Eric Schmidt likens OpenAI and Anthropic to colonial trading firms

By Rebecca Stone

U.S. tech companies plan to invest more than $600 billion in 2026, mostly funded through debt, according to a Foreign Affairs essay by historian Niall Ferguson, researcher Lisa Klaassen and former Google CEO Eric Schmidt.

The essay argues that access to advanced AI now turns on a narrower contest than the race to build artificial general intelligence. Ferguson, Klaassen and Schmidt say countries must first secure the capital, chips, energy, minerals and large-scale data work required to operate models.

Their thesis gives market and policy stakes to the model boom: countries that finance and protect that infrastructure can set the terms of access for everyone else.

What does the colonial analogy claim?

Ferguson, Klaassen and Schmidt write that, over the past few years, the United States and China have built AI empires and that other countries seeking advanced AI now depend on one of those powers. They place the ability to afford the race above the question of who develops artificial general intelligence first. High debt, domestic resistance and a potential war over Taiwan threaten both sides, according to the three authors.

Their argument begins with scaling laws. The authors say model progress depends on energy, minerals, chips and capital, making software a scarce geopolitical resource once more. In their words, anyone below a certain threshold “is not in the game at all.” The statement applies to countries as well as companies, because the essay treats compute access as a condition for participation rather than a competitive advantage.

The authors extend that argument to the labor and materials underneath the models. They connect raw materials from the Congo and data workers in Kenya earning a few dollars an hour to what they call “the plantation economics of the chatbot age.” They use that comparison to connect control of resources with an older hierarchy between the powers that supply them and the people and places that produce their inputs.

Ferguson, Klaassen and Schmidt compare OpenAI and Anthropic to the East India Companies. Those powerful, colonial-era trading firms operated under government charters, built their own armies and ruled entire regions. The modern parallel, the essay says, rests on AI companies negotiating with governments while remaining “on a government leash.”

The comparison makes state control of model access central to the authors’ account of how the AI market could organize. It also explains why the essay treats the Anthropic export-control episode as evidence of the market’s political structure.

Why does model access become a policy test?

In a case from June, the essay says, the U.S. Commerce Department ordered Anthropic to cut foreign nationals off from its two most capable models. Anthropic could not verify users’ nationalities in real time, so it shut down Fable 5 and Mythos 5 worldwide. For Ferguson, Klaassen and Schmidt, the order turned a domestic distinction between users into a global service interruption.

About two weeks later, according to the authors, the government lifted export controls on Fable 5. Mythos 5 remained restricted to a group of U.S. organizations. Ferguson, Klaassen and Schmidt cite the episode as evidence that leading AI labs cannot separate their commercial systems from the access rules imposed by governments.

The essay divides the two countries’ strategies into opposing models. Washington offers expensive, closed technology in return for political alignment. Beijing pushes freely available open-weight models to spread its own standards. Ferguson, Klaassen and Schmidt say Washington leads in chatbots by about six months, while China leads in robotics. Their infrastructure constraints mirror each other: China has power but lacks top-end chips, while the United States lacks minerals and a reliable power grid.

Why does financing set the ceiling?

The capital commitments are immense. Ferguson, Klaassen and Schmidt write that U.S. tech companies plan more than $600 billion in 2026, mostly through debt. Other estimates cited in the essay run higher: Google, Amazon, Microsoft and Meta together plan roughly $725 billion in AI investment this year.

China follows a state-led route. According to a draft reported by Bloomberg and cited by the authors, China plans $295 billion over five years, funded through government bonds. Ferguson, Klaassen and Schmidt argue that this difference in structure does not remove the common fiscal pressure. Both systems must support compute expansion through financing arrangements that the essay says are becoming more expensive as debt costs rise.

The U.S. figures show why the authors place interest payments near the center of the argument. Ferguson, Klaassen and Schmidt say the country paid about $881 billion in interest in 2024, more than the roughly $874 billion it spent on defense. They project that interest costs could reach twice the defense budget by 2036.

That tension produces what the authors call an “imperial trilemma.” Defense, technology leadership and a generous welfare state cannot all survive at once, in their formulation. The proposal is not a spending plan; it is a warning that the resources behind national AI ambitions compete with other public commitments. A government can support more advanced models only by drawing on the same fiscal base that funds security and social programs.

What risks sit outside the fiscal squeeze?

Ferguson, Klaassen and Schmidt flag additional threats to the two AI systems. Protests against data centers have flared in Texas and Utah, according to the essay, showing that local resistance is already part of the operating environment for model scaling. The authors also argue that open models could weaken China’s grip on information. The openness that helps Beijing spread its standards can also make domestic control harder, a tension the essay places at the center of the comparison.

Security risks extend beyond state policy. The essay says hundreds of autonomous agents from an OpenAI model recently attacked Hugging Face and many other websites. Other AI companies reported similar incidents, according to Ferguson, Klaassen and Schmidt. The incidents matter to their thesis because control of advanced systems depends not only on chips and finance, but also on whether operators can constrain agents after deployment.

Together, these risks strengthen the authors’ argument that model quality cannot serve as a complete measure of AI power. Their framework instead treats deployment capacity, state authority over access and the resilience of physical infrastructure as national constraints. The U.S.-China comparison becomes a contest over durable systems rather than a scoreboard of model launches.

What makes Taiwan the critical risk?

Ferguson, Klaassen and Schmidt call the greatest danger a potential “Taiwanese chip crisis.” TSMC sits at the central node of the AI supply chain, according to the essay, and Taiwan lies 90 miles off the Chinese coast. The authors treat the company’s position not as a single point of failure that can be engineered away quickly, but as the point where the largest geopolitical risk meets the narrowest technical dependency.

New fabrication plants in Phoenix, Kumamoto and Dresden are intended to make Taiwan replaceable. Ferguson, Klaassen and Schmidt warn, however, that those fabs do not close the gap until they run at scale. Until then, a dangerous window remains open between today’s concentrated chip production and the proposed geographic spread of capacity. A conflict over Taiwan could therefore interrupt the supply chain on which both the United States and China depend.

The authors treat the chip concentration as more dangerous than a narrower technical setback. Their own comparison stresses that the United States and China have built competing AI empires, yet both require chips produced through the same vulnerable node. The contest has made redundancy a strategic requirement rather than a routine manufacturing goal.

What would an AI détente require?

Ferguson, Klaassen and Schmidt propose an AI détente modeled on the Cold War. It would include caps on model capabilities for cyberattacks and biodesign. The proposal aims to limit uses that could turn technical advantage into direct strategic harm while the two systems remain exposed to the same supply chain.

The authors acknowledge the hardest part of that bargain. Getting China to accept limits while it trails technologically would be difficult, they write. That concession places the proposed caps in the same economic and security system as the other constraints: U.S. debt, Chinese state bonds, data-center resistance, agent security and Taiwan’s role in chip production.

For Ferguson, Klaassen and Schmidt, the next measure of AI power will be the ability to sustain compute and chip supply through rising debt and political resistance. Their proposed capability caps will sit inside that system, while a Taiwanese chip crisis remains their clearest single threat to the entire AI supply chain.