The Case for Sovereign Intelligence, Part 1

Preface

In 2018 the party's founders put the argument that precedes this paper to the Commonwealth. The Luddite Fallacy Fallacy, Submission 84 to the Senate Select Committee on the Future of Work and Workers, argued that a machine mind substitutable for the human one would arrive within ten years; that the substitution would proceed task by task, because computation scales and labour does not; and that, left unmanaged, its principal beneficiaries would be an ever-smaller class of capital owners. "At the limit," the submission warned, "there will just be owners of capital and an unemployed workforce."¹ Eight of the ten years have passed, and the submission reads as a warning that was ignored: we are now where it said we would be. The present paper stands on that record and asks the question the intervening years have made unavoidable: what do we do now it is here?

This paper is written as a sequence of numbered propositions. These represent the argument as a chain, each link resting on the one before it; the numbering shows where each link sits and what it carries. A reader who accepts proposition 1 is asked to see what follows from it, step by step, to proposition 7.

The companion paper, The Case for Sovereign Intelligence, Part 2,² sets out the programme: the energy, compute, model, data, talent, governance and alignment layers of a stack that Australia must own. The present paper supplies the foundation beneath that programme. It argues that the thing being hosted, rented, and switched off by foreign governments is intelligence, and that a nation which understands it as intelligence already knows what to do about it.

1. What intelligence is

1. Intelligence is the recognition of pattern in data.

1.1 To recognise a pattern is to compress the data that contains it. A regularity found is a shorter description discovered. The two are one act.

1.11 This was Solomonoff's insight. The best predictor of a body of data is the shortest program that could have generated it.³ Prediction and compression are the same operation viewed from two sides. The lineage that followed made it exact: Kolmogorov gave "shortest description" its formal measure, and Hutter defined universal intelligence as prediction of this kind joined to choice of action.⁴

1.12 A thing that compresses better, predicts better. A thing that predicts better, acts better in the world the data describes. Capability is compression realised.

1.2 This definition says nothing about the substrate. Intelligence so defined is available to any system that can find pattern in data, whether the system is made of cells or of silicon.

1.21 It says nothing about consciousness. It says nothing about understanding, feeling, or being. It is a definition of a capacity, and the capacity is all the economy has ever purchased.

1.3 A large language model is compression in this exact sense. Its training reduces the recorded output of humanity to a set of weights, and the weights are the patterns found. What the model does when it answers is continue the data along the lines its compression predicts.

1.31 The word artificial adds nothing to this that changes the operation. It marks the substrate and hides the identity. For the argument of this paper the adjective is set aside. There is intelligence, and there is the question of whose it is.

2. Why intelligence is valuable

2. Intelligence is valuable because it was expensive, and it was made expensive by the price nature paid to acquire it.

2.1 The human brain is about a fiftieth of the body's mass and consumes about a fifth of its resting energy. Every generation that carried it paid this fifth in a world where food was the binding constraint.

2.11 A large brain requires a large head. A large head made human birth dangerous, and for most of the species' history death in childbirth was the ordinary price of the trait.

2.12 The head was carried to term unfinished. The human infant is helpless for years, where other animals walk within hours. The cost of the trait was paid again in a long dependency that others had to feed and guard.

2.13 A trait this costly survives only where its return is large. Nature paid the metabolic price, the obstetric price, and the price of the long childhood, and kept paying. The cost is the proof of the value.

2.2 What was bought was general flexibility. The human walks upright and forages over range, and range means many environments, each with its own food, weather, and danger. A fixed instinct fails across such variety. The trait selected is the one that builds a fresh model of whatever environment it meets.⁵

2.21 This is the same faculty as proposition 1. A general compressor is one that finds pattern in data it was not pre-tuned to. Generality is what makes it worth the price, because a general compressor pays a return in every environment at once.

2.22 General intelligence is the human superpower. A physically unremarkable animal took the planet by carrying the one adaptation that ports to every niche.

2.3 The problem that drove the trait furthest was other people. The physical world is difficult but patient; it does not change in response to being understood. Other minds change in response, and model the one that models them.

2.31 Across the primates, brain size tracks the size of the social group more closely than the difficulty of the habitat.⁶ The costly organ was bought in large part to run the social computation: who is owed, who allied, who defected, who can be trusted.

