There is a persistent assumption embedded in how most people talk about artificial intelligence: that it exists outside the normal constraints of the physical world. It scales through code. It improves through iteration. Its limits are technical, not material. That framing was understandable, because for a long time it was close enough to true. Digital growth appeared largely decoupled from physical friction. The infrastructure grew quietly in the background, and the constraints that mattered were financial or computational, not material.
That separation is closing. And the timing is not incidental.
Behind every model is a physical system: data centres, power supply, cooling infrastructure, land, specialised hardware, and network capacity. The more capable the models become, the more those requirements expand. This is not metaphor. It is a direct, compounding relationship. Training large models and running them at scale requires continuous, high-density power draw over sustained periods. Unlike traditional computing loads, which absorb bursts of activity, AI workloads operate at high utilisation for extended durations. Grids were not designed for that profile.
Then there is cooling. AI systems generate heat that has to be managed, and that management requires additional energy and, in many cases, significant volumes of water. In regions already navigating higher temperatures or water stress, this is not a peripheral concern. It is a site selection factor, a regulatory question, and in some cases a community one. The more advanced the system, the more pronounced the requirement.
Location follows the same logic. Data centres need physical space, connectivity, and proximity to reliable power. They are increasingly positioned not simply where demand exists, but where infrastructure can support them. This is already reshaping investment patterns in ways that are not yet fully visible. Regions that can offer energy stability and policy clarity are becoming more attractive. Those that cannot risk being bypassed regardless of other advantages. The competition between jurisdictions for AI infrastructure is, at its core, a competition for grid capacity.
The materials picture sharpens this further. AI depends on specialised hardware concentrated in a small number of global supply chains. Any disruption, whether geopolitical or environmental, has a direct impact on capacity. This is not theoretical. It is visible in how governments are treating semiconductor access as a strategic question distinct from other industrial inputs, in how long-term power agreements are being secured ahead of scarcity rather than in response to it, and in how data centre projects are increasingly tied to energy availability rather than market demand alone.
Taken together, these factors reframe AI from a software story into a resource story. And that reframing matters, because AI is being positioned simultaneously as a driver of productivity and efficiency while increasing demand on systems that are themselves becoming more constrained: energy, water, materials, infrastructure. These two forces are moving in opposite directions at the same time.
That does not mean AI slows. It means it becomes selective. Where it is deployed, how it scales, and who can access it will increasingly be determined by physical conditions, not just technical capability.
The future of AI will not be determined solely by breakthroughs in models. It will be shaped by energy grids, supply chains, climate conditions, and infrastructure decisions that most people tracking AI are not watching. The story is not that AI is accelerating. It is that it is accelerating inside a world that has its own limits, and those limits are not waiting for the technology to catch up.
