Sometime around 2022, a class of statistical models called large language models began to perform tasks — write essays, generate images, write code, hold open-ended conversations — that had been considered defining markers of human intelligence for the entire history of computing. The release of ChatGPT that November put a usable version in front of a hundred million people inside two months, the fastest consumer adoption ever recorded. By 2025 the leading systems were passing graduate professional examinations, contributing meaningfully to mathematics research, and being embedded in nearly every white-collar workflow. The pace of capability gain caught essentially everyone, including the researchers building them, off guard.
The technology has moved, almost from the start, into a geopolitical frame. The training of frontier models requires the largest data centres ever built, vast quantities of high-end semiconductors (most designed in the US by Nvidia and fabricated in Taiwan by TSMC), and electricity at the scale of small countries — a supply chain whose chokepoints concentrate in a handful of firms and one geopolitically exposed island. The United States has imposed sweeping export controls — from October 2022 onward, tightened repeatedly — on the most advanced chips and the lithography equipment to make them (most of it from the Dutch firm ASML), to slow Chinese progress; China has responded with a state-directed industrial policy of unprecedented scale to achieve self-sufficiency, while open releases like DeepSeek showed in early 2025 that the gap was narrower than assumed and that algorithmic efficiency could partly route around hardware bans; the European Union has chosen to regulate first with the AI Act and is being lapped by both on frontier capability. The strategic stakes are not subtle: if AI continues to accelerate, the country whose firms lead may compound advantages in productivity, military capability, and scientific research at rates the international system has not previously had to absorb. The downside risks — systemic accidents, deliberate misuse, alignment failures, mass displacement of cognitive labour — are also being actively negotiated, in public and in private, more or less for the first time in real time, by people who concede they do not fully understand what they have built.
AI is plausibly the first general-purpose technology since electricity whose deployment is being shaped by export controls, treaty negotiations, and national-security review before its full economic and social effects have arrived. The compute build-out alone — hundreds of billions in data centres, straining power grids — is reordering capital markets and energy policy. Whether it is governed well, or at all, will be one of the defining stories of the next decade.