The Compute Cartel: Who Controls AI's Critical Infrastructure

Sometime in the past three years, without a formal treaty or a cartel declaration, the world's most consequential industrial resource quietly changed hands. It is not crude oil, rare earth oxides, or enriched uranium. It is measured in petaflops, priced in the hundreds of thousands of dollars per unit, and fabricated almost exclusively on a single island off the coast of mainland China. The concentration of advanced AI compute — the raw capacity to train and run frontier models — inside a handful of corporations, one dominant chipmaker, and one irreplaceable contract manufacturer has created a chokepoint that no strategic planner in Washington, Brussels, or Riyadh fully anticipated, and that none has yet devised a coherent plan to dissolve.

The architecture of this dependency has been assembling itself in plain sight. NVIDIA's H100 and successor B200 accelerators, which together underpin virtually every serious frontier-model training run, are fabricated on TSMC's 3-nanometre process node in Taiwan — a supply chain so geographically concentrated that a single adverse weather event or military contingency could halt global AI development for the better part of a year. Washington moved to weaponise that bottleneck with its October 2022 export-control package, tightened it materially in October 2023, and extended the perimeter again in December 2024 to close loopholes that had allowed restricted chips to reach Chinese data centres via third-country intermediaries. Meanwhile, the hyperscaler oligopoly — Microsoft and OpenAI, Google DeepMind, and Amazon with Anthropic — has absorbed the available H100 and B200 allocation at a pace that leaves sovereign governments, universities, and smaller commercial operators rationed or priced out entirely, accelerating a rush of state-backed compute initiatives from the UAE's G42 and Saudi Arabia's newly launched HUMAIN programme to the EU's network of AI gigafactories and the UK's sovereign AI infrastructure commitment.

The full article maps the emerging geopolitics of compute through the lens of a simple but under-examined thesis: that access to advanced AI accelerators now functions less like a market in electronic components and more like access to a strategic energy reserve, with all the statecraft, supply anxiety, and great-power rivalry that analogy implies. It will argue that electricity — the unglamorous second-order constraint of data-centre power draw, already pushing grid operators across Virginia, Ireland, and Singapore toward hard capacity limits — is fast becoming as binding a chokepoint as the silicon itself, and that the nations which move earliest to secure both the chips and the gigawatts will hold a structural advantage in the next decade's economic competition that will prove extraordinarily difficult for latecomers to close.