On July 1, Virginia became the first US state to impose a dedicated electricity consumption tax on data centres.
As countries race to establish their leadership in artificial intelligence, there are billions of dollars being invested in AI data centres and agentic AI. While India launched the India AI Mission to build GPU compute capacity, datasets and start-up ecosystems, the United States is pursuing large-scale AI integration through initiatives such as the Genesis Mission.
Governments all over the world are drafting national AI missions, enterprises are fine-tuning foundation models, and start-ups are shipping agentic systems that don't just answer questions, but act — booking, coding, negotiating and deciding. The instinct is understandable: whoever builds the most capable AI first captures the economic upside.
The global AI race is accelerating, and everyone wants a piece of the pie.
However, underneath this sprint, sits a harder question few are addressing — what is this intelligence actually made of, and what happens when the world tries to build enough of it for everyone at once?
The honest answer is that AI is not a purely digital product. It is a physical one, assembled from silicon, copper, rare-earth elements, cobalt, lithium and water. Treating its growth as limitless because it "lives in the cloud" is the mistake sitting at the centre of the moment.
The Mineral Layer Nobody Budgets For
A few years ago, the resource conversation around AI was mostly about a handful of labs training a handful of giant models. That era is over. The current phase is defined by proliferation — thousands of organisations fine-tuning models for narrow use cases and deploying agentic systems that run continuously, rather than answering one prompt and going idle. Across organisations, this creates a permanent new load on the grid.
Public debate about AI's footprint has largely settled on two familiar concerns — electricity and water. Both matter. But a third input gets far less scrutiny, even though it may prove harder to solve: the minerals that make the hardware possible in the first place. Every GPU cluster and cooling system depends on a supply of critical minerals — gallium and germanium for semiconductors, cobalt and nickel for batteries, rare-earth elements for magnets, and copper in quantities large enough that AI infrastructure is now among the fastest-growing sources of new copper demand globally.
Unlike energy, which can, in principle, be decarbonised, these materials are finite and geographically concentrated. A company can offset a data centre's power use, but not a rare-earth mine, especially when hardware is replaced every few years.
Growth At Any Cost Undermines The Growth Itself
There is a temptation to treat sustainability as a constraint to be addressed later and that framing has it backwards. Resource strain does not sit outside the AI growth story; it is increasingly the thing that limits it. A small number of countries dominate refining of the critical minerals AI hardware depends on, so any disruption quickly becomes a global shortage. The industry's habit of chasing marginal efficiency gains through constant hardware upgrades increases mineral throughput rather than easing it. And unlike carbon accounting, which now has recognised frameworks, there is no equivalent standard for tracking mineral use or hardware circularity across AI's lifecycle. Most organisations are scaling AI without tracking one of its largest physical inputs—a gap Virginia's new tax highlights.
What Sustainable Growth Requires
Genuine adoption of a technology depends on a few deliberate shifts: smaller, fine-tuned or distilled models which meet a need rather than a large model for every task; hardware longevity through better utilisation, scheduling and refurbishment, slowing the pace at which fresh extraction becomes necessary; measuring the full lifecycle, not just the power bill, so provenance and end-of-life recycling sit alongside energy and carbon in what gets audited; and shared industry standards built now, before scarcity forces the issue, much as carbon accounting converged on common frameworks after years of fragmented reporting.
Countries and companies competing the hardest for AI leadership are still measuring success in the way earlier industrial races were measured — by speed and scale. But the technology being built this time depends on materials that cannot be manufactured, only extracted, are finite and in unevenly distributed quantities. The organisations still standing a decade from now will likely be the ones that treated resource discipline as a strategy from the outset, not as a cost to work around. Sustainable AI growth is not the cautious alternative to ambition. Judging by Virginia announcements, it may be the only version of this race with a finish line worth reaching.
.png)
