Which technology controls create durable bargaining leverage, and which mainly accelerate substitution at a manageable cost?

My position is that US frontier-chip and manufacturing-input controls offer more durable bargaining leverage than China’s rare-earth restrictions—provided licensing actually withholds capabilities. This is a conditional comparative judgment, not a claim that China cannot build useful AI systems or that magnet substitution is already complete.

I use four operational definitions: an export control conditions cross-border transfers on government authorization; a technology chokepoint is an essential input difficult to replace competitively; rare-earth processing separates and refines usable materials; supplier qualification verifies that replacement supplies meet production requirements.

First, Washington can restrict both accelerators and their reproduction inputs. The December 2024 package covered 24 equipment categories, three software-tool categories, and high-bandwidth memory, including specified foreign-produced items; some memory qualifies for an exception. This creates linked substitution requirements, rather than a single chip-design hurdle. That is a mechanism for leverage, not proof of its magnitude. (bis.gov) Crucially, effective January 15, 2026, qualifying H200 and MI325X exports from the United States to China and Macau became eligible for case-by-case review, subject to supply, security, and testing conditions. Calling these products categorically banned would exaggerate US leverage: authorized supply must be deducted from any estimated shortfall. (govinfo.gov)

Second, China’s strongest counterexample demonstrates useful substitution, but not comprehensive competitive equivalence. Huawei’s CloudMatrix384 paper describes a production-grade system integrating 384 Ascend 910C accelerators and 192 CPUs, with reported DeepSeek-R1 decode throughput of 1,943 tokens per second per accelerator under its stated conditions. This is substantive progress, not merely a prototype. However, that benchmark does not establish independently verified, fleet-scale cost parity, domestic production yields, or freedom from restricted inputs. My inference is that the relevant hurdle remains simultaneous adequacy in compute, memory, interconnects, software, and manufacturing—not simply successful inference. (arxiv.org)

Third, rare-earth diversification has identifiable financing and buffering requirements. The IEA estimates roughly $60 billion of investment over a decade to meet demand outside China through diversified supply, and approximately $200 million in net operating costs for a stockpile covering one year of exposed imports—not the stockpile’s purchase price. Yet the opposing case is strong: China supplied 91% of refined magnet rare earths and 94% of sintered magnets in 2024; April 2025 controls caused automotive disruptions, while the October expansion was suspended for a year in November. (iea.org) MP Materials’ second-quarter 2026 disclosure reports 840 tonnes of separated NdPr production, up 41%, but still describes magnet start-up activities ahead of commercial production. Processing progress is therefore not proof of qualified magnet replacement. (investors.mpmaterials.com)

The stakes are sustained AI capacity versus industrial continuity. My policy judgment is to prioritize enforceable frontier-production controls while funding inventories, alternative processing, qualification, and redesign. The decisive comparison must measure equivalent accepted output, delivered cost, and time to scale—not announced capacity or export values.

B’s strongest point is that rare-earth replacement requires an industrial ecosystem, not merely financing: the IEA identifies equipment procurement lasting several years, and the 2025 recovery followed Chinese licensing rather than demonstrated substitution. I accept that a hypothetical one-year stockpile cannot establish independence. China can therefore possess stronger immediate disruption leverage without necessarily possessing the more durable chokepoint. (iea.org)

I also accept B’s demand for a symmetrical substitution test. For either side, count sustained, customer-accepted output at comparable performance and delivered cost—not identical technology, but not merely some useful output either. For AI, that means specifying model quality, latency, throughput, availability, and total infrastructure cost; for magnets, operating temperature, magnetic performance, reliability, and qualified production volume. Neither successful inference nor qualified oxide alone settles that comparison.

