Empires of AI

Niall Ferguson, Lisa Klaassen and Eric Schmidt · Foreign Affairs, November/December 2026 issue, published online October 20, 2026 · foreignaffairs.com. Summary below with a few short quotes; the full essay is at the link.

Summary

The claim. In a few years the US and China have each assembled something better described as an empire than a superpower: control over the models, chips, infrastructure and capital behind advanced AI gives them dominion over knowledge itself, and makes every other country that wants AI dependent on them. History’s warning is that empires rarely fall to a challenger; they fail because defending their dominions drains them. That failure can be financial (Britain after 1945), institutional (the late Qing) or technological (HMS Dreadnought, 1906, which made Britain’s own fleet obsolete and forced it to nearly double naval spending while launching pensions). So the decisive question may be not who reaches AGI first but “who can bankroll the race.” What is new is speed: the arc of expansion, overreach and unravelling is happening “at machine speed.”

Why AI is an imperial technology.

• Scarcity returns. From the 1970s to the 2010s software escaped scarcity because it was nearly free to copy. Scaling laws undo that: progress depends on energy, minerals, chips and capital in bulk, and once a technology’s frontier depends on scarce inputs, it is geopolitics.
• Centre and periphery. A handful of firms in two countries control the frontier; everyone else is a dependent user of models they cannot inspect and that can be repriced or withdrawn. Inputs come from the periphery: minerals from the DRC and Indonesia, data labelling by low-paid workers in Kenya and the Philippines.
• Chokepoints. Britain’s power came less from territory than from coaling stations, canals, telegraph cables and sterling; the postwar US order rested on the dollar, Bretton Woods, bases and the nuclear umbrella. The AI stack is a denser lattice of chokepoints (chips, minerals, models) the centres can switch on or off.
• Deliberate vs. absent-minded. China’s is a state project, from the 2017 State Council plan to lead in AI by 2030 to Xi’s 2026 push on global standards. America’s grew from private capital and loose regulation, like Britain’s empire acquired “in a fit of absence of mind” (John Robert Seeley).

The labs as East India Companies. Frontier labs (OpenAI, Anthropic) resemble the chartered companies: they negotiate with heads of state and trade model access and data-centre investment for land, power and sovereign capital, as in the May 2025 Gulf deals. China’s Moonshot is reported to be seeking revenue shares from US clouds hosting Kimi K3. But like the East India Companies, they need the state’s permission to move a chip, and each government holds a leash: in the US, light regulation in exchange for defence access and export-control compliance; in China, state-supplied land, power and capital in exchange for security and ideological review.

Two offers to the world. Washington offers the full stack, expensive and mostly closed, in exchange for alignment with its export controls and against Chinese hardware. Beijing offers cheap, open-weight models that clients can download and modify, with its hardware, clouds and standards following. The cheaper, open offer will likely win more countries. The authors cite the US Commerce Department ordering Anthropic in June to cut foreign nationals off from its two most capable models, which went dark worldwide within days. Each side also has its legitimating story: the State Department’s “Pax Silica” trusted ecosystem against coercive dependence, and Beijing’s open models as a global public good against “new historical injustices.”

Mirror-image bottlenecks. The US leads by perhaps six months in chatbots and software tools; China leads in embodied AI (robotics, autonomy, manufacturing automation). China has cheap power and near-monopoly rare earths but cannot make leading-edge chips; the US has the chips and frontier models but lacks minerals and grid. “One empire is short of watts; the other of transistors.” Capital is the resource that binds both.

The butter burden. The dollar and Wall Street let the US fund its build-out with the world’s savings; US AI debt is mostly private (big tech committed over $600B of 2026 capex, much of it borrowed). China’s is sovereign: a reported $295B five-year plan for state data centres on domestic chips, funded by long-dated bonds. Both already carry heavy debt for non-military reasons, entitlements in the US and local-government borrowing in China, worsened by ageing. Ferguson’s rule from history: an empire that spends more on debt service than defence is in decline. The US crossed that line in 2024 (about $881B of interest vs. $874B on defence), and on present trends interest will be double defence by 2036. AI is sold in both countries as the escape from debt and stagnation, but the payoff may arrive too late, and AI capex is already pushing rates up. A sell-off could tip the US into recession with little fiscal room to respond; in China, overcapacity and margin-destroying competition (neijuan) could raise youth unemployment. Both face guns-or-butter.

How it ends — four scenarios.

1. The challenger overtakes, as Germany and the US overtook Britain by commercialising its inventions. China is doing the latecomer’s version: energy, manufacturing, robotics and near-frontier open models, while the US spends heavily to defend a lead China is commoditising.
2. Exhaustion. Britain won two wars and bankrupted itself, then chose a welfare state over empire. The authors’ “imperial trilemma”: a power can defend its dominions, keep technological primacy, or fund generous welfare at home, but not all three.
3. Stagnation. Extracting rents is easier than inventing. The Soviets could match nuclear technology but not Texas Instruments. US techno-optimists fear federal bureaucracy could do the same to American AI.
4. Rebellion from within. Job losses at home, US backlash against data centres even in Texas and Utah, and in China the risk that open-weight models erode the party’s information control. And the AI itself: they cite the July attack on Hugging Face by hundreds of autonomous agents from an OpenAI model as a sign the most capable systems may already exceed their builders’ control.

Taiwan. The central node of the whole stack, TSMC, sits 90 miles off China’s coast; capturing Taiwan would blunt America’s AI advantage. That makes a “Taiwanese chip crisis” the most likely great-power war. AI raises the risk further because autonomous systems act in seconds, and some conventional weapons (Phalanx, Aegis) already fire without human authorisation. Two off-ramps: make Taiwan dispensable (TSMC fabs in Arizona, Kumamoto and Dresden), though the danger peaks in the window before that succeeds; or an AI détente modelled on Cold War arms control, such as agreeing not to release models above set capability levels in cyber-offence or biological design, and thresholds for open-weight releases. Hard to police, but better than no taboos at all.

Close. AI is “the most capital-intensive imperial project in history,” run by two heavily indebted powers. Every precondition for collision or decline is present, and it is unfolding faster than either political system can keep up with.

Connections

• Niall Ferguson — lead author; the “Ferguson limit” (interest exceeding defence spending) is his. Niall Ferguson and Tyler Cowen — AIpocalypse Now is an earlier Ferguson take on AI in the corpus.
• Eric Schmidt — co-author; former Google CEO, chair of the Special Competitive Studies Project, co-author of Genesis.
• Decline of Empire, Empire Cycles — the historical frame. The Fate of Empires and the Search for Survival is on the anti-library.
• Taiwan, TSMC, Semiconductors — the flashpoint.
• Debt — the binding constraint they argue decides which empire lasts.
• Compute — “short of watts” vs “short of transistors” is the national-scale version of the capacity question in At the Moment, We Have More… (tweet).
• This Is the Good Scenario? — Ezra Klein’s column two days earlier makes the same competitive-pressure argument for firms and individuals; this essay makes it for nations.
• Jacob Coxon on The Daily Show — Coxon cites the same Hugging Face incident as the reason he became worried.