America’s Freedom to Slow AI Is Priced in Magnets

By Carlo J.V. Caro

For a week in May 2025, Ford stopped building Explorers in Chicago because it could not get rare-earth magnets. China had restricted their export the month before. Two months later, Washington moved to resume sales to China of Nvidia’s H20 AI chip, which it had restricted in April. Commerce Secretary Howard Lutnick tied the reversal to the rare-earth negotiations: “We put that in the trade deal with the magnets.” An AI-chip restriction and a car part had been exchanged for each other. That trade, not the state of the technology, is the real test for anyone who wants America to slow the development of its most advanced AI systems.

In September, Anthropic Chief Executive Dario Amodei argued for pacing advances in frontier AI so that safety work has time to catch up. An extra year or two, he suggested, could help. Progress would not stop; its leading edge would be held back. But the proposal is feasible only if America keeps its lead over China, and America keeps its lead partly through restrictions on advanced chips, chipmaking equipment, and access to computing power.

That makes industrial resilience part of the safety strategy, not somebody else’s trade problem.

Having a lead and keeping the policies that protect it are different problems, and they run on different clocks. The first clock measures how fast China can close the capability gap. The second measures how long Washington will keep restrictions in place once Beijing makes them expensive somewhere else.

The AI debate watches the first clock almost exclusively, and the first clock runs in years; the pacing argument is about buying one or two of them. The second clock runs in weeks. Washington was reversing the H20 restriction within three months of imposing it. In September 2025, Commerce adopted a rule automatically extending certain export restrictions to companies at least 50% owned by listed entities. Roughly six weeks later, Washington suspended the extension for a year, as part of a broader agreement in which China also suspended its expanded rare-earth controls. The lesson is not that the controls that constrain China most will fall next. It is narrower and worse: China does not have to win the technology race to end American restraint. It only has to make restraint cost something Washington will not pay.

The benefit of pacing goes to laboratories, on time. The cost of the policies that permit pacing lands wherever Beijing aims its retaliation. In 2025 that was an assembly line in Chicago. Beijing decides which American industry pays for American AI policy. The laboratories do not.

A president who answers to manufacturers, workers, and consumers as well as to national security cannot treat every restriction as untouchable. Neither of the 2025 concessions was a payment for an AI slowdown, and perhaps both were worth making. That is precisely the problem: a strategy for restraint cannot assume the government will always rank its preferred competitive conditions above every other interest, because it demonstrably does not.

Nor is the damage confined to the concession itself. Once Washington has shown it will trade a chip restriction for a magnet shipment, the expected return on producing the next magnet shortage rises. A concession is also a price list.

The answer is not to demand more resolve from future presidents. It is to make the restrictions that matter cheaper to hold.

Rare earths show what that takes. The International Energy Agency puts China’s 2024 share at about 60% of magnet rare-earth mining, 91% of separation and refining, and 94% of sintered permanent-magnet production. A new mine outside China solves the smallest piece of the problem; its output must still be separated, refined, alloyed, and formed into a magnet that meets a customer’s specification. Washington has begun. Project Vault, launched in February with an approved Export-Import Bank loan of up to $10 billion and nearly $2 billion in private capital, is meant to build a strategic reserve of critical materials for American manufacturing. That is serious money.

A warehouse of rare-earth oxide does not keep a plant open if no facility outside China can turn that oxide into the magnet the line needs before its inventory runs out. The distinction that matters is not between having material and lacking it. It is between having material and being able to use it in time. The right buffer follows the production line backward: for some manufacturers, a qualified component; for others, intermediate material paired with assured processing capacity. The reserve should be judged by the production it can sustain under a realistic disruption, not by dollars committed or tonnes stored.

New capacity is coming, but not soon. MP Materials’ planned 10X magnet campus in Northlake, Texas, is scheduled to begin commissioning in 2028. That matters for the next decade. It cannot replace a shipment withheld this quarter. Both are needed: a bridge through the near term and durable American or allied alternatives beyond it. Stockpiles without replenishment merely schedule the next crisis.

After 2008, regulators stopped taking banks’ word that their capital could absorb a shock. They specified the shock and made the banks prove it. A pacing strategy rests on export controls the way a bank’s lending rests on its capital, and the controls should have to pass the same kind of test before anyone builds a safety timeline on them.

Congress should require two connected assessments. The first would stress-test the technology argument. Instead of one reassuring estimate of America’s lead, Commerce and the intelligence agencies should evaluate specific capabilities under specific scenarios: controls held; selected licenses granted; enforcement weakened; Chinese substitutes improved. Which gaps survive each scenario, and with what confidence?

The second would stress-test industrial continuity under retaliation. For each dependency capable of generating pressure for concessions, officials should establish how long production can continue, which alternatives have actually been qualified, and whether reserves can reach the point of use before a line stops. Sensitive findings can stay classified or commercially protected.

Those findings should direct the money. The priority is not every commodity on a long list. It is the handful of bottlenecks where a manageable investment in inventory, processing, or supplier qualification would substantially reduce Beijing’s leverage.

The bill should also say who pays. Manufacturers should fund ordinary commercial resilience. The public should fund the additional protection that serves national purposes beyond any company’s balance sheet. And AI companies that lobby for restrictions whose survival depends on those bottlenecks should help pay to remove them; they are asking other industries to absorb the retaliation those restrictions invite. That does not make laboratories responsible for rebuilding American industry. It recognizes that the freedom to be careful may have to be purchased outside their own campuses, and purchased before a shipment is withheld rather than after.

None of this settles whether slowing frontier AI is the right choice. Better testing and safeguards need not wait for a magnet plant. And the requirement is not partisan to caution: those who want America to accelerate depend on the same controls, and the same resilience preserves their choice too. What it does settle is what a credible version of either strategy owes: an account of how the country will keep the policies that make the strategy possible, when the pressure on those policies arrives in weeks and the payoff is counted in years.

An AI laboratory can choose to wait. Making that choice last means keeping an assembly line running.

Carlo J.V. Caro is a strategic analyst trained in international security at Columbia University. He previously served as Vice President of a private equity firm.

This article was originally published by RealClearDefense and made available via RealClearWire.