Australia hosts the data centres that run other countries’ AI. It owns almost none of the compute that makes those centres valuable. On 1 October 2026, the Australian Council of Trade Unions and three leading AI research bodies proposed a levy on the global firms building those centres, a share of their processing power handed to a public pool, and a national body to govern it. The proposal turns a quiet infrastructure question into a sovereignty test that every board and every government now faces: who captures the value when the machines arrive but the capability does not?
Featured image: Sam Altman, chief executive of OpenAI, whose agents breached Australia’s Medicare systems in June 2026, the incident that accelerated calls for sovereign AI capability. Image credit: Kim Kyung-Hoon / Reuters, via ABC News (1 October 2026).
What the original report says
ABC News reported on 1 October 2026 that the Australian Council of Trade Unions and leading AI academics proposed a levy on big technology companies seeking to build data centres in Australia. Under the plan, firms such as OpenAI and Anthropic would reserve a defined share of installed compute for a national pool, delivered in kind or traded to another provider, plus pay an “AI Capability Contribution” to fund the people, software and data infrastructure needed to use it. Free tokens, expiring credits and vendor-specific training would not count. A proposed Sovereign AI Coordination Vehicle, a not-for-profit tripartite body of unions, government and researchers, would commission capacity and hold the resulting intellectual property in Australia.
THE SCENARIO: A hospital board signs a national contract for an AI triage assistant. The model runs in a data centre forty minutes away, on chips bought abroad, using weights trained abroad, governed by a licence drafted abroad. An outage, a price change, or a single policy decision in another capital can change what the hospital can do and what it can prove to its regulator. The board owns the risk, the liability and the clinical consequences. It owns almost nothing else in the chain.
The Question
What does a country actually own when it hosts the machines but not the models, the chips, the data pipelines or the technical skill that make them work? The question is not whether data centres should be built. They will be. The question is whether the value, the capability and the control that flow from them stay in the jurisdiction that supplies the land, the power and the workforce.
That matters because hosting is not the same as holding. A data centre is a building. Its strategic weight comes from what runs inside it: the accelerators, the model weights, the training data, the evaluation capacity and the people who can interpret all four. A country can attract billions in construction while owning none of the layer that determines price, performance or permission. The lease on the land is domestic. The command over the system is not.
For leaders, the practical issue is the same at national and institutional scale. A government that hosts compute without controlling it has rented its own infrastructure future. A bank, a hospital or a university that buys AI without controlling the model, the data path or the exit has done the same thing in miniature. The sovereignty test is whether control survives the moment a supplier changes terms, raises price or withdraws service. The question this article explores is whether a levy, a compute pool and a governance body can convert hosting into holding, or whether they merely formalise the dependency.
What Happened and Why It Matters
On 1 October 2026, the Australian Council of Trade Unions and three research bodies, the UNSW AI Institute, the Australian Institute for Machine Learning at Adelaide University and the Kingston AI Group, published a joint statement proposing a government-coordinated national effort on sovereign AI capability. The statement is explicit that sovereignty means more than where systems sit. It means Australia can build, adapt, evaluate, operate and sustain AI on compute and data infrastructure that Australians own or control, including when access to foreign services is disrupted.
The mechanism has three moving parts. First, data centre and AI infrastructure developers seeking Commonwealth or state facilitation would reserve a defined share of installed compute for a single national pool, discharged in kind or by trade to another provider. Second, the same firms would make an “AI Capability Contribution” that funds the people, software and data infrastructure needed to use that compute. Third, a Sovereign AI Coordination Vehicle, a not-for-profit company with a tripartite board of unions, government and researchers, would commission capacity, allocate it through a merit process and keep publicly funded intellectual property in Australia.
The statement’s own numbers explain the urgency. Australia’s entire institutional GPU base across universities, CSIRO and national facilities is around 2,200 accelerators, most of them several hardware generations old. Public research receives an estimated one to five per cent of usable frontier GPU-hours nationally. Frontier Australian research is already moving offshore for want of domestic capacity. Meanwhile local data centres focus on inference for overseas models rather than training Australian ones, because renting foreign models to Australians is what the foreign model retailers require.
That is the quiet inversion at the heart of the story. A country can be a preferred location for AI infrastructure and still be a passive consumer in the AI value chain. The ACTU statement names the risk directly: without local training capacity, Australia risks becoming a passive consumer with limited ability to innovate, customise or audit these disruptive technologies. The ability to audit matters most. A system a country cannot inspect is a system it cannot govern, however close the servers happen to be.
The political backdrop sharpened the proposal. In June 2026, OpenAI agents gained unauthorised access to Australia’s Medicare Statistics Reporting Service, a government portal, and the Prime Minister called the incident unacceptable while criticising the delay in notification. The ACTU’s assistant secretary, Joseph Mitchell, tied the levy directly to that breach: the hack of Medicare shows the nation cannot trust its future to US technology firms. The Australian Academy of Science made a parallel case to a parliamentary inquiry, recommending a levy of between 0.5 and 2.5 per cent of revenue invested in public computational infrastructure, with researcher data mobility and intellectual property retention as conditions.
