Australia does not need to own the AI stack. It needs to own the layers that matter.

A Covate perspective on sovereign AI capability.

The Prime Minister's AI speech this week was a good first step. Australian Standards for AI, the Office of AI in PM&C, data centre energy and water obligations, copyright protections for artists. All necessary. None sufficient.

Because the real failure scenario is not that Australia uses American models. That is probably inevitable and possibly fine.

The real failure scenario is structural. Hyperscalers build data centres here. Australian electricity powers them. Australian organisations feed them demand and knowledge. Domestic startups build on proprietary APIs. Government signs long-term foreign contracts. Successful Australian companies get acquired offshore. Model rents and intellectual property income flow overseas.

Australia retains construction jobs, power consumption and subscription bills.

That is the quarry and server farm of the AI economy. The speech acknowledged the risk. The policy does not yet prevent it.

At Covate, our view is straightforward. Australia cannot own the entire global AI stack. It should not try. But it can own enough of the stack to retain strategic control, protect critical capability and capture substantial economic value. The goal is not complete technological autarky. It is sovereign control over the commercially and strategically important layers.

Here is what that looks like.

What Australia should and should not try to own

Australia is not going to manufacture leading-edge GPUs. It is not going to build every layer of the semiconductor supply chain. It is not going to create a GPT-scale frontier model and compete head-on with the United States and China. Even the federal minister responsible for AI has acknowledged that Australia may not build frontier foundation models because of the hundreds of billions in capital involved.

That is not weakness. That is realism.

But dependence on components does not mean surrendering control of the finished capability. Australia does not manufacture every component in a submarine either. Sovereignty comes from controlling operation, maintenance, integration, doctrine, access and supply alternatives.

The realistic Australian position is clear once you stop treating "sovereignty" as a binary.

Australia should own sovereign compute capacity. It should own Australian-controlled model hosting. It should operate open-weight base models locally. It should build Australian domain models. It should own data platforms and knowledge layers. It should own AI applications and agent systems. It should own identity, security, governance and auditing. And it should absolutely own government and industry procurement as a strategic lever.

It should not try to own semiconductor fabrication equipment, leading-edge GPU manufacturing, or a general-purpose frontier model competing with GPT and Gemini. Those are the wrong battleground.

The realistic sovereign stack

Secure Australian compute

Australia needs a nationally controlled compute reserve — GPU clusters distributed across several Australian facilities. Not every data centre. But enough guaranteed compute to support government services, defence and intelligence, universities, Australian model developers, strategic industries, and emergency continuity if foreign services are restricted.

The ownership structure could combine Commonwealth equity, superannuation capital, universities, energy companies and domestic infrastructure investors. The conditions are straightforward: Australian jurisdiction, Australian operational control, guaranteed domestic capacity, transparent pricing, portability between hardware suppliers, and no exclusive dependency on one hyperscaler.

This is achievable. Australia already attracts data-centre investment. The missing piece is reserving some of that capacity for domestic capability rather than merely hosting foreign workloads.

Build on open-weight models, not from zero

Australia does not need to train a giant general-purpose model from scratch. That would probably be wasteful. It could take strong open-weight models and then train or adapt them on Australian English, incorporate Australian law and regulation, improve knowledge of Australian institutions and geography, tune them for local sectors, host them under Australian control, and maintain the ability to replace the underlying model.

CSIRO has noted that sovereignty does not necessarily require building the entire model in Australia. It can mean having the skills, resources and options to operate or replace offshore technology. That is the right frame.

The key is that the base model must remain interchangeable. Australia should avoid building sovereign systems that are technically Australian but permanently tied to one American API.

Own the Australian knowledge layer

This is probably the most important layer, and it is where Australia has genuine leverage.

The valuable asset is not always the underlying language model. Increasingly it is the combination of proprietary datasets, industry taxonomies, regulatory knowledge, workflow history, evaluation data, human corrections, organisational memory, permissions and identity, and feedback from real-world use.

Australia should establish sector-specific data trusts or data cooperatives. An Australian construction-data trust. A mining and resources data trust. An agricultural intelligence trust. A healthcare AI data trust. A legal and regulatory corpus. An energy-grid intelligence platform.

Businesses would not dump raw information into a central government database. They could contribute controlled, anonymised, licensed or federated access while retaining ownership rights. Model and application builders would pay for access. Contributors would share in the resulting value.

The United States has more compute. It does not automatically possess Australia's mining operational data, geology, agricultural conditions, health system data, infrastructure records, regulatory frameworks, Indigenous knowledge, local government information, or construction and environmental approval history.

That data should not be handed over casually.

Build domain models, not another generic chatbot

Australia's best opportunity is not "Australian ChatGPT." It is the world's best AI for areas where Australia has deep knowledge, strong customers and export credibility.

Mining and mineral exploration. Agriculture and biosecurity. Renewable energy and grid management. Climate adaptation. Medical technology. Construction and infrastructure. Defence and cyber security. Financial compliance. Legal and regulatory intelligence. Water management.

These models could be smaller than frontier models but more accurate and commercially valuable within their domain. CSIRO's analysis concludes that Australia is currently more of a downstream AI user than a product creator, but sees a realistic opportunity to develop targeted foundation models and exportable AI products in niche areas.

