed a sovereign supplier base. The same approach has, over the last fifty years, sustained the Australian shipbuilding, ammunition, defence-electronics and pharmaceutical industries; we propose only that it be extended to AI, which is now an industry of comparable strategic significance.

13.3 Public Reporting and Parliamentary Oversight

The third addition is public reporting and parliamentary oversight of the sovereign AI programme as a whole. The programme is large enough, expensive enough, and consequential enough to require continuing democratic supervision, on the model of the parliamentary committees that supervise the intelligence services and the defence forces. We propose a Joint Standing Committee on Sovereign AI, with security clearance sufficient to receive classified briefings on capability and adversary activity, with statutory access to the audit findings, and with a mandate to report annually to Parliament on the state of the programme. This is the institutional fact that converts the programme from an executive-branch initiative that is vulnerable to the three-year electoral cycle and the personalities of individual ministers, into an enduring national capability.

The governance layer, in summary, is the layer at which the country either takes the programme seriously or does not. The other layers can be built; they can be funded; they can be staffed; and they will, without governance, drift into the same patterns of foreign capture that the present arrangements already display. Governance is not the icing on the cake. Governance is the structural reinforcement that allows the cake to bear weight.

14. The Alignment Layer

The alignment layer is the layer at which the model's values, priorities and behavioural constraints are encoded. It is the layer at which the policy of the country becomes, quite literally, the instruction set of the machine.

The popular discussion of AI alignment is dominated by two preoccupations. The first is safety alignment in the sense of the global AI-safety research community: ensuring that very powerful future systems do not, through misspecification or instrumental convergence, take catastrophic actions against human interests. We treat this preoccupation seriously and we accept that a fraction of the sovereign research programme should be allocated to it. The second is behavioural alignment in the sense of content moderation: ensuring that models do not generate output that is illegal, defamatory, sexually exploitative of minors, dangerous in operation, or otherwise unfit for general distribution. We treat this preoccupation seriously and we accept that the sovereign utility must implement it competently.

Neither preoccupation is, however, the alignment question that motivates the present paper. The alignment question we raise is sharper. Whose values are encoded in the model the country uses? When an Australian asks the foreign model a question about Australian history, Australian law, Australian land use, Australian electoral politics, Australian foreign policy or Australian cultural practice, the answer the model returns reflects the alignment training the model received, performed by foreign trainers, against priorities set by foreign management, in accordance with foreign legal and cultural norms. The answers are usually defensible. They are not, in the strict sense, Australian answers. They are foreign answers about Australia, with whatever residual systematic biases a foreign training process leaves in place.

This matters in three ways.

The first is a matter of accuracy: foreign-trained models systematically know Australian facts less well than they know facts about the foreign training-data dominant cultures (American, English, increasingly Chinese). The deficit is not large, but it is real, and it affects the model's competence on Australian problems in proportion to how Australian the problem is.

The second is a matter of priority: the foreign model is aligned to priorities chosen by the foreign trainer, and these priorities are not always congruent with Australian priorities. A model whose alignment training prioritises American First Amendment norms will respond differently to questions about Australian defamation law, the Racial Discrimination Act or the Australian electoral integrity provisions than a model aligned to Australian norms would respond. A model whose alignment training prioritises Chinese political norms will, in well-documented and visible ways, respond differently again. Neither misalignment is malicious. Both are real.

The third is a matter of cultural transmission. A population that interacts daily with a model aligned to a foreign culture gradually adjusts its own thinking to that alignment. The effect is not large per interaction but it compounds across millions of interactions per day across the population, and it is the same mechanism by which Hollywood, the foreign-language broadcasts of the BBC and the foreign news wire services have, for a century, shaped what Australians find familiar. The shaping is not sinister; it is mechanical. The country that does not have its own model will, in the long run, find that its citizens think a little more like the people who built the foreign one.

The alignment layer therefore has a sovereignty dimension that no amount of safety or moderation work can fill. The companion paper compresses the point into a single line: a model's alignment is its flag. Australian alignment requires Australian alignment training: priorities chosen by Australian institutions, performed by Australian alignment researchers, embedded into models hosted on Australian compute. The work is non-trivial, it requires constitutional convention-style decisions about what, exactly, the country wants to align to, but it is the work the sovereign programme cannot avoid. Here, as in §13, the question is not whether the work will be done. It will be done; the only question is by whom, and to which polity's priorities. We propose that the polity be ours.

15. Reconstruction: The Programme

We have set out, layer by layer, what each part of the sovereign AI stack must consist of. This section assembles the layers into a single integrated programme. We are deliberately brief; we are not, in this paper, costing the programme to two decimal places, which is a separate exercise. We are giving the reader who has followed the layered argument a single page in which to see the whole.

