The AI infrastructure gap in Australia
Australia is one of the most advanced digital economies in the Asia-Pacific region. But when it comes to AI infrastructure, there is a structural dependency problem.
The majority of Australian businesses running AI workloads today are running them on infrastructure controlled by three US hyperscalers: AWS, Microsoft Azure, and Google Cloud. These providers have data centre regions in Australia, which creates a common misconception: that using an Australian region means the workload is sovereign.
It does not.
A hyperscaler region in Sydney is a physical facility, but it is not an Australian legal entity. The control plane, the software layer that manages compute allocation, data routing, access control, and billing, is operated by the US parent entity. The terms of service are governed by US law. The US CLOUD Act applies to data held by US companies regardless of where that data physically sits.
Australia has no equivalent domestic AI infrastructure at hyperscaler scale. That gap exists for reasons of capital, market size, and timing. But it has strategic consequences that are now becoming visible as AI moves from experimental to operational across regulated industries. The question is not whether local sovereign AI infrastructure should exist. It is who will build it and when Australian organisations will use it.
Risks of running AI on offshore infrastructure
The risks follow directly from the legal and technical reality of foreign-controlled infrastructure. They are not theoretical.
Sovereignty exposure. The US CLOUD Act, enacted in 2018, allows US authorities to compel disclosure of data held by US companies anywhere in the world. For AI workloads, this extends to training data, fine-tuning datasets, inference inputs, model outputs, and audit logs. An Australian business running AI on a US hyperscaler operates under US legal reach, regardless of which region the compute is allocated to. Selecting the ap-syd-1 region does not change which legal entity owns the control plane.
Compliance risk. The Privacy Act 1988 regulates cross-border disclosure of personal information. The Security of Critical Infrastructure Act 2018 places obligations on entities in sixteen asset classes, including health, financial services, energy, transport, and communications. APRA’s CPS 230 prudential standard creates third-party risk management requirements for APRA-regulated entities that procure technology services. None of these obligations are suspended because the workload is labelled AI. AI workloads amplify the risk, because they ingest and act on sensitive data at a scale and speed that manual processes do not.
Latency. AI inference latency is not only a performance issue. For real-time applications, it is a functional one. Decision-making AI systems used in fraud detection, patient triage, logistics optimisation, or financial trading require sub-100ms response times. Offshore routing cannot reliably provide this. Local AI infrastructure is a technical prerequisite for these use cases, not just a compliance preference.
Support and accountability. When something goes wrong with an AI system, the support escalation path matters. Offshore support structures are slower, less accountable to Australian regulators, and less responsive to compliance requirements that arise under Australian law. An infrastructure provider with Australian-based support staff, subject to Australian employment and privacy law, is a materially different accountability arrangement.
What sovereign AI infrastructure looks like
Sovereign AI infrastructure is not generic cloud with an Australian data centre address. It is purpose-built for AI workloads, locally governed, and structured to satisfy Australian regulatory requirements.
GPU-dense compute. AI workloads, particularly model training and large-scale inference, are not CPU-bound. They require purpose-built GPU hardware. Modern AI-grade infrastructure, including NVIDIA H100 and A100 class hardware, draws between 5kW and 10kW per rack equivalent, compared with 3-5kW for standard compute. The facility’s power supply and cooling systems must be designed to sustain these loads continuously.
Local data backbone. AI workloads move large volumes of data between storage and compute. A sovereign AI infrastructure stack requires high-bandwidth, low-latency interconnects between storage and GPU clusters, all within Australian jurisdiction. This shapes the design of the facility and the location choices for compute and storage tiers. It is not a generic networking requirement.
Power and cooling at scale. AI compute is thermally dense. Facilities running AI-grade GPU infrastructure require advanced cooling, including direct liquid cooling or high-density air cooling, and reliable power infrastructure engineered for continuous high-draw workloads. Tier 3+ certified facilities at established Australian colocation providers are designed to meet these requirements. Standard enterprise data centres are not.
Locally governed operations. The facility and its operators must be subject to Australian law. Support staff must be located in Australia and accountable under Australian employment and privacy frameworks. AUD-denominated contracts with an Australian legal entity clarify jurisdiction and establish a basis for enforcement under Australian law. These are governance requirements, not commercial preferences.