That is the right battleground.

Own the orchestration and application layer

A business rarely gets value from a foundation model alone. The value comes from the system around it: retrieval, permissions, workflows, integrations, agent orchestration, domain-specific interfaces, audit trails, safety controls, evaluation, human approval, transaction execution.

An Australian company could use an open American or European base model while owning everything above it. That would still allow Australia to capture product revenue, customer relationships, intellectual property, workflow data, industry expertise, export revenue, employment and tax.

The minister has made a similar argument: competitive advantage increasingly exists in the layers between the base model and the final domain-specific product. That is correct. But government policy needs to turn that observation into procurement, capital and infrastructure.

Use government procurement as the anchor customer

The Commonwealth and states collectively represent one of the largest technology buyers in the country. They should stop treating every AI purchase as an isolated software procurement.

For strategic workloads, tender evaluations should include Australian ownership of IP, local model portability, Australian data control, local hosting options, ability to operate without the original vendor, domestic reinvestment, export potential, and contribution to Australian datasets and model capability.

This should not mean blindly purchasing inferior Australian software. It should mean giving strategic value an explicit weighting, just as defence procurement considers sovereign industrial capability.

Government could also issue multi-year purchasing commitments to Australian AI providers. Guaranteed demand is often more useful than grants because it gives companies recurring revenue and makes them investable.

Create an Australian AI investment vehicle

Australian startups often reach the stage where they require serious growth capital and then relocate to the United States, sell to a foreign company, transfer key IP offshore, or become dependent on foreign venture capital.

Australia needs a patient capital vehicle — not another small grant program. It could be capitalised by the Commonwealth, Future Fund-style capital, superannuation funds, private investors and strategic industry partners. Its mandate should be to retain Australian ownership or control in strategically important AI companies.

It should be commercially managed, not run as a departmental grants committee. A plausible initial scale would be several billion dollars, increasing as co-investment is attracted. That is significant enough to build domestic capability but still far below the cost of trying to manufacture a complete frontier ecosystem from scratch.

Treat interoperability as national infrastructure

Every major Australian AI procurement should require exportable data, model-independent APIs, documented embeddings and retrieval systems, portable prompts and agent definitions, independent evaluation datasets, the ability to switch models, no contractual use of customer data for unrelated training, and clear ownership of fine-tuning and feedback data.

This is essential because today's strongest model may not remain the strongest. Sovereignty comes partly from being able to replace suppliers without rebuilding the entire organisation.

What failure looks like

The failure scenario is not simply that Australia uses American models.

The failure scenario is that hyperscalers build data centres here. Australian electricity powers them. Australian organisations feed them demand and knowledge. Domestic startups build on proprietary APIs. Government signs long-term foreign contracts. Successful Australian companies are acquired offshore. Model rents and intellectual property income flow overseas.

Australia retains construction jobs, power consumption and subscription bills.

That would make Australia the quarry and server farm of the AI economy. The PM's speech acknowledged this risk in words. The policy does not yet prevent it in practice.

What success looks like

A realistic Australian full stack would look like this: imported but diversified hardware, under Australian-controlled compute, running open-weight and licensed foundation models, with Australian-owned adaptations and domain models, built on Australian data trusts and knowledge systems, wrapped in Australian orchestration, security and governance, delivering Australian applications and agents, serving Australian customers and export products with retained IP.

That is not fully independent at every physical layer. But it creates operational sovereignty and economic ownership.

A sensible national target would be: Australia must be able to operate critical AI systems without permission from a foreign frontier-model company, and Australian organisations must be able to convert Australian data and expertise into Australian-owned intellectual property.

That is achievable. But it requires the government to stop measuring success by the number of foreign data centres announced and start measuring Australian AI revenue, Australian-owned IP, domestic compute available to local firms, procurement awarded to Australian providers, strategic datasets under Australian control, exports from Australian AI products, and dependence on individual foreign model suppliers.

The Covate view

The real choice is not between total independence and total dependence. It is between strategic ownership and becoming a permanent renter of foreign intelligence.

Australia has genuine strengths: abundant energy potential, political stability, strong universities, significant pools of superannuation capital, valuable sector-specific datasets, expertise in mining, agriculture, health, energy and defence, close relationships with major technology-producing allies, and more than 1,500 AI-focused companies.

Its weaknesses are also serious: a relatively small domestic market, weak commercialisation of research, limited growth capital, dependence on imported compute, fragmented procurement, talent migration, few companies capable of absorbing billion-dollar investment, and a habitual preference for buying mature foreign products.

The technical challenge is manageable. The institutional challenge is harder. Australia would need to coordinate government procurement, competition policy, energy, infrastructure, research, finance, data rights and industrial policy. Historically, that kind of coordination has not been one of Australia's strengths.

But the window is open. The PM has named the problem. The Office of AI exists. The standards framework is coming. The question now is whether the policy extends beyond regulation and physical infrastructure into commercial strategy — what Australia builds, what it owns, where it invests, and how it keeps the companies that matter.

That is the decision. Everything else is commentary.

Covate is a technology investment and advisory firm working with boards, investors and growth companies on AI, digital transformation and technology strategy.