Empires of AI: Why Eric Schmidt Compares OpenAI and Anthropic to East India Companies

Eric Schmidt, the former Google CEO, has put his name to one of the more provocative framings of the AI race this year. In Empires of AI, published by Foreign Affairs on October 20, 2026, he joins historian Niall Ferguson and researcher Lisa Klaassen to argue that AI is "by its nature an imperial technology." The US and China, they say, have each assembled a new kind of empire, and everyone else is a dependent. The Decoder summarized the piece under a headline comparing OpenAI and Anthropic to modern colonial trading companies.

The essay is opinion, not data. But it comes from a prominent insider, it lands in the middle of a year of export bans, data-center protests and record capital spending, and it gives builders a useful lens on why the models they depend on can change access overnight. This post walks through the argument, separates what is sourced from what is interpretation, and ends with what it means in practice.

TL;DR: the argument at a glance

Why do the authors call AI an imperial technology?

Chips moving along a supply track, showing the AI chokepoints in the Empires of AI essay

The authors give several reasons. The first is scarcity. From the 1970s to the 2010s, they write, software escaped the logic of finite resources because it could be mass-produced at almost zero cost on cheap silicon. Scaling laws end that reprieve: model performance climbs with parameters, data and compute, so progress depends on energy, critical minerals, advanced chips and capital at massive scale. Once a frontier depends on scarce inputs, they argue, production becomes geopolitics. Below a certain threshold of energy, capital and compute, a country "is not a weaker player; it is not in the game at all."

The second is a center-periphery structure. Control of frontier AI is concentrated in a handful of firms in the US and China. Everyone else is a dependent user, "whether a government running public services on a model it cannot inspect or a firm building products on an interface that can be repriced or withdrawn overnight." Inputs flow the other way: minerals from the Democratic Republic of the Congo and Indonesia, and data labelers in Kenya and the Philippines paid a few dollars an hour, which the essay calls "the plantation economics of the chatbot age."

The third is chokepoints. The authors note that British power in the nineteenth century came less from territory than from controlling pipelines such as coaling stations, the Suez Canal, undersea cables and sterling finance. The AI stack, they say, offers an even denser lattice, with chips, minerals and models that the metropoles can switch on or off.

How do OpenAI and Anthropic fit the East India Company analogy?

This is the passage that made headlines. The essay says the great corporation-states of the colonial era minted coins, fielded private armies and bargained with kings as equals. Today's counterparts, in the authors' words, are frontier AI companies such as OpenAI and Anthropic, which command the human talent, data, models and finance on which advanced intelligence depends. They are "arbiters of knowledge," increasingly determining what is discovered and how fast.

The analogy has two halves. The labs bargain with governments that lack a frontier model of their own, trading model access, compute allocations and data-center investment for land, power, sovereign capital and regulatory accommodation. The essay reads the May 2025 Gulf deals, where chip and AI firms offered their most advanced technology in exchange for petrodollars and cheap energy, as imperial concessions in all but name.

But the labs are also, like the East India Companies and the Rothschild banks, subordinate to the home government. They cannot move a chip or build a data center without state authorization, and they operate under an export-control regime that can throttle any of them.

The Anthropic example from June

The authors use a concrete case. In June, they write, the US Commerce Department told Anthropic to suspend all foreign nationals' access to its two most capable models, and those models went dark worldwide within days. We covered the ban at the time in our report on the US government's order against Fable 5 and Mythos 5, the Mythos 5 trusted-partner arrangement, and the executive order that followed. According to The Decoder, export controls on Fable 5 were lifted about two weeks later, while Mythos 5 stayed restricted to a group of US organizations; see Is Fable 5 back? for that thread and the original Fable 5 and Mythos 5 launch.

For a developer, this is the most practical sentence in the essay: a model your product depends on can be withdrawn by a government decision, not by a pricing page.

Two empires, two strategies

The essay contrasts the American and Chinese approaches, and several of its claims are worth listing with their attribution.

The authors argue that an empire built from low-cost components with no barriers to access is likely to win the allegiance of most economies, while a premium restricted one attracts only those who can afford it and are permitted access. That echoes themes we have followed: the China AI playbook of free models and cheap compute, and the SemiAnalysis comparison of China's and America's compute capacity.

The essay also notes Reuters reporting that Moonshot AI is in talks with Amazon, Microsoft and Alphabet about sharing revenue from hosting its Kimi K3 model. The authors present this as evidence that the imperial dynamic is not purely one-directional.