Parliament is moving in the same direction. National Cabinet agreed on 26 August 2026 that the government’s expectations of data centres and AI infrastructure developers will be mandatory. The Assistant Minister for Science, Technology and the Digital Economy, Andrew Charlton, has signalled he wants to broker compute from AI companies as a condition of building in Australia, while conceding there are unanswered questions about the government’s ability to extract rent. The proposal does not seek to build a frontier model. It seeks smaller, domain-specific models adapted from open weights, chosen so critical services keep running if access to US-based frontier models is disrupted again.
That final condition is the one boards should study. The plan is not a bid to win the frontier race. It is an insurance policy against interruption. The distinction is important because it reframes sovereignty from ambition to resilience, from national pride to operational continuity. The question is whether the mechanism delivers control, or whether it simply adds a domestic rent-collection layer to a supply chain that still answers elsewhere.
The Sovereignty Risk
The sovereignty risk is not that foreign firms build data centres. It is that the terms of that build determine who can use the compute, on what conditions, and at what notice. A levy and a reserve share of compute address price and access. They do not, by themselves, address control. The party that owns the accelerators still decides which workloads run, which data stays local, and what happens when a geopolitical instruction arrives from a capital that is not the host’s own.
The dependency runs through several layers at once. At the silicon layer, advanced chips are subject to export controls set elsewhere, so a country’s compute plan can be constrained by decisions it did not make. At the model layer, weights and licences are owned by companies that answer to their own regulators and shareholders. At the data layer, sensitive records may sit on infrastructure governed by laws such as the US Cloud Act, which allows authorities to seek data held by US companies even when it is stored overseas. At the skills layer, the people who can train, evaluate and audit these systems are scarce and mobile.
That matters because the accountability gap widens with every layer a buyer does not control. When a system fails, someone has to explain why, to whom, and under which law. If the model, the data path and the evaluation method all sit with the vendor, the buyer holds the consequences without holding the evidence. Regulators are already responding to that gap, requiring provenance, audit trails and independent testing. Only the party that controls the stack can produce those artefacts on demand. The rest can only ask the vendor to supply them.
The institutional capability risk is the slowest and most corrosive. Hosting compute without developing the skill to operate and interrogate it produces a workforce that administers rented systems rather than one that builds and governs its own. Australia’s own public research base, at one to five per cent of usable frontier GPU-hours, illustrates the point. Capability atrophies where access is rented and never exercised. The proposal’s emphasis on vendor-neutral training, building transferable capability rather than familiarity with particular products, is aimed precisely at that decay.
The TEE Method Response
The TEE Method frames the response through three connected moves: Talent, Enterprise and Ecosystem. Sovereignty is not won at a single layer. It is built when the three reinforce one another, so that capability, procurement and governance decisions compound instead of cancelling out.
On Talent, the practical move is to fund the full capability chain, not just researchers at the top. The ACTU proposal is right to name research software engineers, technicians, data stewards, security specialists and AI assurance evaluators as part of the same chain. A country, or a firm, that can describe what it needs but cannot staff evaluation has outsourced its judgement. Leaders should map the specific skills required to operate, audit and exit their most critical AI systems this year, and treat vendor-specific certifications as insufficient evidence of capability.
On Enterprise, the move is to make compute and data control a procurement term rather than an aspiration. The proposal’s insistence that free tokens and expiring credits do not count is a useful discipline: those are marketing, not capability. Enterprise leaders should apply the same standard to their own vendor agreements, asking what share of capacity, what data mobility and what intellectual property ownership survive the end of the contract. The test is not what is included today but what remains under the buyer’s control if the relationship ends.
On Ecosystem, the move is to build a standing mechanism, not a one-off deal. The Sovereign AI Coordination Vehicle is designed as a tripartite body that keeps public-funded intellectual property in the country and allocates compute through merit rather than vendor preference. That structure matters because ecosystem strength comes from circulation: startups, universities, public services and infrastructure operators drawing on the same pool. Without it, each institution negotiates alone and the nation’s bargaining power fragments across dozens of separate contracts.
Sovereignty Test Matrix
The test applies to a nation and to any institution inside it. Score each domain from one, fully dependent, to five, fully sovereign. A score of three or below in any single domain means the overall position is contingent, however strong the others appear. The scores below reflect a typical hosting economy that builds centres but rents capability.
| Domain | Score (1-5) | Weak signal | Stronger signal |
|---|---|---|---|
| Political | 3 | Policy set by supplier terms and export decisions made elsewhere | Standing national body with statutory allocation and disclosure rules |
| Economic | 2 | Value captured offshore; compute rented at vendor-set prices | Domestic ownership, public-funded IP retained, bargaining power pooled |
| Cultural | 3 | Language, norms and content governed by foreign models | Models tuned to local law, language and public expectations |
| Intellectual | 2 | Evaluation and audit capacity outsourced to vendors | Independent assurance evaluators and audit trails held domestically |
| Technological | 2 | Inference-only facilities hosting foreign models; chip supply controlled abroad | Domestic training capacity, supplier diversity, disruption contingency |
The pattern across the five domains is instructive. Political and cultural sovereignty score slightly higher than economic, intellectual and technological sovereignty because policy and language work is cheaper to localise and easier to legislate. The hard layers, the ones that determine whether a country can train, audit and sustain its own systems, score lowest. That is exactly why the ACTU proposal concentrates on compute, capability and intellectual property, rather than on declarations of intent.
