The microchip war: why export controls bite and how China responds

01The chokepoint: why leading-edge chips are a single-supply-chain story

Modern chips are not really manufactured by any one country — they are manufactured by a relay team in which each runner has a near-monopoly on one leg. Design tools and core intellectual property are concentrated in the United States; the most advanced logic wafers are etched almost entirely in Taiwan by TSMC, with Samsung in South Korea as the only meaningful alternative; extreme-ultraviolet lithography machines come from exactly one company, ASML of the Netherlands; and the specialized chemicals, photoresists, and polishing supplies skew toward Japan. For three decades this division of labor worked because every participant needed every other participant. That interdependence was assumed to be self-protecting: nobody would weaponize a chain they depended on. The 2018-2026 period tested that assumption directly, and the result reshaped industrial policy in Washington, Beijing, Brussels, and Tokyo. Understanding the chokepoints — not just the headline bans — is the only way to understand why the controls have teeth and where they grind.

02The tool monopoly: ASML, EUV, and why lithography cannot be copied quickly

The single most consequential chokepoint is a machine. To pattern the finest features on cutting-edge logic chips, manufacturers use extreme-ultraviolet lithography, and ASML is the sole company on Earth that builds EUV systems — each a $150-200 million assembly of hundreds of thousands of parts sourced from a specialized supplier network that took decades to assemble. EUV light is generated by hitting tin droplets with lasers fifty thousand times a second; the mirrors that focus it are the smoothest surfaces ever made. This is why the Netherlands' 2023 decision to require export licenses for ASML's advanced EUV tools mattered more than most sanctions: there is no second supplier to buy from. China's own lithography maker, SMEE, remains generations behind, and duplicating ASML means rebuilding not just one machine but the entire optics, metrology, and materials ecosystem around it. The next-generation High-NA machines, at a reported $350-400 million each, extend the lead further. A monopoly like this cannot be sanctioned around quickly — it must be engineered around, and engineering it is measured in decades, not quarters.

Approximate share of leading-edge (10nm-class and below) logic wafer capacity; compiled from SIA/BCG and industry capacity estimates. Values are estimates, not measurements.

03The control regime: what the US and allies restricted, and when

The control regime arrived in waves. In 2018-2019, the US placed ZTE and Huawei on the Entity List, cutting named companies off from American technology. The structural shift came on October 7, 2022, when the Commerce Department published comprehensive rules restricting not just advanced chips but the tools, know-how, and even US-person labor supporting China's advanced-node development — an explicit move from case-by-case denials to country-class capacity denial. The Netherlands and Japan joined in 2023, licensing advanced deposition, etch, and lithography tools from ASML, Nikon, and Tokyo Electron. October 2023 updates closed loopholes around Nvidia's China-specific A800 and H800 accelerators. December 2024 added controls on high-bandwidth memory and added roughly 140 Chinese entities to the lists. In 2025, the framework evolved again with the AI-diffusion rule and its subsequent revisions, while China answered with its own leverage: export licensing on gallium, germanium, graphite, and rare-earth processing — inputs the West also depends on. The regime is now a negotiated, escalating system rather than a one-time wall.

04China's countermove: SMIC, mature-node volume, and domestic toolchains

China's response has been to spend, substitute, and saturate. State funds led by the Big Fund III — raised at about 344 billion yuan in 2024 — bankroll the domestic toolchain, while foundry leader SMIC demonstrated the workaround path: its second-generation 7nm-class process, identified in the Huawei Mate 60 Pro teardown by TechInsights in 2023, uses DUV multi-patterning to approximate what EUV does natively, at lower yield and higher cost. The strategy accepts a permanent lag at the leading edge while dominating the mature nodes (28nm and above) that power cars, appliances, and industrial electronics — and where China's expanding capacity now sets global price floors. Domestic toolmakers (Naura, AMEC), memory makers (YMTC, CXMT), and designers (HiSilicon) form an increasingly complete, if lagging, stack. The gap is real but narrowing on Beijing's own terms: not parity at 3nm, but sufficiency — enough domestic capability that sanctions can no longer dictate what Chinese industry can build, even if it dictates how fast.

05Measuring the gap: process nodes, yields, and AI accelerator performance

How big is the remaining gap? In fabrication, the leading edge is unambiguous: TSMC and Samsung ship 3nm-class production while SMIC's most advanced verified node is 7nm-class, and Taiwan still accounts for the overwhelming majority of leading-edge logic capacity. In accelerators, the picture is subtler. Huawei's Ascend series and Nvidia's licensed China market chips compete on paper specifications, but AI training performance depends on the whole system — interconnect, memory bandwidth (where December 2024 HBM controls bite hardest), and software ecosystems like CUDA that have a decade of tooling behind them. Yield is the quiet multiplier: DUV multi-patterning produces fewer good dies per wafer, which raises unit costs exactly as volumes try to scale. Independent benchmarks of Chinese accelerators show competitive inference results and harder sledding at frontier-training scale. The honest summary: China can build nationally significant AI systems today, but building them costs more, integrates worse, and scales more slowly than the unrestricted alternative — which is precisely the outcome the controls were designed to produce.

Major US and allied semiconductor control actions, compiled from BIS and government announcements. The 2021 gap reflects no comparable structural milestone.

06The cost of decoupling: prices, supply chains, and third countries

Neither side escapes the bill. Export controls cost US firms one of their largest markets: Nvidia's China revenue share fell from roughly a quarter of revenue to a small fraction as licensing tightened, and the company now writes off China-specific inventory when rules change mid-cycle. China pays a premium for domestically substituted tools and chemicals, and its mature-node buildout pressures global pricing in exactly the segments where others planned to compete. Third countries are discovering that decoupling has gravity: economies from Malaysia to Vietnam to Mexico absorb relocated assembly and testing, while US allies balance security commitments against billions in China revenue — the Netherlands and Japan most visibly. Consumers see the cost slowly, through higher component prices and duplicated supply chains that would once have been shared. The deeper cost is efficiency itself: the pre-2018 chip industry was arguably the most optimized supply chain in economic history, and every percentage point of redundancy now being engineered into it is a real resource diverted from somewhere else.

07Scenarios: escalation, leakage, and the long-run equilibrium

Three futures dominate the planning documents. Escalation: controls tighten to a full technology denial regime — the 2025 diffusion-rule fights point this way — which maximizes friction, accelerates Chinese substitution, and hardens a two-bloc chip world with parallel standards, toolchains, and ecosystems. Leakage and drift: enforcement erodes at the edges — transshipment through third countries, licensed legacy tool exports, grey-market accelerators — and the regime becomes porous enough to slow but not stop China's advance, its effectiveness measured in quarters of delay rather than decades. Managed equilibrium: both sides formalize the division of labor they have stumbled into — China dominant in mature nodes, the US-led bloc at the leading edge — and negotiate verification regimes the way arms controllers once did. The lesson of 2018-2026 is that the chip war is not won by any single ban or breakthrough. It is a grinding industrial contest where the decisive assets are decades old: a Dutch optics monopoly, a Taiwanese fab network, an American design-software ecosystem, and a Chinese determination to make the chokepoints irrelevant. Whoever accumulates them fastest, wins the next decade.