15.1 What Is Built

The programme builds, in concurrent rather than sequential streams over a 1–3 year horizon:

• A sovereign baseload energy capability comprising firmed renewables, gas peaking and a nuclear pathway (small modular reactors in the near term, with thorium molten-salt as a longer-term option), with the statutory prohibitions on nuclear power repealed, a credible siting decision made, and a first-of-kind reactor in the licensing process within the parliamentary term.
• A sovereign hyperscale operator owning at least one and ideally two data-centre footprints of leading-edge compute, sufficient to absorb federal AI workloads and to host the open-weight model deployments described below.
• A sovereign AI utility offering free-at-point-of-use AI to every Australian, hosted on the sovereign operator, running open-weight frontier models fine-tuned and aligned under Australian governance.
• A distributed sovereign neocloud network of substation-co-located regional data centres, enterprise nodes and edge deployments, running the distillation pipeline of §10.6 and hosting the compact specialist models close to the industries that use them, fed by the subsidised compute harvest of §10.7.
• A specialist-model programme in mining, agriculture, environmental management, mineral exploration, marine science and remote medicine, each delivering a domain-specific Australian-originated model, distilled from the hosted open-weight giants and continuously trained on Australian data, exported globally.
• A sovereign data programme with three components: a stronger personal-data regime, a deliberate operational-IP protection strategy via the utility, and a national strategic-data catalogue feeding the specialist-model programme.
• A talent retention and repatriation programme structured around the work the rest of the programme creates, supplemented by direct research funding and university-industry coupling.
• A governance regime comprising sovereign audit, sovereign procurement preference, and parliamentary oversight via a Joint Standing Committee on Sovereign AI.
• A sovereign alignment capability producing Australian-aligned model deployments from the open-weight base models, with priorities set by Australian institutions and embedded by Australian alignment researchers.

15.2 What Is Not Built

We do not propose, and we are explicit about not proposing:

• A from-scratch Australian frontier model. The economics do not justify it; the open-weight alternatives are sufficient; the resources are better deployed elsewhere in the stack. The national foundation models of §10.6 are a different artefact: mid-size, domain-anchored, and affordable on neocloud hardware. Those we build.
• A monopoly AI provider. The sovereign utility is the floor, not the ceiling. Private deployment, foreign provision and self-hosted alternatives all remain available, and the Hayekian search of §5 requires that they do.
• A UBI-style transfer programme. The augmentation of the citizenry through the utility, combined with the work the rest of the programme creates, is the alternative to UBI, not its complement (§5.4).
• A protectionist trade regime. Sovereign procurement preference is not a tariff. The country remains open to foreign technology, foreign investment and foreign expertise; it merely insists that the operating layer be Australian where strategic considerations require it.

15.3 Sequencing and Funding

The sequencing constraint is set by the energy layer: firm baseload, and nuclear in particular, takes time. The remainder of the programme can and should proceed in parallel rather than wait. The compute, model, data, talent, governance and alignment layers can each begin within months and reach meaningful operational scale within the parliamentary term. The energy layer sets the long horizon; the other layers set the near-term deliverables.

The funding is significant but not large by national-infrastructure standards. The compute footprint is comparable in capital cost to a single major freight project. The energy programme is, by 2025 prices, comparable to a frigate programme. The utility, the specialist models, the talent and governance lines are smaller still. The cumulative cost is, in the order of magnitude, a fraction of the annual defence budget; the cumulative return, on any plausible accounting of preserved sovereignty, is many multiples of the annual defence budget. The case is not financially difficult. It is politically difficult, because it requires deliberate choice in a system that has, for a generation, preferred drift.

15.4 What Happens Without It

We have set out, in §§1–6, what happens to a country that does not build a sovereign AI capability during the speciation window. We will not repeat the argument. We summarise it in a single line: the country becomes a customer of the new productive species, indefinitely, on terms set elsewhere, with its commonwealth claim unenforceable, its operational intelligence leaking abroad, its strategic data capitalised by foreign holders, its talent producing for foreign laboratories, its alignment performed by foreign trainers, and its capacity to defend any of these conditions diminishing year by year. The trajectory is not catastrophic; it is gradual; it is, in the absence of a deliberate choice to do otherwise, almost certain.

The end state has a name, because we have seen it before. Economies whose wealth comes out of the ground rather than out of their people stop needing their people, and everything else follows: the government is funded by royalties rather than by wages, so it answers to the diggers and not to the wage-earners. The Congo holds some of the richest mineral ground on Earth and its people are among the poorest. Venezuela sits on the largest proven oil reserves in the world. Equatorial Guinea has, on paper, a rich country's income per head, and nearly all of it belongs to one family. The instructive exception is Norway, which struck the same oil as everyone else and remains one of the best places on Earth to live, because Norway owned the oil, in common, and banked the proceeds for its people. The curse was never the resource; it is who owns it. AI is the next resource. The question this paper answers is whether Australia holds it the way Norway held its oil, or leases it the way the Congo leased its copper.

The window in which the choice can be made remains, for the moment, open. We estimate it at 1–3 years, and the upper bound is generous. After it closes, the residual scope for sovereign action will be confined to managing the consequences of decisions taken elsewhere. This paper exists because the window is open and because we believe the country has not yet noticed that it is closing.