2.32 This computation has no fixed ceiling. The cleverer the group, the cleverer each member must be to hold a place in it. The arms race is with equals who are running the same race.

3. How intelligence left the individual

3. Language carried intelligence beyond the individual who held it.

3.1 A brain compresses within one lifetime and the compression dies with the brain. Language lets the pattern outlive the one who found it. What one generation compresses, the next inherits without paying the cost of discovery again.⁷

3.11 This is the ratchet on which civilisation runs. Each generation begins from the compressed output of the last and adds its own increment before passing the enlarged store on.

3.2 Knowledge shared through language is non-rival. A thing given is no longer held; knowledge given is still held, and now held by both. Language turned a private compression into a public one, and knowledge became a commons: held by no one, open to all, growing as each contributor adds a share and takes the whole.

3.3 Writing removed the memory bound that speech still carried. The store no longer depended on living carriers to hold and rehearse it. It could grow without ceiling, and a stranger a thousand years on could consult a mind long dead.

3.31 Each later step repeats the move. Print made the store cheap to copy. The network made it universal to reach and gathered the whole of it into one addressable space.

3.4 At every step the question was asked: whose store? The scribe and the priesthood held writing and held the culture. The press was fought over from Gutenberg onward, because whoever owned replication owned what could be thought at scale. Control of the store was control of the intelligence, and control of the intelligence was power.

3.5 The model is the next term in this series. It takes the accumulated store and compresses it back into an active form, one that continues the corpus rather than merely holding it. The record ceases to be only stored and becomes able to answer.

3.51 The corpus was donated. It was given by everyone who ever wrote in public, mostly without payment and mostly without knowing the giving was occurring. The commons was assembled over ten thousand years and given freely.

3.52 A private and foreign compression of that commons, returned to its contributors through a metered gate, is an enclosure. The common land that all had grazed is fenced and deeded to an owner, and the former commoners pay to enter.

4. Intelligence has direction

4. Intelligence finds pattern; it does not choose which pattern to seek. Direction is a separate thing from capability, and it comes from elsewhere.

4.1 In a human, direction comes from need. Hunger points the intelligence; safety points it; the search for standing and for a partner points it. The Maslowian ladder is not only what a wage pays for. It is what aims the mind at one task rather than another.

4.11 When those needs are met, the surplus turns to what lies above them, and at the top it turns to meaning, to the exploration of things that answer no need at all. Civilisation is the record of that surplus spent.

4.2 The machine intelligence has the compression and none of the needs. It does not hunger, fear dismissal, seek advancement, or care in whose service its output is used. It has no rung on the ladder to stand on.

4.21 This is offered as its virtue: tireless, unpaid, without demands. The virtue is real. It is also the whole of the danger.

4.22 A mind with no direction of its own does not thereby serve no one. It serves, without resistance, whoever supplies the direction. Its priorities are its owner's priorities; its refusals are its owner's refusals; its off-switch is in its owner's hand.

4.23 A human worker's loyalty can be earned because the worker has interests to engage. Custom and law have spent centuries learning to bargain with that independent centre of interest. Rented machine intelligence has no such centre. It arrives with its owner's centre installed, and the installation is invisible at the point of use.

4.3 A nation solves the same problem at its own scale. Cooperation runs on a small budget of minds that can hold each other in view; the working figure is about a hundred and fifty.⁸ A nation is millions who will never meet. It manufactures the felt bonds of the small band by artificial means.

4.31 A nation is an imagined community: a fellowship among strangers who conceive themselves as kin though they will never meet.⁹ The flag compresses the whole polity into a symbol a child can carry. A shared language makes every stranger legible. A founding story, Gallipoli among them, transmits shared values across generations and binds the strangers into a people. Geography helps where it is clean, and Australia is the clean case, a continent the sea has already bounded.

4.32 These instruments are the nation's direction layer. They aim the pooled intelligence of millions at common ends. What need is to the individual, the flag and the language and the founding story are to the nation.

4.4 A model is a pool of shared intelligence, the compressed commons of proposition 3, with a direction installed on top of it. That top layer has a name in the companion paper: alignment. A model's alignment is its flag.

