The views expressed here are my own and do not reflect those of any client, employer, or affiliated organization.
In my post on the Iran war’s second-order consequences, I traced the systems failure logic — the way the war disrupted chip input supply chains, degraded Pacific deterrence, and accelerated Taiwan’s strategic recalculation. The rare earth dimension of that story deserves its own treatment, because it reframes the US-China technology competition in a way that chip export controls alone don’t capture.
The US strategy on AI competition has rested on a hardware denial premise: if China can’t access the most advanced chips, it can’t develop frontier AI, and the US maintains technological superiority. It was a reasonable theory. It hasn’t worked.
DeepSeek is the evidence. Using Nvidia H800 chips — the export-controlled, deliberately degraded version the US permitted to sell into China — a Chinese lab produced a model that is competitive with GPT-4 class systems. The architectural choices that made this possible weren’t workarounds or compromises. They were innovations forced by constraint. A Mixture-of-Experts design that activates only 37 billion of 671 billion parameters per token, dramatically cutting compute load. A Multi-head Latent Attention mechanism that shrinks memory cache requirements by roughly 40 percent. FP8 precision training that halves memory compared to standard approaches. The entire system was pretrained on 14.8 trillion tokens using 2.788 million GPU hours — a fraction of what US labs spend on comparable training runs.
Hardware scarcity didn’t stop China’s AI development. It made Chinese AI development more efficient. That efficiency gap is now a strategic asset.
China understood this. What it did next was more sophisticated than building better models on worse chips. It moved upstream.
Semiconductor fabrication depends on rare earth inputs at multiple process steps. Dysprosium and terbium go into the permanent magnets in HDD motors and precision actuators. Lutetium and yttrium are used in thin-film deposition, polishing compounds, and target materials central to leading-edge logic and memory fabrication. These aren’t optional ingredients or legacy process holdovers. They are embedded in the manufacturing steps that produce the chips that go into every data center the US is now racing to build.
China controls over 85 percent of global rare earth processing capacity. In April 2025, it imposed export licensing requirements on seven rare earth families. In October, it expanded those controls with an extraterritorial provision that is genuinely novel: any product made outside China using Chinese-origin rare earth materials or Chinese rare earth technologies now requires an export license from Beijing’s Ministry of Commerce before it can be sold to a third country. The US flipped the upstream logic on China by restricting chip design tools and lithography access. China flipped it back by claiming jurisdiction over the raw materials that go into the chips themselves.
The one-year suspension negotiated in November 2025 has quieted the immediate alarm. It expires in November 2026. Nothing about the underlying leverage has changed.
This is where the Taiwan argument becomes more complicated than the military framing usually captures. My earlier post on Taiwan focused on the credibility collapse following the Iran war and the asset redeployment that left a physical deterrence gap in the Pacific. That analysis stands. But there is a slower, less dramatic, and arguably more durable form of coercion operating in parallel.
TSMC — Taiwan Semiconductor Manufacturing Company, the foundry that manufactures the world’s most advanced chips and whose fate I wrote about at length after the Iran war began — has approximately one to two years of rare earth reserves on hand. After that, continued operation requires supply from sources that largely don’t exist yet at the required scale. Australia’s Lynas Rare Earths operates the largest non-Chinese refining facility outside the mainland. Diversification efforts are underway in the US, Canada, and Japan. None of them will be at sufficient scale before Taiwan’s buffer runs out if China decides to enforce its export controls rather than suspend them.
Taiwan’s leadership is not naive about this. The rational calculus for a small island whose economic survival depends on a manufacturing process dependent on inputs controlled by its largest neighbor and primary security threat is not a comfortable one. Military deterrence addresses the kinetic scenario. It doesn’t address a scenario in which the inputs stop flowing and the fabs go quiet by attrition rather than by invasion.
China does not need to take Taiwan by force to absorb its most strategically valuable asset. It needs Taiwan to conclude, over time, that accommodation is more rational than resistance. The rare earth leverage, combined with the deterrence gap the Iran war created, combined with the demonstrated willingness to use economic tools as coercive instruments, makes that conclusion easier to reach each year. The noose tightens not through violence but through the accumulation of facts that make resistance increasingly costly to sustain.
The US export control strategy gave China a constraint. China converted it into an engineering advantage and a strategic one. That asymmetry is not going to be resolved by building more data centers.
This is the second in a three-part series on AI infrastructure and its supply chain vulnerabilities. The next post examines helium — the invisible input almost nobody in the AI debate is talking about.