Government and industry momentum
Australian government policy is shifting, and infrastructure decisions are being re-evaluated against sovereignty criteria at both Commonwealth and state levels.
The National AI Centre, operating under CSIRO, is the federal government’s primary mechanism for building Australian AI capability. Its mandate includes supporting local AI research, publishing governance frameworks, and working with industry on sovereign AI deployment. Its existence reflects a policy recognition that offshore AI dependency is a strategic risk that requires active intervention.
Government AI procurement is also changing. AI systems procured by Commonwealth and state agencies are increasingly subject to data sovereignty assessments as part of the evaluation process. For government-facing vendors, and for businesses supplying regulated industries, the ability to demonstrate sovereign AI infrastructure is becoming a procurement differentiator. It is no longer sufficient to assert that data remains in Australia. Agencies are asking which legal entity controls the infrastructure and whether CLOUD Act exposure exists.
Industry investment is responding. Local infrastructure providers are deploying GPU-dense AI-ready compute at tier-certified Australian data centre sites across Sydney and Melbourne. Open-weight AI models, including Meta Llama, Mistral, DeepSeek, and Qwen, can be run on this local infrastructure without routing inference through offshore API endpoints. That removes one of the most significant sovereignty gaps in typical AI deployments: the inference call that leaves Australian jurisdiction even when underlying data storage is locally compliant.
What businesses should ask their infrastructure provider
Not all claims of sovereign AI infrastructure are equal. These questions distinguish genuine sovereign infrastructure from marketing language.
Where is the legal entity? The data centre address is not the relevant question. Ask which legal entity owns and operates the service, and which jurisdiction’s law governs the contract. A US-headquartered company with a local facility is still a US-headquartered company.
Is the control plane local? Who manages compute allocation, access control, and data routing? Is that management layer operated by an Australian entity, or by an offshore parent? The control plane jurisdiction determines CLOUD Act exposure.
What are the support arrangements? Is support provided by Australian-based staff? Can the provider engage directly with Australian regulators if a compliance question arises? Offshore tier-1 support with Australian escalation is not the same as Australian-based operations.
What certifications apply? Can the provider demonstrate data centre certifications relevant to regulated workloads, such as ISO 27001 or tier ratings that establish physical security and operational standards? Certifications provide an independently verified baseline; verbal assurances do not.
Is the contract AUD-denominated? AUD billing signals that the contracting entity is Australian and that the commercial relationship is structured under Australian law. It is not definitive, but it is a meaningful indicator.
What are the data residency guarantees? Does the provider contractually commit that data will remain within Australian jurisdiction, including for backups, disaster recovery, and support access? Is that commitment enforceable under Australian law, with clear contractual remedies?
A provider that answers all of these questions clearly and specifically has built sovereign infrastructure with intention. Vague or incomplete answers are meaningful data points in their own right.
Australia’s AI capability will be built on its AI infrastructure. Organisations that get the infrastructure right, locally governed, purpose-built, and compliant by design, are the ones that will be able to operate AI systems without constraint when it matters most.
Related reading: What is sovereign AI? A guide for Australian businesses and Choosing an AI infrastructure partner: what to look for.
Frequently asked questions
What does “sovereign AI infrastructure” actually mean? It means AI compute, storage, and control systems owned and operated by an Australian legal entity, under Australian law, with no foreign parent able to compel access to the data. A local address alone does not qualify.
Is an AI workload sovereign if it runs in an Australian cloud region? Not necessarily. A hyperscaler’s Sydney region is a physical facility, not an Australian legal entity. The control plane, billing, and terms of service are typically still governed by the US parent company, which keeps CLOUD Act exposure in place regardless of region.
What is the US CLOUD Act, and why does it matter for AI workloads? It is US legislation that lets US authorities compel US-incorporated companies to produce data they hold anywhere in the world, including training data, prompts, and model outputs stored on Australian servers. It applies based on the operator’s legal jurisdiction, not the server’s physical location.
Why is government AI procurement increasingly requiring sovereign infrastructure? Commonwealth and state agencies are adding data sovereignty assessments to AI procurement, driven by policy bodies like the National AI Centre. Vendors are now expected to show which legal entity controls their infrastructure, not just assert that data stays in Australia.
