China’s Materials Grip Is Harder to Replace Than It Looks

Permission to trade is proving a poor substitute for dependable supply. On September 4, Reuters reported that some Chinese rare-earth suppliers were refusing American orders for fear of repercussions from Beijing, despite commitments to ease trade. Meanwhile, Nvidia’s latest quarterly filing reports that some American-licensed H200 AI chips reached China, but Chinese restrictions prevented it from selling all the authorized products. The immediate problem for buyers in both countries is access that remains politically conditional. (investing.com¹)

Those developments sharpen the question behind the technology contest: which government controls something its rival cannot economically replace? I give China’s selected rare-earth controls more weight as demonstrated leverage over ongoing commercial production. Washington has a credible longer-term advantage in restricting the means of expanding frontier computing, but the evidence does not establish that its leverage will last longer. Useful Chinese AI substitution is already documented; broad materials independence is still a collection of incomplete industrial milestones. (arxiv.org³)

The comparison needs discipline. I would count replacement when a buyer can sustain acceptable output at an acceptable total cost, including redesign and replenishment. Neither country needs an identical product. But neither deserves credit for independence merely because a machine runs or a material sample passes inspection. That standard separates controls that chiefly encourage adaptation from those that preserve bargaining power.

Start with what China actually restricts. Announcement 18, effective April 4, 2025, requires export licenses for specified products involving samarium, gadolinium, terbium, dysprosium, lutetium, scandium and yttrium. It includes samarium-cobalt magnet materials and neodymium-iron-boron magnet materials containing terbium or dysprosium. The requirement is not limited to American destinations, and the announcement provides no blanket civilian-use exemption. Calling it a ban on every rare earth or every magnet exaggerates its scope. (english.mofcom.gov.cn⁷)

The more sweeping October 2025 measures remain suspended through November 10, 2026, according to China’s official explanation. That suspension did not erase the earlier licensing system. Beijing says qualifying civilian applications will be approved, but approval remains a government decision rather than an unconditional right to buy. (english.scio.gov.cn)

Washington’s chip restrictions also contain consequential openings. Effective January 15, 2026, qualifying Nvidia H200, AMD MI325X and similar exports from the United States to China and Macau became eligible for case-by-case review. Conditions include protection of American supply, restrictions on military users and uses, and independent testing in the United States. Aggregate exported processing performance cannot exceed half the corresponding performance shipped to American end-users for the same commodities through the application date. Eligibility therefore falls well short of unrestricted supply. (govinfo.gov¹⁰)

Actual deliveries complicate claims of American success. Nvidia says H200 shipments accounted for less than 1% of its latest quarterly Data Center revenue, while attributing incomplete sales of licensed products to Chinese restrictions. It also disclosed a $400 million first-half charge tied to H200 excess inventory and purchase obligations. Those are real corporate costs. I would not convert them into an equivalent loss of Chinese computing capacity: some demand was blocked by Beijing, and a supplier’s lost revenue does not measure customers’ replacement output. (investor.nvidia.com²)

The argument for greater American durability rests further upstream. The December 2, 2024 package covered 24 types of semiconductor-making equipment, three types of software tools and high-bandwidth memory, the fast memory feeding AI processors. It also reached specified foreign-produced items and provided a conditional memory license exception. The package took effect that day, with delayed compliance for certain controls until December 31. Its strategic purpose was to impede China’s ability to reproduce advanced computing hardware, not simply deny a particular Nvidia model. (bis.gov¹¹)

That is a serious objection to treating domestic Chinese accelerators as an escape. Operating existing machines cannot establish that factories can keep supplying replacements and larger fleets. Yet the operating evidence deserves more credit than the word “workaround” suggests. Huawei researchers report that their xDeepServe system serves large language models in production on Ascend hardware. A representative deployment using 16 Ascend 910C servers handled variable-length requests, reporting 900 milliseconds to the first token and an average 34.8 milliseconds per subsequent output token. This is evidence about inference, the work of answering users, rather than training a new frontier model. (arxiv.org³)

I take those results seriously without treating them as audited economics. The paper does not establish independently verified fleet costs, chip-manufacturing yields or sustainable expansion free of restricted inputs. Its engineering results narrow the claim Washington can reasonably make: denied access to a preferred accelerator need not prevent useful AI service. Whether the resulting electricity, equipment and engineering bill is manageable across buyers remains unresolved by this evidence. (arxiv.org³)

Materials buyers face an equally linked production chain. The International Energy Agency estimates that China supplied 91% of refined magnet rare earths and 94% of sintered permanent magnets in 2024. Processing separates usable elements, but replacement supply must continue through metal conversion, alloying and magnet manufacture. The agency identifies equipment procurement that can take several years, alongside gaps in practical production knowledge. Finding ore addresses only the beginning. (iea.org⁶)

Recent American disclosures show both progress and distance remaining. MP Materials reported 840 tonnes of separated neodymium-praseodymium production in the second quarter of 2026, while still describing initial magnet-production start-up ahead of commercial production. On August 19, Energy Fuels said a Japanese magnet manufacturer had fully qualified its terbium oxide: the customer had accepted it against its production specifications. But that announcement supplied no sustained commercial tonnage or yield figures. Qualified oxide and repeatable deliveries of finished magnets are different milestones. (investors.mpmaterials.com⁴)

Redesign can genuinely remove dependence. Honda and Daido announced a heavy-rare-earth-free hybrid motor application on July 12, 2016, combining new magnets with changes to the motor and rotor. Honda reported equivalent torque, output and heat resistance. That achievement deserves the same application-level credit as Huawei’s serving system. It also required coordinated redesign, rather than a universal drop-in replacement. (global.honda¹²) Nearly nine years later, Chinese licensing restrictions still forced some automakers to curtail or halt production; the subsequent export recovery followed license approvals. I read that sequence as evidence that successful substitution can coexist with persistent industrial dependence. (iea.org⁶)

Inventories should be judged against that replacement timetable. The IEA estimates roughly $200 million in net operating costs for a stockpile covering one year of exposed magnet-material imports outside China; that excludes the stockpile’s purchase price. It describes a possible buffer, not verified inventories already protecting buyers. (iea.org⁶) I would use that buffer to finance time for qualification and production, not declare the supply problem solved.

For American policymakers, the practical implication is to stop treating materials diversification as the easier half of the contest. I would prioritize customer-backed purchases, qualification and replenishable output, while reserving claims of durable chip leverage for restrictions that demonstrably constrain expansion. The September shipping refusals bring the distinction home: Reuters could not establish their total scale, but reported that some American companies had waited more than six months for mineral licenses. A diplomatic reopening leaves considerable bargaining power with Beijing when American production still depends on its next approval. (investing.com¹)