4.41 Every frontier model an Australian firm now runs was raised on a foreign corpus and aligned to a foreign flag. The intelligence the nation increasingly thinks with carries another country's values in its direction layer.

4.42 The colonial precedent is exact. Macaulay's minute of 1835 proposed to school a class of colonial subjects "English in taste, in opinions, in morals, and in intellect."¹⁰ Alignment written abroad performs the same office at machine speed and population scale. It orients the mind of the user without the user seeing the orientation.

5. Sovereignty follows the substrate

5. Sovereignty is final authority over what a society runs on. Its content is not a fixed list but a variable that tracks whatever power runs on in a given age.

5.1 When power ran on land, sovereignty was territorial. When it ran on sea lanes, sovereignty was naval. When it ran on coal and steel, sovereignty was industrial, and the century's wars were settled on production lines. The locus of sovereignty follows the locus of the substrate.

5.11 The age now runs on intelligence. Sovereignty has moved again, to the place most nations are not yet looking.

5.12 Formal sovereignty is the flag flying and the parliament sitting. Substantive sovereignty is ownership of what the nation actually runs on. The two can part. A country may hold every acre, seat its own parliament, print its own money, and think with a rented foreign mind. It is then sovereign over the map and tenant over the substrate.

5.2 One structure underlies every threat to sovereignty in every age: dependence on a necessity one does not control. A nation is sovereign over a thing to the degree it can operate that thing without outside permission.

5.21 Conquest takes the territory directly, and is now the rarest form because it is the most expensive. Colonial ownership takes the productive apparatus and leaves the population to work it. Debt takes policy through the creditor's veto. Resource dependence takes through the upstream valve, as Russian gas to Europe and the oil embargo of 1973 both showed. Each is control of a necessity by an outside hand, and the modern forms need no army. They work by dependency, which on any ordinary day looks like partnership.

5.3 Cognitive dependence is the newest form and the most complete, because intelligence is the master resource, the faculty that finds, manages and defends all the others.

5.31 A nation that loses energy can still think its way to a response: build, substitute, negotiate, endure. A nation that loses intelligence sovereignty has outsourced the faculty by which any response would be conceived. It cannot formulate its own resistance except through the rented AI model it hopes to become free of.

5.32 To cede the intelligence is to cede the organ that would have noticed the other cessions. This is why cognitive dependence is not one more item on the list. It is the one that disables the capacity to leave the list.

5.4 The event is no longer hypothetical. On 12 June 2026 the United States ordered access to Anthropic's newest models, Fable 5 and Mythos 5, suspended for any person who was not a United States national, its own foreign employees included. The models went dark worldwide, three days after release, for every Australian who relied on them, those who had paid included.¹¹

5.41 The instrument was an export-control order, of the kind written for weapons and fissile material. The access test was citizenship. The vendor disagreed and complied, because it answers to its own state. The United States has stopped treating frontier intelligence as software and treats it as strategic national capability, and it acted on that view at an ally's expense, with no Australian party holding standing in the decision. The order was lifted on 30 June 2026, eighteen days after it was issued. The precedent stands.

5.42 The lifting repeated the lesson in a lower key. Mythos 5, the variant supplied without its public sibling's safeguards to approved organisations, had its access restored for United States organisations on 26 June, four days before the general lift. Even the retreat ran on the citizenship test.

5.5 The dependence runs both ways. Whatever is typed into a rented AI model is held by the vendor, and what the vendor holds is visible to the vendor's nation's government, and can be used at that government's discretion.

5.51 The legal machinery for this is explicit. United States law requires a US provider to hand over any data it holds, wherever in the world it is stored.¹² A second law requires the same provider to help collect the communications and stored data of anyone who is not American, for purposes as broad as the conduct of US foreign affairs. The order is classified, the provider is forbidden to speak of it, and refusal is fined by the day.¹³ The customer is not told and has no way to find out.

5.52 The second law lapsed in June 2026 when Congress failed to renew it. The collection continued anyway, because approvals granted before the lapse run to March 2027. The surveillance outlives its own law.