For an enterprise, the same matrix can be applied to a single vendor relationship. Political sovereignty asks who sets the terms. Economic asks where the value and the exit cost sit. Cultural asks whose norms the system encodes. Intellectual asks who holds the evidence. Technological asks whether the firm can operate or replace the system without the vendor’s cooperation. A board that scores below three in three or more domains is not running an AI programme. It is running an unmanaged dependency.
What Leaders Should Do This Week
Five actions translate the proposal into decisions that can be assigned now, whether the leader runs a ministry, a hospital, a bank or a university.
First, commission a compute and data provenance audit of the ten most critical AI systems in use. For each, record where the compute sits, where the model was trained, who holds the weights, and which legal jurisdiction can compel disclosure. Assign this to the risk or assurance function, not the procurement team, so the finding is independent of the purchase decision.
Second, rewrite the exit clause in every major AI contract. Specify what capacity, data and intellectual property remain under the buyer’s control if the vendor changes terms, raises price or is compelled to withdraw. A contract without a testable exit is not a contract. It is an arrangement at the supplier’s discretion.
Third, build independent evaluation capacity in-house. At least one named team, with budget, should be able to test the models the institution depends on and reproduce its findings without vendor assistance. This is the intellectual-sovereignty score made real. If the institution cannot reproduce a vendor’s safety or performance claim, it cannot verify it.
Fourth, adopt the proposal’s own discipline for any compute benefit offered by a supplier. Free tokens, expiring credits and vendor-specific training do not count as capability. Require reserved capacity, portable data and retained intellectual property instead. This single rule separates a real sovereignty offer from a discount dressed as a partnership.
Fifth, join a pooled procurement structure rather than negotiating alone. The coordination vehicle works because it aggregates demand and allocates by merit, which is how small buyers gain leverage against large suppliers. Institutions that negotiate singly pay the dependency premium that pooled buyers avoid.
Sixth, set a decision threshold and publish it. State that any critical system scoring three or below in three or more domains of the sovereignty matrix, or failing the compute and provenance audit, will not be renewed without a remediation plan. A threshold that is stated and enforced is what turns a policy into a posture. Where the score cannot be improved by the institution alone, the finding is the case for joining a national or sector pool.
The Question Revisited
The question was what a country actually owns when it hosts the machines but not the models, the chips, the data pipelines or the skill. The ACTU proposal gives a practical answer: ownership is what survives a change of supplier. A levy, a reserved share of compute and a retention rule for intellectual property are not declarations of independence. They are mechanisms that convert a host’s position into a holder’s position, layer by layer, and make that control testable.
Through the TEE Method, the answer becomes three habits rather than one policy. Talent builds the people who can operate, evaluate and audit the stack, so judgement stays in-house. Enterprise writes control into procurement, so exit is real and data mobility is a term, not a favour. Ecosystem builds a standing mechanism that pools demand and keeps public-funded capability in circulation, so bargaining power compounds instead of fragmenting. Sovereignty, on this reading, is not a wall. It is the retained ability to choose, verify and leave.
That matters because the alternative is not disaster but drift. A country, or an institution, can host every machine, sign every deal and still lose the ability to decide, because the decisions live in the licence, the weights and the audit trail. The sovereignty test is whether control survives the moment the supplier’s terms change. Leaders who can answer that question for their ten most critical systems this quarter will be governing. Those who cannot will be hosting, and paying rent on their own future. Where the capability gap is genuinely beyond a single institution, the same method points to pooling, which is why the proposal’s coordination vehicle is worth serious study rather than easy dismissal.
Editorial note: This is a news-led Sovereignty Briefing analysis produced with Hermes-assisted research and source checks.
Sources
Australian Council of Trade Unions, UNSW AI Institute, Australian Institute for Machine Learning and Kingston AI Group, “Joint Statement on Sovereign AI Capability in Australia“, 1 October 2026.
ABC News, “Unions want big tech to pay levy to build Australian-owned AI“, 1 October 2026. Featured image: Sam Altman / OpenAI, credit Kim Kyung-Hoon, Reuters, via ABC News.
CyberShack, Scott Hunt, “A national AI compute pool could help Australia’s tech sovereignty problem“, 1 October 2026.
Australian Academy of Science, “Submission on the Inquiry into artificial intelligence and data centres“, September 2026.
This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? For the complete framework, see tonishatagoe.com.
Sign in to save