Executive Summary
Australia's National AI Plan (2 Dec 2025) frames a whole-of-government approach across three goals: capturing opportunities, spreading benefits, and keeping Australians safe. The Plan consolidates multiple existing and committed initiatives and is best read as an acceleration-of-diffusion and safety program rather than a sovereign attempt to underwrite frontier-model training at hyperscale.
Over 2026–2035, the binding constraint on domestic AI capacity is the scale, efficiency, and utilisation of Australia's data-centre fleet, because this determines the ceiling on deployable compute.
To support practical planning, this whitepaper provides a transparent conversion from electricity (TWh/year) to AI service capacity. Under a reference scenario:
For Australia, the highest-leverage national strategy over 2026–2035 is rapid diffusion of productivity gains, domain adaptation on Australian data where appropriate, development of assurance/evaluation capability, and resilient infrastructure to host and operate AI systems domestically when sovereignty and security require it.
Context and Policy Landscape (2025–2026)
AI is now a general-purpose capability that reduces the marginal cost of knowledge work, software production, content generation, and many forms of decision support. The practical question for Australia is not whether frontier models will exist, but:
- How quickly institutions can adopt them safely
- How much domestic infrastructure exists to host, adapt, and operate models where sovereignty is required
- Whether energy, network, and data-centre planning can keep pace with demand
The National AI Plan consolidates government actions intended to lift research translation, workforce capability, ecosystem development, and adoption, while framing responsible practice and safety. In parallel, the Digital Transformation Agency's AI Plan for the Australian Public Service 2025 focuses on practical government adoption, capability uplift, and safe deployment within the APS.
1.1 What is funded and what it practically buys
The Plan's funding and program lines are material for adoption capacity (skills, pilots, shared tools, guidance) but are small relative to sovereign hyperscale data-centre and accelerator capex. As a consequence, the Plan is most credibly interpreted as a diffusion accelerator: it reduces adoption friction, provides guardrails, helps crowd in private investment, and builds trust frameworks that allow scaling.
- Capture opportunities — Enable AI adoption across sectors
- Spread benefits — Ensure equitable access and workforce transition
- Keep Australians safe — Establish responsible AI practices and governance
Infrastructure Reality: Data Centres, Electricity, and Efficiency
Domestic AI capability is constrained by the same infrastructure as any large-scale digital service: approvals, fibre, water and cooling, grid capacity, and the ability to finance and operate facilities at high utilisation. Electricity is the most binding shared input because it caps both capacity and cost.
2.1 Why efficiency and utilisation dominate cost
Efficiency gains compound: a 10% improvement in PUE, a 10% improvement in utilisation, and a 10% improvement in model efficiency together deliver roughly 27% more useful output per TWh. At national scale, this makes operational discipline as important as capital investment.
Facilities in hot climates, sites with constrained water access, and operators without sophisticated workload scheduling will systematically underperform. Over a decade, early attention to efficiency will meaningfully change how much AI capacity Australia can field per dollar of electricity and per dollar of capex.
Quantifying AI Service Capacity: From TWh to Tasks
3.1 Reference scenario and assumptions
| Parameter | Value | Notes |
|---|---|---|
| PUE | 1.56 | Industry average (Uptime 2024) |
| GPU TDP | 700 W | H100/H200 class |
| Host/network overhead | 20% | Additional per-GPU system power |
| Grid power per GPU-eq | 1.31 kW | 700W × 1.20 × 1.56 |
| Inference throughput | 20 tok/s | Large model, mixed batch |
| Utilisation | 80% | Well-managed inference cluster |
3.2 Conversion equations
GPU-equivalents per TWh:
GPU-eq = 10⁹ kWh / (P_grid × 8760 h)
Token capacity:
Tokens/year = GPU-eq × tok/s × seconds/year × utilisation
3.3 Tier table (planning envelope)
| Tier | TWh/yr | Tokens/yr | Workplace Tasks/yr | Deep-work Tasks/yr |
|---|---|---|---|---|
| Small | 0.1 | ~3.1T | ~1.2B | ~0.5B |
| Medium | 1.0 | ~31T | ~12.3B | ~5.1B |
| Large | 5.0 | ~155T | ~62B | ~26B |
| National | 21.4 | ~663T | ~265B | ~110B |
What This Enables by Sector (2026–2035)
4.1 Where value concentrates and what constrains scaling
Sector benefits depend primarily on workflow integration, evaluation, and governance. High-volume, document-heavy domains can realise measurable gains quickly when AI is connected to systems-of-record under strong control planes.
