The AI Race: Who Owns the Chips, the Energy and the Rules?

A report prepared by the US Semiconductor Industry Association with Boston Consulting Group in 2021 estimated that 92% of the world's capacity for the most advanced chips, those below 10 nanometres, lies in Taiwan, and 8% in South Korea. The report described this concentration as "single points of failure" that natural disasters or conflicts could disrupt. Every advanced model you read about today passes through those chips before it reaches you.

The chip: one narrow link

On 7 October 2022 the US Department of Commerce announced rules restricting the export of advanced computing chips and semiconductor manufacturing equipment to China, and required licences for exporting manufacturing equipment to chip fabrication plants there, with a presumption of denial for facilities owned by Chinese entities. The chip has thus become an instrument of policy and not merely a commodity. The figure above dates from 2021, and the picture may change as factories and restrictions change; what matters is the pattern of concentration, not today's exact percentage.

Energy: a physical constraint

The International Energy Agency estimates that data centres consumed about 415 terawatt-hours in 2024, roughly 1.5% of global electricity consumption, and that they could reach about 945 terawatt-hours in 2030 in its base case, around 3%. The agency notes that the United States and China together account for nearly 80% of the growth in this demand. This is a scenario estimate, not a certainty, but it warns that AI competes for electricity just as it competes for chips. For more, see the article "Artificial Intelligence on Physical Ground: Electricity, Water and the Cloud".

Models: a count, not quality

Stanford's AI Index (2025) counted 40 notable models produced by the United States in 2024, 15 by China, and 3 by Europe combined. This is a count of models, not a measure of their quality; the same index notes that DeepSeek-V3 achieved high performance using fewer computational resources than many leading models, and that leading American models have generally been more compute-intensive.

Europe: a bet on rules

The European Union chose to begin with legislation. The AI Act entered into force on 1 August 2024; its prohibitions began to apply on 2 February 2025, and the obligations for general-purpose models on 2 August 2025. The deadlines for high-risk systems were then postponed by an amendment that entered into force on 27 July 2026, so that they now run from 2 December 2027 for sensitive uses and from 2 August 2028 for regulated products. The European example shows that rules themselves need review and adjustment.

The Gulf: from announcement to capacity

In May 2025 HUMAIN, owned by Saudi Arabia's Public Investment Fund, announced a deal with Nvidia to supply 18,000 GB300 chips for a single supercomputer, with an intention to supply "several hundred thousand" over five years and to build data centres with a capacity of up to 500 megawatts. These are announcements, and turning them into operating capacity requires skills, data, electricity and regulation. On the human and linguistic condition, see the articles "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models" and "Localising AI: Cultural Context, Not Interface Translation".

What remains?

The race is not decided by one link, and we do not yet know who will advance. The sounder question is: which links do we need to master, in which do we take part, and how do we protect ourselves from the failure of any one of them? Then we take the means and rely on God.