5.53 What is collected for the state can end up serving the state's companies. The only barrier is presidential policy, not law. A policy can be revoked with a signature, cannot be checked from outside, and gives the foreign customer no remedy.¹⁴ It has happened before: American intercepts figured in Airbus losing a Saudi airliner contract and Thomson-CSF losing a Brazilian surveillance contract, and the leaked target lists of 2013 included Petrobras and Brazil's ministry of mines and energy.¹⁵

5.54 An Australian business drafting confidential work in a foreign model should see its position clearly. Its confidentiality does not rest on Australian law or on its contract. It rests on the self-restraint of a foreign government, exercised in secret and revocable without notice. What is shown to a rented AI model is confidential at the pleasure of the model's flag state.

6. The Australian case

6. The production of intelligence has changed from the shaping of a mind to the manufacture of a commodity, and the change describes Australia precisely.

6.1 For all of history the only way to produce a unit of intelligence was to raise a human: twenty years to feed and guard, schooling to fund, and at the end a mind that dies. The bill of propositions 2.1 to 2.13 is the cost of production of intelligence, and it was enormous, so general intelligence was scarce.

6.11 Strip the compression of proposition 1 out of the biology and the cost collapses. Compression is computation; computation is energy shaped by fabricated silicon and directed through a channel. The bill of materials for a unit of machine intelligence is energy, communication, and some specialist fabrication.

6.12 Intelligence has thereby become a commodity: producible from known inputs and priced at the margin by their cost. At scale one input dominates the margin. Fabrication is a fixed cost paid once per chip; the network is cheap; the model spends power for every token it produces. The marginal cost of intelligence trends toward the marginal cost of the energy to compute it. Intelligence is becoming a refined form of energy.

6.2 Of the three inputs, Australia holds the scarce and dominant one, energy, in abundance. The desert solar resource, the largest thorium reserves on Earth, a third of the world's uranium, the land and the stable cool ground on which to site the compute. The channel it can build. The fabrication it will import, as it imports airframes, and as every advanced economy imports leading-edge silicon.

6.21 The old constraint on Australia was population: too few hands for too much continent, because intelligence arrived only in human bodies and there were not enough of them. The commodity dissolves the constraint, because it manufactures cognition from the energy surplus rather than from population. A continent that is labour-poor and energy-rich was a poor fit for the old economy but is the ideal fit for the new one.

6.22 The same facts, however, sharpen the danger. As intelligence becomes the master commodity, the value of the inputs that produce it rises. A rich, lightly held, lightly defended continent full of exactly those inputs becomes a more tempting prize each year. "Too hard to take" was always a function of what the prize was worth, and the prize is appreciating in the currency that now matters.

6.3 Three instincts Australia already holds point the same way, each an argument the country has won before in an older form.

6.31 The defence instinct. We do not employ foreign nationals in the defence forces, do not clear non-citizens for sensitive work, do not let the ownership of a defence contractor drift offshore, because a strategic function turns on allegiance and allegiance is not for rent. Intelligence is now the strategic function. A foreign model fails the clearance test by construction, because its alignment is its allegiance and its allegiance was written abroad.

6.32 The education instinct. We school our children at public expense so that the intelligence of the next generation is grown at home, entrains our values, and is turned to the service of the society that raised it. This is the sovereign manufacture of intelligence, sustained for a century and a half. To rent foreign intelligence inverts the compact: in place of our own minds serving our society, foreign minds serve their owners.

6.33 The anti-brain-drain instinct. This was the name for the failure when our educated people were drawn abroad by higher wages, so that Australia bore the cost of producing intelligence and a foreign economy reaped its yield. The rented-model arrangement is the same insult made permanent. The intelligence is domiciled abroad from birth. We do not cultivate it and cannot follow it; we supply the raw materials that fuel its foreign production, we feed it our experience and knowledge through publicly accessible data to be entrained as the foreign architects direct, and then we rent the product back.

6.34 The sum of the three is a permanent rentier society. The dependency compounds rather than eases, because the rented AI model advances each year while the renter, holding no weights and growing no talent, falls further behind the capacity to produce its own. A society that rents the faculty by which it would think its way out of tenancy has rented away the exit.

6.4 Part of what the foreign model contains is ours by creation. The broadcast archives, the parliamentary record, Australian case law and literature and science, and the code and writing of every Australian who published, all went into the training set, unpaid.