National Security, Resilience, and Economic Competitiveness
5.1 Security-relevant capability categories
Security implications span: intelligence analysis and threat detection; secure communications and cryptography; critical infrastructure monitoring; and sovereign data processing requirements where foreign hosting creates unacceptable risk.
5.2 Energy, infrastructure, and resilience
AEMO's 2025 ESOO highlights the need for substantial investment in transmission and generation capacity to maintain reliability under changing demand. The CEFC report similarly highlights the scale and coordination challenges created by rapid data-centre growth. Stakeholders should treat power and cooling as strategic inputs, with resilience planning integrated into AI operating models.
5.3 Macroeconomic implications
AI diffusion can raise productivity by lowering the marginal cost of many clerical and analytical tasks. Gains are not automatic: the binding constraints are workflow redesign, workforce enablement, and measurable quality control. The upside is broad diffusion across services-heavy sectors; the downside risk is uneven adoption and concentration of capability.
International Comparison: US, China, and Australia's Positioning
The 2026–2035 decade will be shaped by global scaling of AI infrastructure and supply-chain security.
6.1 United States: capital formation and semiconductor industrial policy
U.S. industrial policy explicitly aims to expand leading-edge manufacturing clusters and advanced packaging capacity. The Department of Commerce announced a CHIPS incentives award to TSMC (Nov 2024), and NIST announced $1.4B in final awards for the National Advanced Packaging Manufacturing Program (Jan 2025).
6.2 China: scale, constraints, and substitution under controls
China is investing heavily in domestic compute supply and substitutes under external constraints. BIS maintains export controls on advanced computing and semiconductor manufacturing items to the PRC. The OECD documents how these constraints interact with market structure and accelerator supply.
6.3 Australia: comparative advantage and realistic ambition
Australia is unlikely to outspend the U.S. or China on frontier-model training infrastructure. The higher-return strategy is to: rapidly diffuse productivity gains; host and operate AI systems domestically where security requires it; build assurance capability as a national asset; and align energy, planning, and skills to support sustained data-centre growth.
In practice, Australia's strategic position is as an excellent deployer and adapter of frontier models, with selective sovereign capability where dependence is unacceptable.
Risks and Stakeholder Actions
The National AI Plan emphasises responsible practices and harm mitigation. At scale, safety and performance are operating properties, not policies: they must be engineered into control planes, evaluation, and change management.
7.1 Risks that dominate at scale
- Reliability: Hallucinations and brittle behaviour
- Security: Prompt injection, tool compromise, data exfiltration
- Privacy & Sovereignty: Sensitive data handling and cross-border processing
- Workforce: Role redesign and accountability gaps
- Infrastructure: Energy and cooling constraints
7.2 Actions stakeholders can take now
Early movers typically do four things:
7.3 What to measure
Calculation Notes (Reference Scenario)
Electricity-to-GPU conversion uses:
GPU-eq per TWh = (10⁹ kWh) / (P_grid (kW) × 8760 h)
where P_grid is an all-in grid power assumption per GPU-equivalent (GPU + host/network overhead, multiplied by PUE).
Token capacity uses:
Tokens/year = GPU-eq × (tok/s) × (s/year) × (utilisation)
All non-hardware parameters (host/network overhead, utilisation, throughput, tokens per task) should be replaced with organisation-specific benchmarks.