Two AI Empires, and the Question That Decides Them: Who Can Afford to Keep Paying?

The situation

In a few years, the United States and China have each built something that behaves less like a tech industry than like an empire. Whoever makes the leading AI models — and owns the chips, power plants, and capital behind them — sets the terms on which every other country gets to use them. Without a single threat or alliance, these two can make everyone else dependent on them.

The model: how empires actually end

An empire is a polity whose authority reaches over vast territories and peoples without their consent — a center commands, a periphery obeys. Empires recur throughout history for an economic reason: economies of scale. Pool resources across huge distances and you can deliver security and order more cheaply than a patchwork of rivals can. That efficiency is the engine of expansion.

But the useful lesson isn't how empires rise; it's how they end. Absent a stronger rival, empires rarely fall to invasion. They fall because holding the periphery costs the center more than the center can bear. That strain takes three forms, each with a clear historical case:

• Financial. Britain won both world wars and then had to liquidate its empire to pay the bill — the debt forced London to choose a welfare state at home over dominions abroad.
• Institutional. The Qing dynasty collapsed in slow motion because it could not industrialize, reform its bureaucracy, or renew its own legitimacy.
• Technological. In 1906 the Royal Navy launched HMS Dreadnought, the first "all big gun" battleship, so superior it instantly made every other battleship — including Britain's own fleet — obsolete. A century of accumulated advantage evaporated, because everyone now restarted from the same line. Germany could suddenly compete. Britain nearly doubled naval spending from 1900 to 1914 to stay ahead, just as it tried to fund its first old-age pensions.

That last case is the warning for this one. The decisive question in the US–China AI race may not be who reaches human-level intelligence first, but who can afford to keep paying for the lead.

The mechanism: why AI maps onto this shape

Two features make AI behave like silver did for Habsburg Spain or oil did for the American century — not merely a product that enriches its owner, but the operating system everyone else runs on.

1. Scarcity has returned. From the 1970s to the 2010s, software escaped the economics of empire: copy it for nearly nothing, run it on cheap, abundant chips. That reprieve is over. Modern AI follows scaling laws — a model improves roughly in proportion to its number of parameters, the data it trains on, and the raw computing power ("compute") spent training it. All three demand enormous energy, rare minerals, advanced chips, and capital. The moment a frontier technology depends on vast scarce physical inputs, making it becomes geopolitics.

2. A built-in center and periphery. Frontier AI is controlled by a handful of firms in two countries. Everyone else is a user who cannot see inside the model they depend on and can be repriced or cut off overnight. The inputs, meanwhile, come from the edge: cobalt and nickel from Congo and Indonesia; low-paid "ghost workers" in Kenya and the Philippines labeling training data for a few dollars an hour. Extract at the edge, refine in the metropole, sell back at a markup — the oldest imperial shape there is.

And like older empires, the whole thing hangs on chokepoints, not territory. British dominance in the 1800s came less from owning land (much of it worthless) or enforcing intellectual property (barely tried) than from controlling shipping lanes, undersea cables, and finance. Control the chokepoints — chip supply, fabrication plants, cloud infrastructure, capital — and you don't need to own what runs downstream.

Put the two halves together and the decline mechanism becomes concrete: scarcity plus chokepoints give the center its power, but holding that lead means paying the Dreadnought bill — spending without limit to stay ahead of a rival who can reset the race with one breakthrough.

What's different: the speed

The imperial arc — expansion, overreach, strain, unraveling — used to take centuries. Rome had hundreds of years; Britain had one. The AI empires are running that cycle at machine speed. Their endowments differ: the US leads in capital markets, frontier labs, and transistors; China leads in state direction, cheap energy, and the sheer electrical watts to run everything. The open question is whether either converts its strengths into a durable order — or whether both run aground paying for the race.

The other view

This is an interpretation, and a contested one. The strongest counter-case: AI may not stay scarce. If training costs fall sharply, if smaller open-weight models prove "good enough" for most tasks, or if the leading edge decays within months rather than years, the empire analogy breaks — dependence only bites when the center controls something the periphery genuinely cannot reproduce. Dreadnought cuts both ways: a breakthrough that erases your rival's lead can erase yours just as fast. What is factual is the extreme concentration of frontier models and advanced-chip production today. What is interpretation is that this concentration hardens into lasting dependence rather than erodes.

Where else this appears

The "scarcity returns, geopolitics follows" mechanism isn't confined to AI. Watch for it wherever a technology shifts from cheap-and-copyable to input-constrained: battery supply chains are already reorganizing great-power politics around lithium and cobalt. And in your own working life it holds in miniature — whoever controls a chokepoint (the one database, the one vendor relationship, the one skill nobody else has) holds power over everyone downstream, no ownership required.

What to watch

• The cost curve. If training and inference costs keep falling fast, the periphery escapes. If they hold or rise, dependence deepens.
• Chokepoint fights. Export controls on advanced chips, restrictions on chipmaking equipment, and electricity bottlenecks are the live edges of this empire.
• Who blinks on spending. The Dreadnought parallel says the loser may be whoever can't sustain the capital burn — watch for a power pulling back on data-center buildout under fiscal strain.

Expectation: ~65% that frontier AI capability remains concentrated in US- and China-based firms through at least 2030, with other states acting as dependent users rather than producers. The likeliest thing to falsify it is not a political challenger but a technological one — a cheap, open model that collapses the scarcity the whole structure rests on.

Distilled from Foreign Affairs