Chips and Sovereignty: Who Holds the Keys to Computing?

AlexNet, which won the image-classification competition in 2012, was a network of about 60 million parameters, trained on two GTX 580 graphics cards for five to six days (Krizhevsky et al.). The winning entry, an ensemble of seven networks of the same kind, reached a top-five error of 15.3%, against 26.2% for the nearest rival. The story was not a new algorithm alone, but hardware that computes thousands of similar operations at once, which is what neural networks need.

Since that day the chip has been part of the story. Four nodes show why.

Manufacturing: one heavy node

Many design companies own no factories and rely on specialised foundries. In the second quarter of 2025, TrendForce put global foundry revenue at about US$41.7 billion, of which about US$30.2 billion went to TSMC, or 70.2%, while Samsung's share was about 7.3%. The figure should not be read as a share of 'advanced chips' alone; it is a share of all foundry revenue.

The machines: a unique technology

The finest circuits can be etched only with extreme ultraviolet (EUV) lithography machines. The Dutch company ASML says this technology is 'unique to ASML' and that it alone kept working on it until it reached a viable system, after investing more than six billion euros over seventeen years. So the chain is narrow at design, manufacturing and the machine alike, though the firms differ at each link.

When the chip becomes policy

On 7 October 2022 the US Bureau of Industry and Security announced licensing rules for exporting advanced chips and manufacturing equipment to China. The rules of 17 October 2023 widened the controls to cover more types of equipment, a new performance-density threshold, and licence requirements for more than forty additional countries through which chips might be diverted. They also added thirteen entities to the Entity List. These rules change, so anyone who wants the current position should consult the Bureau's page. This article does not assess their economic effect, and that effect is disputed.

Electricity: the other constraint

The International Energy Agency names AI as the most important driver of growth in data-centre electricity use (its figures for 2024 and its 2030 scenario are in the article "The AI Race: Who Owns the Chips, the Energy and the Rules?"). Estimates then widen considerably: between 700 and 1,700 terawatt-hours in 2035. The agency also considers that around 20% of planned data-centre projects could be delayed if grid risks are not addressed. Whoever has the chip without enough electricity cannot run it.

Our place in the chain

HUMAIN was launched in May 2025 under the Public Investment Fund to build the links of the stack, from data centres to models. These are announced plans, and it is too early to judge their results; its aims and partners are set out in the article "AI Diplomacy: Whoever Writes the Rules Helps Shape the Game".

The editorial reading is that making the most advanced chips is not a project that can be bought ready-made, and that neighbouring positions in the chain are more attainable: hosting computing, supplying energy, and building models that understand our language and context. In all of this the wager rests on skills, because a facility is built over years and run by those who understand it, and because buying equipment alone leaves the keys with others. Hardware ages within years; trained minds remain.