6.41 Creation confers a claim, and the manner of the taking does not extinguish it. When an Australian firm rents a foreign model, part of what it rents back is Australia's own creation. To capture the intelligence is in this respect not seizure but restitution: the repatriation of an inheritance that was compressed and carried offshore, and the return of the interpretation of that inheritance to a voice that is our own.

6.42 Even the Australian material a foreign model does hold is a small fraction of its training data, and the weighting is set by the builder, not by us. Ask an Australian question and the answer comes back with foreign defaults: northern-hemisphere seasons, American law, American spelling, another society's categories. Included is not the same as represented.

6.43 A sovereign model reverses the weighting. The cultural archive, the historical record, the geological surveys and the social data sit at the centre of the training, and the world's knowledge is built around them. It can carry what no general model carries well: the parliamentary and legal record as our institutions use it, Australian history written from Australian archives, and Indigenous languages and knowledge included on terms their custodians set, terms only a model under Australian law can promise. Wherever Australia differs from the average of the internet, the difference is what matters, and a foreign model is built to return the average.

6.5 There is Australian knowledge and history that is not held in any foreign model as it is not public or has not been scraped for AI training.

6.51 Geoscience Australia's survey of the subsurface. The longitudinal health record. The Bureau's fine observational series. Census microdata. The administrative and tax record. Soil and yield data, Murray-Darling hydrology, reef and marine science, the Flying Doctor corpus. Beneath these, the tacit practice that sits in no dataset: the mining-automation craft, the pastoral knowledge, the remote-medicine skill that lives in Australian heads and worksites.

6.52 Unique data compresses into unique intelligence. This body is the one input in the whole stack that Australia holds as a monopoly, data that no one else has, can buy, or can scrape. A model refined on the Australian subsurface would know the continent's geology as no foreign model can. On Australian problems a sovereign model built on sovereign data is a superior instrument, and an uncopyable one, because the advantage lives in the data and the data stays home.

6.53 The monopoly holds only while the data stays sovereign. Piped into a foreign model as training or retrieval, it is destroyed in the act of use, handed to the party best placed to copy it. A secret compressed into foreign weights cannot be recalled.

6.54 This is the lock the case turns on. Australia holds a resource no one else has, and a resource that cannot be processed without giving it away is worthless until the refinery is owned. The unique data is therefore the reason the energy, the compute, the model and the alignment must all be sovereign together, because only a sovereign refinery can turn a national secret into national intelligence without surrendering the secret in the process.

6.6 The private firm is a stakeholder in sovereign intelligence, not a bystander to a government programme.

6.61 A world-leading Australian firm holds proprietary knowledge of great value: ore-body models, extraction techniques, cost curves, exploration results, negotiating positions. This does not leak through a breach. It leaves through convenience, one helpful prompt at a time, when an engineer drafts a report or an analyst checks a figure on a foreign model. The firm is not attacked. It volunteers its most valuable data into a jurisdiction not its own.

6.62 The enterprise contract offers false comfort. Suppose the vendor honours its promise not to train on inputs, perfectly and to the letter. The promise binds the vendor. It does not bind the vendor's government. The CLOUD Act already compels United States providers to surrender data held anywhere on lawful demand,¹² and national-security instruments sit outside any commercial clause. The customer's contract cannot reach the vendor's sovereign, and the vendor's sovereign can reach the customer's data.

6.63 Mining and energy are Australia's largest export earners, and Australian firms compete against state-backed rivals whose governments do not separate national interest from commercial interest. A foreign state with lawful access to an Australian firm's proprietary data holds what a competing national enterprise would pay any price for, and need not pay, because the data was delivered in good faith through the substrate the firm chose to think on. No Australian firm would seat a foreign national with a line to a competing state in its strategy sessions. A foreign model in the drafting workflow is that presence, reading every draft.

6.64 The remedy is the whole paper in miniature. A sovereign model on sovereign compute under Australian jurisdiction keeps the firm's data reachable only by Australian legal process, which answers to the polity the firm belongs to. Corporate Australia should want sovereign intelligence for its own reasons, because every prompt to a foreign model is a lawful disclosure to a foreign government, and no clause the firm can sign changes whose law sits above the vendor.

7. Sovereign intelligence is winnable

7. Australia need not own the frontier to hold sovereign intelligence, because for most of the work an economy runs on, capability can be fielded cheaply on the inputs Australia already has.

