Chip War author Chris Miller: China is 5–6 years behind TSMC, but shell company smuggling reached millions of chips

Summary

Chris Miller, author of Chip War, argues the current US-China technology competition is a modern replay of Cold War-era strategic dynamics, where access to computing power has always determined a nation's geopolitical position.

China's Semiconductor Gap

SMIC, China's leading chip fabricator, is 5 to 6 years behind TSMC in producing processor chips at quality and at scale. Chinese chipmaking equipment firms are even further back, roughly 10 years behind equivalents like ASML, Applied Materials, and Tokyo Electron. The consequence is stark: China's leading-edge fabs still run predominantly on Western tools, and last year China accounted for approximately half the global revenue of the world's major chipmaking equipment vendors.

On the memory side, the picture is more competitive. YMTC in NAND and CXMT in DRAM have come close to cutting-edge capability, though volume production remains a constraint. Huawei is considered highly capable in chip design, including rapid progress on GPUs, but is ultimately bottlenecked by SMIC's manufacturing limitations.

China's largest import by value is semiconductors from Taiwan, exceeding even oil imports. That dependency is the central driver behind years of state-backed investment. Unlike solar or EVs, where competitive pressure across hundreds of firms accelerated development, the chip sector's capital intensity has concentrated progress around a small number of state-owned players, with SMIC holding the same dominant domestic position today that it did a decade ago.

Shell Company Smuggling at Scale

A shell company network discovered and reportedly shut down in 2024 procured several million chips from TSMC, with the hardware appearing to have been diverted to Huawei. Miller describes the scale as genuinely surprising. Enforcement expectations had been set around dozens, hundreds, or at most thousands of units. Millions represents a failure of export control enforcement that undermines the credibility of the entire rule set. The US government faces a compounding problem: broad rules that are not seriously enforced produce worse outcomes than no rules at all.

Cloud-based circumvention is an emerging parallel concern. Know-your-customer requirements at US cloud providers are currently far weaker than those applied in financial services. Miller expects tightening regulation on large compute purchases within the next two years. Malaysia has drawn particular attention as a GPU transit point, with House proposals already circulating to redirect GPU allocations toward US cloud providers and away from third-country operators.

China's AI Infrastructure Is Underutilized

Despite significant chip accumulation through both domestic production and smuggling, anecdotal evidence consistently points to low utilization rates at Chinese AI data centers. State-owned telecoms managing cloud infrastructure are cited as a key source of inefficiency. Miller finds it notable that firms like DeepSeek publicly cite compute constraints given the volume of chips known to be physically present in China. The pattern suggests inefficient decentralization rather than the coordinated buildout the country's strategic posture implies.

The passport confiscation of DeepSeek researchers, mirroring earlier treatment of rare-earth experts, is flagged as a materially negative signal for China's AI talent ecosystem. Restricting international travel is a structural disincentive to research productivity.

TSMC Arizona and the Cost of Onshoring

TSMC has publicly confirmed that manufacturing yields at its Arizona fab match those in Taiwan, a meaningful milestone. However, unit costs in Arizona remain materially higher than in Taiwan, a gap that will narrow but not close. TSMC has announced plans for six additional fabs beyond the one currently operational and one under construction, though the timeline is unspecified. Tariff dynamics and customer pressure from Nvidia and the hyperscalers are the primary forces driving that expansion commitment.

The premium cost of US-based production is expected to be absorbed across the supply chain rather than concentrated at any single point. Using an iPhone as a proxy, a 30% increase in TSMC content cost on a roughly $50 chip input into a $1,000 device is not economically disruptive. Similar logic applies to Nvidia server economics.

Semiconductor Talent Is Not Power-Law Distributed

Unlike foundation model research, where a small number of individuals command offers reportedly reaching $100 million against starting salaries around $300,000, chip fabrication expertise does not concentrate in a few critical hires. Every vertical within the process stack requires deep specialization, and a fab missing 10% of that knowledge base produces chips that do not work. When TSMC brought its Arizona facility online, it required hundreds of Taiwanese engineers on-site for a node that was already two generations behind its leading-edge process. Replicating TSMC requires replicating a substantial fraction of its workforce, not recruiting a handful of principals.

Much of SMIC's progress over the past decade is directly attributable to recruiting engineers from Western, Taiwanese, and Singaporean semiconductor firms, illustrating that human capital transfer at scale is the actual mechanism of capability diffusion.

Intel and the Middle East

Incoming Intel CEO Lip-Bu Tan faces two structurally linked problems: how to configure the relationship between the manufacturing and design businesses, and how to recover market share lost to AMD in both PC and data center CPUs. A clean split between the two divisions is viewed as an oversimplification given the multi-year dependency the foundry business has on Intel's design pipeline and vice versa.

The UAE, backed by capital from hydrocarbon revenues and a stated ambition to build post-oil industries, secured a commitment of 500,000 GPUs per year following President Trump's visit in mid-2025. Miller's caution is that building data center infrastructure and affording GPUs is the straightforward part. Building viable businesses at the application layer on top of that infrastructure is the harder and more valuable problem, one that most governments investing in AI compute have not adequately addressed.