Why "who can bankroll the race" is the key line

The headline claim of the piece is that the race may not be decided by breakthroughs. Their historical template is the Royal Navy's Dreadnought of 1906, which made the existing fleet obsolete, set off an arms race that nearly doubled UK naval spending between 1900 and 1914, and collided with new social spending such as the 1908 Old Age Pension Act.

The Decoder's summary adds the numbers the authors use for today:

• US corporations alone plan to spend more than $600 billion on AI in 2026, mostly debt-financed. The Decoder separately reports estimates around $725 billion for Google, Amazon, Microsoft and Meta combined.
• China's state-driven plan reportedly totals $295 billion over five years, funded by government bonds, based on a Bloomberg report of a draft. Treat that as a reported draft, not a confirmed program.
• US interest payments in 2024 (about $881 billion) exceeded defense spending (about $874 billion), and by 2036 interest could reach twice the defense budget, according to the authors.

From this they derive an "imperial trilemma": defense, technology leadership and a generous welfare state cannot all be sustained at once. We keep finance out of this blog unless it changes what builders pay or can access, but this one does: if AI capex is debt-funded and politically contested, pricing, rate limits and availability are tied to forces well outside any lab's roadmap.

What could stall both empires?

A power pylon feeding small houses, illustrating the energy limits on Empires of AI

The essay lists several risks beyond money:

• Data-center backlash. It cites protests against data centers in Texas and Utah. Our coverage of the ABC/Washington Post/Ipsos poll on data centers and the Thane, India data-center protests shows the same resistance elsewhere.
• Open models weakening Chinese information control. The authors suggest freely downloadable models could erode Beijing's own grip.
• Loss of control of the technology. The Decoder says the essay points to recent incidents involving autonomous agents built on OpenAI models.
• Taiwan. TSMC, the central node of the supply chain, sits about 90 miles off the Chinese coast. New fabs in Phoenix, Kumamoto and Dresden are meant to make production replaceable, but a risky gap remains until they run at scale. Our reporting on Taiwan making AI chip exports to China a crime and the October chip supply chain results give the near-term context.

What do the authors propose?

Per The Decoder, the authors suggest a Cold War style AI detente, including limits on model capabilities for cyberattacks and biodesign. They concede that persuading China to accept limits while it trails technologically would be difficult. The paywalled full text may add detail that we could not verify, so we are relying on the essay's open introduction and the Decoder summary for the proposal.

What is verified and what is interpretation?

It helps to sort the essay's claims.

Critics could reasonably push back on several points. Frontier labs are far more constrained by competition than the East India Companies were by their charters. Many countries hold meaningful AI capacity, including Europe and the Gulf states. And an author's analogy is not evidence of a mechanism. We also note that Schmidt has been an investor and adviser in the AI sector, which is context for a reader, not a reason to dismiss the argument.

What this means for people who build with AI

1. Plan for access risk. The June episode showed a flagship model can disappear worldwide within days. Keep an abstraction layer and a tested fallback model. Our guide to open-source alternatives to Fable 5 and GPT-5.6 is a starting point.
2. Expect price and limit volatility. Debt-funded capex and energy limits are the constraint the essay highlights; they show up for users as usage caps and repricing.
3. Watch the open-weight track. If the authors are right that open models are a state-backed strategy, the supply of cheap capable models will keep growing, with governance questions attached for regulated users.
4. Treat sovereignty as a procurement question. Governments and firms that cannot inspect a model they depend on are exactly the dependents the essay describes.
5. Read the original. The full text is subscriber-only at Foreign Affairs, but the opening sections are public, and they are worth the time.

If you are building agents on any of these models, the security side matters too: AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

Bottom line

Empires of AI will be quoted for its East India Company line, but its more useful claim is that the AI race is a financing and supply-chain contest with political strings attached. Whether or not you accept the imperial framing, the practical lesson holds: the models you build on sit inside a web of export rules, power grids, debt and chips that can change faster than a product roadmap.

Related reading

• US government bans Fable 5 and Mythos 5: the export-control story
• Is Fable 5 back?
• China's AI playbook: free models, cheap compute
• China 24 GW compute versus US 56 GW
• Taiwan to criminalize AI chip exports to China
• AI chip supply chain: TSMC, Samsung, GlobalFoundries
• Thane data-center protests

Details reflect the Foreign Affairs essay and The Decoder's summary as of October 25, 2026; figures are the authors' or their cited sources' and may change.