Every frontier AI system is built from the same four ingredients, and each one has a geography. First, advanced logic chips — today made in commercial volume only on Taiwan. Second, the machines that make those chips — extreme-ultraviolet lithography tools produced by a single Dutch company, ASML, using mirrors from Germany and lasers from the US. Third, energy — enough baseload electricity, cheaply and reliably, to run data centres the size of small cities. Fourth, talent and data — a small global pool of top researchers and a shrinking supply of new, high-quality training data. Whichever country secures all four wins the decade. Whichever loses even one becomes dependent on somebody else.
TSMC, headquartered in Hsinchu, fabricates the overwhelming majority of the world's leading-edge logic chips — the ones inside Nvidia's AI accelerators, Apple's phones and modern US weapons systems. No other fab on earth currently produces 3-nanometre chips at commercial scale. That single fact reshapes everything else: it is why the US Chips and Science Act poured tens of billions into Arizona and Ohio fabs, why Japan subsidised a TSMC plant in Kumamoto, why Germany funded one in Dresden, and why every Taiwan Strait crisis is also a global AI crisis. Analysts sometimes call TSMC's fabs the 'silicon shield': the world's dependence on them is Taiwan's best deterrent.
You cannot build a modern AI chip without an extreme-ultraviolet (EUV) lithography machine, and EUV machines are made by exactly one firm on the planet: ASML, in Veldhoven, near Eindhoven. Each machine costs around 200 million dollars, weighs 180 tonnes and takes 40 freight containers to ship. Under intense US pressure the Dutch government has, since 2023, blocked ASML from selling its most advanced EUV tools to China. That single Dutch export licence has become one of the most consequential foreign-policy decisions in Europe: a small country with a single company sits on a hard chokepoint of the twenty-first century.
The US does not fabricate most of its own chips, but it designs them — Nvidia, AMD, Broadcom, Apple, Google — and it owns most of the frontier model labs: OpenAI, Anthropic, Google DeepMind, Meta AI, xAI. Washington's strategy since 2022 has been to weaponise that position: export controls on Nvidia's top accelerators to China, the Chips Act to reshore fabs, entity lists targeting Huawei and SMIC, and pressure on allies (Netherlands, Japan, South Korea) to align their own export regimes. The bet is that if China cannot buy the chips and cannot buy the machines that make the chips, it will fall a generation behind — long enough for the US to consolidate an unassailable lead.
Beijing's response has been to try to build the entire stack domestically: Huawei's Ascend accelerators, SMIC's 7-nanometre process, indigenous EUV research, its own frontier models (DeepSeek, Qwen, Kimi, Baidu's Ernie), and a state-directed data-centre buildout under the 'East data, West compute' programme that routes AI workloads to the coal- and hydro-rich west. In 2024–25 DeepSeek's V3 and R1 models showed the West that Chinese labs can compete near the frontier at a fraction of the training cost — narrowing the gap even without top-shelf hardware. The Chinese theory of victory is patience: absorb the sanctions, keep training, and wait until the domestic supply chain catches up.
AI accelerators are useless without huge amounts of high-bandwidth memory (HBM), and HBM is dominated by two Korean firms — SK hynix and Samsung — with Micron in the US a distant third. Japan controls a different chokepoint: the photoresists, ultra-pure hydrogen fluoride, silicon wafers and specialty gases without which no fab in the world can operate. Tokyo demonstrated the leverage in 2019 when it briefly restricted exports of three of those chemicals to South Korea. Seoul and Tokyo therefore sit inside every serious AI strategy, whether they want to or not — and Washington's alliance system is quietly the connective tissue that holds the whole supply chain together.
The UAE and Saudi Arabia have realised something most European capitals have not: in an era where the binding constraint on AI is electricity, energy exporters are also potential compute exporters. Abu Dhabi's G42 is building a 5-gigawatt AI campus (Stargate UAE) with OpenAI, Oracle and Nvidia. Saudi Arabia's Humain, launched in 2025, has struck multi-billion-dollar deals for Nvidia and AMD chips. The pitch to Washington is blunt: give us the accelerators, we will give you the megawatts, and we will align with you rather than with Beijing. It is the same trade Gulf states used to make with oil, redrawn for a world where the scarce commodity is training runs.
Europe has world-class talent (Mistral in Paris, DeepMind's roots in London, ASML in the Netherlands, SAP in Germany) but no frontier lab of American or Chinese scale, and it is short on both cheap electricity and risk capital. Its comparative advantage has become regulation: the AI Act (2024), the Digital Services Act, the Data Act. Whether that is a durable strategy or a slow surrender is the live argument. Draghi's 2024 report on European competitiveness put the choice starkly: massively increase investment in compute and energy, or accept becoming a rule-taker in a technology defined elsewhere.
India is trying to position itself as a third pole between Washington and Beijing: the IndiaAI Mission, a national GPU cluster of tens of thousands of Nvidia chips, sovereign foundation models in Indian languages, and a big bet on being the world's back-office for AI applications rather than the frontier itself. It has advantages the West does not — a massive young workforce, cheap engineering talent, English-language fluency — and constraints it cannot easily fix: an unreliable grid, limited semiconductor capacity, and a strategic need to stay non-aligned enough to keep buying Russian oil.
Draw the AI map and it looks nothing like the old economic map. It runs through a fab in Hsinchu, a machine shop in Veldhoven, a memory line in Icheon, a gas plant in Yokkaichi, a data-centre cluster in Abu Dhabi, a hyperscaler campus in Virginia, a hydro dam in Guizhou. The AI race is not one race but four, layered on top of each other — chips, machines, energy, talent — and no single country wins them all. The interesting question of the next decade is not which country will 'have AI'. It is which coalitions will control which layers of the stack, and what they will demand in return for access.