- A Foreign Affairs essay argues that AI now works like an empire's supply chain. It cites a June US order that took two Anthropic models offline worldwide.
- Companies that build products on one provider's model depend on decisions made by that provider and by governments. Those decisions can change access quickly.
- Leaders should map which products rely on which model, and test whether a second model or an open-weight option could take over.
Renting intelligence means renting someone else's permissions
Many companies treat AI like a utility. They plug in and pay the bill. A recent essay in Foreign Affairs argues that this comfort hides a harder fact. The supplier may not be the one who decides whether the power stays on.
Three authors wrote it: historian Niall Ferguson, researcher Lisa Klaassen and Eric Schmidt, a former Google chief executive. They see the United States and China as building AI empires. Everyone else, they argue, is a dependent user. Their examples include "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public services on a model it cannot inspect or a firm building products on an interface that can be repriced or withdrawn overnight."
The headline claim is geopolitical. The lesson for a business is more practical. Your AI supplier sits inside someone else's chain of command.
What happened in June
The authors give one concrete case. They say a June order from the US Commerce Department led to Anthropic's two most capable models going dark worldwide within days.
The essay excerpt is brief on how. The Decoder, which covered the essay, supplies the mechanism. It reports that the order told Anthropic to cut off foreign nationals. Anthropic could not verify users' nationalities in real time. So it shut down both models, Fable 5 and Mythos 5, for everyone.
The Decoder's account of what came next is mixed. Fable 5 was freed from the restrictions roughly a fortnight on. Mythos 5 stayed restricted to a group of US organizations.
Take that sequence slowly. The order targeted one group of users. The result reached every customer, because the provider had no way to tell who was who at the moment it mattered. Any firm in any country that depended on those models would have lost access. This was an identity-checking gap turned into a full outage.
The wider argument about states and labs
Separate from the June case, the essay makes a general argument about why governments can act on AI labs. It starts with scaling laws. These are the industry's observation that a model improves with more parameters, more training data and more computing power. Progress therefore depends on energy, minerals, advanced chips and capital at large scale. The authors say that once a frontier rests on scarce inputs, producing it becomes a matter of geopolitics.
On that basis they describe labs as less sovereign than they look. In the authors' words, such firms need state authorization to move chips or build data centers. They add that an export-control regime can throttle any lab at will. Each government, they say, has chosen a different leash for its national champions.
This is the authors' framing of the structure, not an explanation of June. The sources do not say chip licences or data-center approvals caused that cutoff.
The authors also contrast the two approaches. They describe the US offer as expensive and mostly closed. They describe China's as cheap, with open-weight models that anyone can download and modify. Open weights means the trained model file itself is published. A buyer who holds a copy does not depend on a provider's switch.
What the essay does not show
This is an argument from three authors, not an audit. One incident does not make a pattern. The essay cites a single cutoff, and the Decoder reports that one model's controls were lifted within about two weeks.
The authors also say it is unclear which kind of AI will prove a bigger source of power. They call the two countries' advantages slight. Nothing here measures how many companies were harmed in June. WebPulse has no data on that, so it makes no claim about scale.
What the case does show is a design fact. Access to a leading model can end for reasons that have nothing to do with your payment history or your provider's wishes.
What leaders should ask
Start with a map. Ask your technology team which customer-facing products and internal processes call a single provider's model. Ask which of those would stop working if that model went offline for two weeks.
Then ask about the fallback. Is there a second provider that has been tested on your real tasks, not only named in a slide? Could an open-weight model handle the basic version of the work, even if it is less capable?
Ask the provider, too, how it would handle a government restriction aimed at some of its users. Can it verify who is using the service? Or would everyone lose access, as happened in June? Check the contract for how fast you would be told.
The essay's wider point is that empires run on chokepoints. For a company, the chokepoint is wherever one supplier's permission is the only thing standing between a product and a blank screen. Find that point before someone else's rules find it for you.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Foreign Affairs (Niall Ferguson, Lisa Klaassen, and Eric Schmidt).
