Building Australia's Sovereign AI Future: The Economics of Independence
Last month, I had the privilege of speaking with Karthik Murugan, AWS’s Head of Data and AI for the Public Sector in Asia Pacific and Japan, at the Canberra Symposium in August 2025. Our conversation centered on a critical question facing Australia: how do we build secure and sovereign large language model capabilities that serve our national interests while navigating the complex economics of AI deployment?
The answer, as we explored, goes far beyond just the initial cost of building AI models – it’s about understanding that inference costs are forever, and without sovereign capability, the economic value from AI adoption flows directly to other countries.
The Hidden Economics: Training vs. Inference
When discussing AI costs, most attention focuses on the astronomical training expenses. The latest models are pushing costs to unprecedented levels: Claude 3.5 Sonnet cost between $10-100 million to train, while Llama 3.1 405B required approximately $60-120 million and GPT-4o’s training costs exceeded $100 million. Recent analysis by Epoch AI shows that leading-edge models are now costing hundreds of millions to develop.
But these figures, while substantial, represent only the tip of the economic iceberg.
The reality is starkly different: inference costs – the ongoing expense of actually running AI models – typically account for 80-90% of the total lifetime cost of any AI system. While training is a one-time investment, inference happens continuously, every time someone queries the model, processes data, or generates content.
This disparity is profound. Research shows that for most companies deploying AI, inference overtakes training as the dominant cost driver within 3-6 months of deployment. Studies indicate that more than 80% of computational demand comes from inference tasks, not training.
Consider the scale: while GPT-4o costs around $2.50 per million input tokens and $10 per million output tokens, organizations processing millions of tokens monthly face substantial recurring expenses that dwarf the original training investment.
The Sovereignty Imperative
This economic reality creates a fundamental sovereignty challenge. When Australia relies on foreign AI models, every query, every analysis, and every AI-powered decision sends value overseas. It’s not just about the direct costs – it’s about the cumulative economic impact of dependency.
The Australian Computer Society estimates that building sovereign AI infrastructure would require $2-4 billion in government and private sector co-investment. While this seems substantial, it pales in comparison to the long-term costs of perpetual dependency on foreign systems.
As generative AI solutions scale across government, healthcare, education, and business, the importance of running lean and efficient inference becomes even more critical. This drives a natural evolution: organisations need to reduce the number of tokens going into and out of models, optimise their usage patterns, and ultimately move toward building their own models or at least fine-tuning models to their specific needs.
Analysis shows that companies are spending more than 80% of their total capital raised on compute resources for AI applications, with inference costs rising significantly due to wider business adoption and increasing model complexity.
Cultural and Security Considerations
Beyond economics, there’s a more subtle but equally important issue: models trained overseas are inherently trained on the culture, values, and languages of other countries. For many use cases, this might be acceptable – we’re accustomed to using US English in business contexts, and Western cultural perspectives often align with Australian values.
But as we move into more sensitive applications – particularly in international security, defense, healthcare and governance – this cultural bias becomes problematic.Studies specifically examining AI bias in national security contexts reveal that AI systems can “exacerbate bias in national security applications” and that “biased algorithms deployed in border control settings might erroneously apprehend or expel innocent individuals.” Research on AI biases in critical foreign policy decisions shows that models “exhibit biases in critical foreign policy decisionmaking domains,” with some showing marked preferences toward escalation in crisis scenarios.
Military AI bias research highlights how “the presence of implicit assumptions around gender, ethnicity, ability and other sensitive characteristics in military AI systems can result in misidentification of threats and non-threats, flawed assessments of humanitarian needs, and invasive surveillance and monitoring practices.”
Moreover, the “black box” nature of these systems means we often don’t understand how training data impacts decision-making. The datasets used to train major models are rarely disclosed, making it impossible to assess potential biases or influences that could compromise security-sensitive applications.
Australia’s Strategic Advantages
Australia has unique advantages that position it exceptionally well for sovereign AI development. Political stability, strong rule of law, abundant renewable energy resources, and trusted relationships with key allies create compelling opportunities to become a regional hub for secure, sustainable AI computing infrastructure.
The renewable energy advantage is particularly significant. Australia’s data centers have unique advantages with abundant solar and wind resources, allowing them to reduce or even eliminate their dependence on fossil fuel-based electricity. Projects like Queensland’s “Supernode” - combining data centers with renewable generation and battery storage - demonstrate how Australia can lead in sustainable AI infrastructure. AWS is investing in 3 new solar farms across Victoria and Queensland to support the data center expansion
Geographic location and political stability position Australia well for data centre expansion and to harness opportunities in Southeast Asia. The recent announcements of multi-billion dollar investments by major cloud providers in Australian infrastructure signal confidence in these advantages. Amazon’s recent announcement of a $20 billion commitment to Australian data center infrastructure over the next 5 years demonstrates global confidence in Australia’s position as a strategic AI hub.
The Path Forward
Australia’s sovereign AI strategy must address three key areas:
1. Infrastructure Investment: Building the computational backbone (which includes cloud) necessary to support domestic AI development and deployment. This includes not just GPUs, but the entire digital ecosystem of skills, governance, and research capabilities. Recent government initiatives like the National AI Capability Plan show commitment to this direction.
2. Economic Realism: Understanding that the choice isn’t between expensive domestic capabilities and cheap foreign alternatives. It’s between building domestic capabilities now or paying exponentially more for foreign dependency over time. As inference costs continue to dominate AI economics, the long-term economics strongly favor sovereign capabilities.
3. Cultural and Security Alignment: Ensuring that the AI systems supporting critical national functions reflect Australian values and priorities, rather than the cultural biases embedded in foreign training data. Research consistently shows that cultural alignment becomes increasingly critical as AI systems are deployed in sensitive applications.
The Window of Opportunity
Recent developments make the timing critical. Analysis shows that Australia faces infrastructure inaction creating strategic risk, while industry experts argue that Australia is at an AI crossroads where it must lead, not lag.
The window for building sovereign capabilities is narrowing. Every day of delay makes the eventual transition more expensive and complex. Recent studies by organizations like ASPI highlight that AI governance frameworks are emerging rapidly, and countries that establish sovereign capabilities early will have significant advantages.
As I discussed with Karthik at the Canberra Symposium, the future belongs to those who understand that AI sovereignty isn’t just about technology – it’s about ensuring the next generation inherits the tools to shape their own digital future, rather than being shaped by someone else’s.
The path to AI sovereignty need not be a choice between complete independence and total dependency. Hybrid approaches that leverage trusted cloud infrastructure while maintaining sovereign control over models, data, and decision-making processes offer a pragmatic middle ground. With cloud providers like AWS investing billions in Australian infrastructure, there’s an opportunity to build sovereign capabilities on world-class foundations.
The economics are clear: inference costs dominate AI spending, cultural biases in AI pose real security risks, and Australia has unique advantages for sustainable AI infrastructure. What’s needed now is the national will to act on them.