16. Conclusion: AI as a Peace Technology

We have argued, across the preceding fifteen sections, that Australia must build a sovereign AI capability. We have argued the case in productive terms, in commonwealth-distributive terms, in Hayekian-search terms, in continental-development terms, and in stack-engineering terms. The case stands on each of these legs separately and more firmly together. We close with a different argument, deliberately reserved to this final section, because it is the argument that lifts the programme out of the category of national interest and into the category of human interest.

16.1 The Demechracy Premise

We have explored elsewhere, in our paper Demechracy: Beyond Dunbar's Number, the proposition that the structural cause of human inter-state conflict is the cognitive limit of the political class. Dunbar's number, the empirical observation that humans can sustain meaningful relationships with approximately 150 other persons, implies that any human governance system above that scale must operate by abstraction, and abstraction in political systems takes the form of categorisation. Other peoples are processed as categories, not as persons; categories admit of stereotype, prejudice and the failure modes of out-group hostility; and out-group hostility, accumulated across the political class of one nation interacting with the political class of another, is the cognitive substrate of war. Wars are fought, in the strict sense, between abstractions held by the leadership classes of the contending polities, while the populations in whose name the abstractions are held bear the costs.

This is a bleak proposition, which has, until now, been unavoidable. The cognitive constraint is human, the human is the political decision-maker, and so the constraint has been baked into the architecture of all governance, at all scales.

The novelty of the present moment is that the political decision-maker need no longer be human at every node. A sovereign AI, suitably constructed, is a governance instrument unconstrained by Dunbar's number. It can, at least in principle, hold the full register of every citizen as a person rather than as a category; it can, at least in principle, model the people of an adjacent polity as persons rather than as a category; it can, at least in principle, perform the kind of detailed, person-by-person, situation-by-situation reasoning that human leaders, however well-meaning, cannot perform at scale. The constraint that produced category-politics at the foundation of the war-making engine is, for the first time in human history, available to be relaxed.

We do not claim that this relaxation will, by itself, produce peace. We claim that it removes one of the structural causes of war, and that no other cognitive technology in human history has been capable of doing so.

16.2 The Implication for Sovereign AI

Sovereign AI, constructed with this implication in view, is not merely an economic instrument or a national-security instrument. It is a peace instrument. The country that builds a sovereign AI capable of treating its citizens as persons rather than categories, and of treating the citizens of its neighbours as persons rather than categories, is the country that has begun to dismantle, in its own institutional architecture, the cognitive substrate of the wars it might otherwise have fought.

This is not a pacifist argument. We are not proposing the unilateral disarmament of the Australian state or the abandonment of its alliances. The decision-theoretic environment in which Australia operates is one in which other states have not yet built such instruments and may not build them for some time, and the Australian Defence Force, the AUKUS arrangement³⁰ and the conventional apparatus of national security may remain necessary for the period in which we are still surrounded by polities operating on the older cognitive architecture. The point is not that Australia disarms. The point is that the sovereign AI programme is, simultaneously, a defensive necessity in the short term and a peace technology in the long term, and that no other technology in the country's reach has both properties at once.

16.3 Demechracy as Export

A sovereign AI built on these principles becomes, over the longer term, an exportable institutional technology. Just as the Westminster system, the trade-union movement, the public-health model, the rule of law and the secret ballot have been adopted by other polities after demonstration in originating ones, the Demechratic model of AI-augmented governance can be demonstrated in Australia and adopted elsewhere by polities that find it persuasive. The export is not coercive. The export is voluntary, by demonstration. A country that, fifty years from now, governs itself with a Demechratic AI on the Australian model is a country that has, voluntarily, removed one of the structural causes of war from its own architecture. An international system in which a meaningful fraction of polities have done so is an international system in which one of the perennial sources of conflict has been, by deliberate institutional design, eroded.

The argument is therefore not that Australian sovereign AI produces world peace. The argument is that Australian sovereign AI is the prototype, demonstrable on Australian soil, of an institutional technology that, at scale, contributes to the structural conditions for peace. The country that builds the prototype carries the moral and intellectual leadership of the export. We are not the largest power and we will not be the largest power; but the institutional ideas with the largest effect on the twentieth century, universal suffrage, the secret ballot, public health, the welfare state, were not, in their origins, the work of the largest powers either. Small countries have, repeatedly, built institutions of global consequence. Australia is in a position to build another.1

16.4 The Choice

The choice the country faces is, in the framing we have set out, not the narrow one of how to procure AI services for the public sector. It is the broad one of whether the country will build, in the closing window, the full-stack sovereign capability that the productive, distributive, search-theoretic, continental and peace-theoretic arguments converge on. The arguments are, in the end, one argument, albeit viewed from different directions. The country that owns its energy, compute, models, data, talent, governance and alignment is the country that has secured the productive base on which its common wealth depends, the augmentation by which its citizens remain in the new Hayekian search, the developmental capacity by which its land is held, and the institutional foundation on which a Demechratic and ultimately peaceful future can be built. If we do not do this we will have, by inaction, ceded each of these goods.

We do not believe our country should make the second choice, nor do we believe our country, properly informed, will make the second choice. The purpose of this paper is to inform, properly, our country that is in the process of making the choice in this short window where a choice is available.

The window is open, but it will not stay open.