There is a great temptation to treat artificial intelligence infrastructure as one enormous investment theme. Data centres, chips, electricity generation, cooling, fibre, cloud platforms and sovereign capability are frequently discussed as though they form a single, inevitable growth story.
They do not. Each sits at a different point in the value chain. Each has different capital intensity, competitive dynamics, customer concentration and exposure to technological change.
The question for Australian investors, infrastructure owners and policymakers is not simply whether demand for AI will increase. It is where the economic value of that demand will ultimately accumulate — and whether the capital invested to meet it can earn an attractive return.
The stack matters
Start with the physical constraints. AI workloads require electricity, land, construction, cooling, specialised equipment and connectivity. Some of those inputs have long development lead times. Others are subject to sudden changes in technology, pricing or supplier power.
That creates very different investment profiles. A location with secure access to power may possess a relatively enduring advantage. A data centre built around a particular assumption about chip density, cooling or customer demand might face more uncertainty. Connectivity can be critical but commercially commoditised unless it is part of a broader differentiated proposition.
It is possible to be on the right side of an enormous technology trend and still make a poor investment.
Three uncomfortable questions
- What exactly is scarce? Is the bottleneck land, grid connection, firm power, equipment, customer commitments, technical capability or something else? Scarcity can change during the investment life.
- Who takes the utilisation risk? An attractive contracted return looks very different from a speculative build dependent on future demand or a concentrated group of counterparties.
- Where does pricing power sit? Demand growth is only half of the equation. If supply catches up, efficiency improves or buyers consolidate, some parts of the infrastructure stack may face margin pressure.
Sovereign capability requires an economic model
The arguments for Australian AI capability extend well beyond financial return: security, resilience, regulatory control and access to sensitive workloads all matter. But these benefits should be explicit, rather than used to obscure the economics. A sovereign strategy needs clarity about the workloads that require local capability, the customers who will use it, the price premium they will accept and the governance needed to protect it.
Australia also needs to think about where it participates in the global architecture of AI. Not every workload must run locally; not every local asset has to serve a national customer. The right answer may be a portfolio of domestic, regional and international capabilities.
From an investment story to an advantage
I would encourage boards to organise decisions around a simple set of lenses: the durability of the bottleneck; contractual quality of demand; speed to revenue; exposure to technology change; and the strategic options created by an asset.
The last is especially interesting. A power-rich site with a credible route to connectivity may support more than one type of customer. A network operator with flexible infrastructure may play a role beyond wholesale capacity. An energy developer may have options across several kinds of industrial demand. These real options can matter as much as the first business case.
AI infrastructure could become one of the defining capital-allocation opportunities of this decade. That makes discipline, rather than enthusiasm, the critical capability.
Further reading: Deloitte Australia TMT Predictions 2026 and Deloitte Global TMT Predictions provide useful context on compute demand and the supporting infrastructure.