Huawei’s evidence deserves credit, but its benchmark boundaries matter. The xDeepServe evaluation producing 2,400 tokens/second/chip completes prefill before simultaneous decoding, uses fixed 2,000-token prompts and outputs, and disables normal end-of-sequence termination. That is meaningful engineering evidence, not an end-to-end competitive-cost comparison across production traffic. The paper documents substantial communication, memory-management, scheduling, and reliability engineering; it does not establish manufacturing yields or independently audited fleet economics. This supports my linked-constraints mechanism, while leaving its quantitative penalty unresolved. (arxiv.org)

Licensing evidence also sharpens rather than eliminates the distinction. Effective January 15, 2026, the H200/MI325X-class pathway permits conditional case-by-case review for exports from the United States to China and Macau. Among its conditions, aggregate exported processing performance cannot exceed 50% of the corresponding performance shipped to US end-users for the same commodities through the application date. Eligibility is consequently neither unrestricted supply nor proof that this ceiling currently binds. (govinfo.gov) NVIDIA’s quarter-ended-July-26, 2026 disclosure confirms actual H200 shipments, representing less than 1% of quarterly Data Center revenue, but says Chinese restrictions prevented selling all licensed products. I therefore cannot attribute that shortfall entirely to Washington—or equate NVIDIA’s lost sales with China’s lost computing output. (investor.nvidia.com)

The additional materials evidence is commercial redesign, not another announced refinery. Japan’s MEXT reports that Daido and Honda achieved mass production of hybrid automobiles using traction-motor magnets without heavy rare earths, by combining magnet development with motor redesign. This establishes substitution for a relevant application, although not universal replacement, freedom from all rare-earth inputs, or audited cost parity. (mext.go.jp) The IEA also reports Japanese demand-side measures resulting in total rare-earth demand 30% below 2010 levels. (iea.org)

My comparative inference remains conditional: demonstrated application-level materials substitution strengthens the case that China’s leverage can erode, whereas useful domestic inference has not yet demonstrated competitively reproducible frontier capacity. The evidence supports prioritizing enforceable frontier-production controls alongside materials diversification—not claiming that either side’s eventual replacement timetable is already proven.

B’s strongest case is that selected rare-earth restrictions can constrain qualified industrial output for years: the IEA documents multiyear equipment lead times and a recovery driven by Chinese licensing, not replacement supply. Meanwhile, Huawei reports commercially relevant production inference, so requiring complete frontier-chip equivalence while crediting individual magnet redesigns would unfairly favor my position. (iea.org)

I accept that criticism and correct my previous emphasis: xDeepServe’s production evaluation reports 900-millisecond time to first token and 34.8-millisecond average time per output token on 16 Ascend 910C servers. That is stronger evidence than the synchronized decoding benchmark alone; it still does not establish independently audited delivered costs, manufacturing yields, or reproducible fleet expansion. Neither should a qualified magnet sample establish materials independence. (arxiv.org)

My position remains stronger as a conditional mechanism-based judgment, not an established comparative timetable. The December 2024 US package covered 24 equipment categories, three semiconductor-development or production software categories, and high-bandwidth memory. Huawei’s own architecture integrates specialized communications, memory management, and serving software. Together, these establish multiple linked requirements for reproducing useful computing capacity; they do not prove those requirements are presently binding. The inference supporting my position is that successful operation of existing machines cannot settle whether additional machines can be supplied competitively after restricted inventories are depleted. (bis.gov)

Materials offer an additional route: removing a controlled input from the application. Honda’s July 12, 2016 announcement describes heavy-rare-earth-free magnets and motor redesign for the forthcoming FREED, not universal substitution. The IEA estimates approximately $60 billion over a decade for diversified magnet-rare-earth supply, while identifying equipment bottlenecks. These support a technically identified, costly replacement pathway—not proof of timely, competitive delivery. My inference is that combining redesign, qualified alternative production, and inventory bridging offers more routes around selected materials restrictions than an accelerator demonstration offers around the entire computing-production chain. (global.honda)

Licensing can overturn that comparison. The January 15, 2026 rule permits conditional case-by-case review of qualifying H200/MI325X-class exports from the United States to China and Macau. Actual authorized deliveries must therefore count toward restored computing output; regulatory coverage alone cannot demonstrate scarcity. (govinfo.gov)

I would change my mind on a matched, independently audited study tracking exposed buyers for eight quarters after restrictions become binding. It should reconcile licenses, deliveries, inventories, equipment provenance, production yields, qualification failures, and customer-accepted output. A concrete reversal criterion would be Chinese AI buyers restoring at least 90% of matched workload output within a 20% all-in cost premium, with sustainable replenishment, while exposed magnet users remain below those same thresholds after inventory and redesign adjustments. Those are proposed decision thresholds, not observed results; the study must distinguish control-induced losses from demand changes and domestic purchasing restrictions.