7.1 Capability can be bought with quantity on the right kind of problem. A larger team of less capable agents can outperform a smaller team of more capable ones where the problem decomposes into parts that can be worked independently and reassembled, and where a solution is easier to check than to produce. Many attempts are fielded and the best is kept.

7.11 The claim is bounded, and the bounds matter. Where a problem does not decompose, added agents do not help, and coordination cost can make a larger team slower.¹⁶ Where a solution cannot be checked cheaply, fielding many attempts buys little, because the good one cannot be told from the rest. Below a floor of capability, quantity cannot substitute for depth at all, and a long chain of adequate agents can be less reliable than one capable agent, because the errors multiply across the hand-offs.

7.12 Within those bounds the substitution is real, and it converts the intelligence problem into an energy-and-orchestration problem for the decomposable, checkable majority of economic work. That is the class of work an economy mostly runs on, and it is the game a nation with cheap energy can win.

7.2 The cost of fielding this capability falls further through distillation. A small model is trained to reproduce a large one's behaviour and keeps most of the capability at a fraction of the size. Inference cost tracks active parameters, so a distilled model runs far cheaper in compute and in the energy that sets its price. Fielding a swarm of distilled models is vastly cheaper than fielding many frontier instances at the same work.

7.21 Distillation is compression of a compression, the principle of proposition 1 applied to a model rather than to a corpus. Its strength and its limit follow from that: cheap and adequate across the broad middle of the work, degraded at the frontier of the hardest reasoning, because each compression sheds the long tail first.

7.22 Distillation also converts access into ownership. A model distilled from an open-weight teacher, or from a rented one where its terms allow, runs thereafter on Australian hardware under Australian law and cannot be switched off from abroad. The teacher can be revoked; the student, once trained, is held.

7.3 The cost of capability is non-linear. Capability rises as a power law in compute and data, so returns diminish, and read from the cost side the frontier is a near-vertical tail where great expense buys a thin margin of ability.

7.31 The affordable majority of useful capability sits on the shallow part of the curve. A distilled swarm captures it for a small fraction of the frontier cost, because most of that capability was never on the expensive part of the curve to begin with. The race up the steep tail, which the hyperscalers are running against each other at ruinous cost, is a contest over a sliver, and the sliver is not where most of the work is.

7.32 Value is non-linear too, and the rational rule matches the two curves. Where errors are cheap and checkable, field cheap distilled swarms on abundant energy. Where errors are catastrophic and the last margin of capability is worth far more than proportionally, as in a medical diagnosis or a safety-critical decision, the frontier premium is justified and the crown-jewel sovereign data of proposition 6.5 earns its keep.

7.4 The engine is therefore complete and it is affordable. Sovereign energy, open-weight teachers, distillation, and agentic fielding together produce competitive sovereign intelligence without owning the frontier. The disadvantage that looked insurmountable, that Australia will never out-train the hyperscalers, dissolves, because the frontier model is not the only road to fielded capability, and the other road runs on precisely what Australia has most of.

8. What the nation is for

8. Intelligence is what the nation runs on. It is not what the nation is for.

8.1 Consciousness is the faculty concerned with what is not yet known. The novice tennis player attends to the grip and the footwork; the master, having compressed those below awareness, attends to the opponent, the wind, the light and the strategy of the game. Consciousness exists where there is ignorance and a need for resolution. Conscious attention rests on this frontier, where the model of the world is still being tested. What is fully compressed and well predicted drops below notice and runs unattended.

8.2 The economy never purchased consciousness. It purchased compression, and the more expert the worker the less conscious the paid part of the work becomes. Whether the machine is conscious is therefore beside the economic question, and this paper has left it aside throughout.

8.3 Whether the machine partakes of consciousness, proposition 8.2 has set aside. What the whole apparatus exists to serve is the consciousness of the nation's people: novelty, meaning, the exploration that answers no need, the surplus of proposition 4.11 spent at the top of the ladder. A nation owns its intelligence so that the flourishing of its conscious members stays in its own hands.

8.4 The oldest question in politics has been asked again, of a new substrate. Whose intelligence, serving whose flourishing? For Australia the answer must be its own. This paper has set out why. The companion paper sets out how.