When the Australian government announced its AI strategy update late last year, sovereign AI compute featured prominently. The argument is straightforward: if AI becomes critical national infrastructure, Australia needs domestic capacity to run AI workloads rather than depending entirely on hyperscaler data centres controlled by foreign companies.

It’s a sound argument. The question is whether Australia can realistically build the compute infrastructure needed, and whether the current approach is addressing the right gaps.

The Current State of Play

Australia’s AI compute capacity is primarily provided by the three major hyperscalers: Amazon Web Services, Microsoft Azure, and Google Cloud Platform. All three operate Australian data centre regions (Sydney, Melbourne, and expanding). They provide access to GPU-accelerated instances suitable for AI training and inference.

For most enterprise AI workloads, this infrastructure is adequate. Australian organisations can run inference, fine-tune models, and deploy AI applications on domestic hyperscaler infrastructure with acceptable latency and data sovereignty for most use cases.

What Australia lacks is large-scale training capacity. Training frontier AI models requires thousands of high-end GPUs running continuously for weeks or months. This scale of compute is concentrated in the United States, with emerging capacity in the UK, Japan, and the Middle East. Australia has no comparable training-scale facility.

The CSIRO operates some of Australia’s most significant public sector compute resources, but these are general-purpose high-performance computing systems not optimised for the specific workloads involved in training large AI models.

What Sovereign Compute Actually Means

The term “sovereign AI compute” gets used loosely. It’s worth distinguishing between several different capabilities.

Data Sovereignty

The most straightforward requirement: ensuring that Australian data processed by AI systems stays within Australian borders and is subject to Australian law. The major hyperscalers already provide this through their Australian regions, though questions remain about metadata, operational data, and access by foreign intelligence services under their home country laws.

For sensitive government and defence workloads, the Department of Home Affairs has established the Certified Cloud Services List, which assesses cloud providers against Australian security requirements. Several Australian-owned cloud providers offer PROTECTED-classified services that meet stringent data sovereignty requirements.

Inference Sovereignty

The ability to run AI models on Australian infrastructure. This is largely achievable today for commercially available models. Organisations can deploy open-source models on Australian cloud infrastructure and run them without external dependencies.

The gap is in specialised hardware. NVIDIA’s latest GPU architectures are allocated preferentially to large customers and specific geographies. Australian cloud providers sometimes face longer wait times for new hardware generations, which can mean running inference on older, less efficient hardware than what’s available offshore.

Training Sovereignty

This is the significant gap. Training large models from scratch requires infrastructure at a scale that doesn’t exist in Australia. Estimates suggest that training a frontier model comparable to GPT-4 or Claude requires 10,000-30,000 high-end GPUs running for several months, representing a capital investment of $500 million to $1 billion in hardware alone, plus substantial power and cooling infrastructure.

No Australian entity, public or private, currently has this capability. The National AI Centre’s strategic plan acknowledges this gap and outlines a multi-year investment pathway, but the timeline and funding remain uncertain.

The Power Problem

Australia’s sovereign compute ambitions run into a physical constraint: electricity.

A large-scale AI training facility consumes 50-100 MW of continuous power. For context, that’s equivalent to powering 40,000-80,000 homes. The data centre industry in Australia is already competing with residential and industrial demand for grid capacity, particularly in Sydney and Melbourne where data centres are concentrated.

The situation is compounded by Australia’s renewable energy transition. AI training facilities need reliable baseload power. While renewable energy is abundant in Australia, its intermittent nature creates challenges for workloads that need to run continuously. Battery storage can bridge short gaps but isn’t yet economical for the multi-day outages that extreme weather events can cause.

Some proposals advocate locating AI training facilities in regions with abundant renewable energy, such as North Queensland with its solar resources. This trades grid proximity for cheap clean energy, but introduces latency, connectivity, and workforce challenges.

What Other Countries Are Doing

Comparing Australia’s approach with peers provides useful context.

United Kingdom: The UK has invested in a national AI compute facility (the Isambard-AI supercomputer) with significant GPU capacity dedicated to research and public sector use. Announced in 2023 and progressively coming online, it represents a direct government investment in training-scale compute.

Japan: Substantial government subsidies have attracted NVIDIA, Microsoft, and domestic providers to build large-scale AI compute facilities. Japan’s approach combines public investment with incentives for private sector infrastructure development.

Saudi Arabia and UAE: Massive sovereign wealth fund investments in AI compute, motivated by economic diversification ambitions. These programs are building infrastructure at scales that dwarf what most Western nations are contemplating.

Canada: The Canadian Sovereign Compute Strategy focuses on research compute and has funded several university-based AI compute centres. The approach is more distributed than the UK’s centralised model.

Australia’s current investment sits well below all of these comparators relative to GDP.

What Australia Should Prioritise

Given finite resources, where should Australia focus its sovereign compute investment?

Inference capacity is the immediate priority. This is where most enterprise and government AI value is delivered. Ensuring reliable, secure, high-performance inference infrastructure within Australia addresses the near-term need. Working with business AI specialists who understand deployment requirements can help organisations make the right infrastructure choices.

Research compute enables local capability development. Australian universities and research institutions need access to GPU clusters large enough to train meaningful models. This doesn’t need to be frontier-model scale, but it needs to be sufficient for Australian researchers to contribute to the field rather than just consuming models trained elsewhere.

Training sovereignty can wait, partially. Building frontier-model training capacity in Australia is expensive and may not be the best use of limited resources. A more pragmatic approach would be securing guaranteed access to training capacity in allied nations through bilateral agreements, while building domestic capacity for fine-tuning and smaller-scale training.

Workforce development matters more than hardware. Australia’s AI workforce gap is more constraining than its infrastructure gap. The talent needed to design, train, and deploy sophisticated AI systems is scarce globally and particularly so in Australia. Investing in workforce development produces compounding returns that hardware investments alone cannot.

The Realistic Assessment

Australia will not train a frontier AI model domestically in the near term. The capital investment, power infrastructure, and workforce requirements are beyond current plans and budgets.

What Australia can and should do is build sufficient sovereign infrastructure to run AI workloads securely within Australian borders, develop research compute that keeps local institutions competitive, and create policy frameworks that ensure Australian organisations aren’t critically dependent on any single foreign provider.

The sovereign compute conversation needs to move from aspirational announcements to practical investment in the infrastructure and workforce that will deliver tangible capability within achievable timelines.