343 Machines That Print AI Chips: New Report Puts Hard Numbers on China's ASML Stockpile

The single most consequential number in the AI chip war this week is not a benchmark score or a funding round. It is 343.

That is the central estimate, published by the Center for Technology & Statecraft (CTS) on October 2, of how many immersion deep-ultraviolet (DUVi) lithography systems Chinese-owned fabs had accumulated by early 2026 — the majority of them ASML’s Twinscan NXT:1980i, a machine that remains legal to sell into China and remains perfectly capable of printing the 7nm-class logic and high-bandwidth memory inside Huawei’s Ascend AI accelerators.

The report, titled DUV Immersion Lithography: The Foundation of the US AI Advantage and authored by Nicholas Brown and Saif M. Khan, is the most detailed public accounting yet of the “DUVi loophole” — the gap in US, Dutch, and Japanese export controls that bans extreme-ultraviolet (EUV) tools but lets older immersion systems flow into the very fabs they were designed to starve.

What the report actually found

The headline numbers, drawn from Appendix E’s import modeling:

• 343 DUVi systems imported for Chinese-owned fabs between 2012 and 2026Q1 (central estimate; 80% model range of 312–383)
• ~270 of them are NXT:1980i, accounting for roughly 80% of the advanced DUVi stockpile at Chinese entities
• 22 NXT:2050i systems, imported between 2021 and 2023 before China-wide restrictions took effect in January 2024, and 6 NXT:2100i systems imported before the ban
• Over $13 billion spent on NXT:1980i systems over 2024–2025 alone, with roughly 90 of the tools imported during that two-year window
• China’s share of ASML’s DUVi shipments surged from 26% in 2022 to 60–70% across 2024Q1–2026Q1

The timing is not subtle. China first began importing NXT:1980i systems in the late 2010s, but imports “drastically accelerated” after the Dutch government announced export controls on more advanced DUVi systems in 2023 — a classic stockpile response to an announced-but-not-yet-effective rule.

Why a “not-quite-EUV” machine matters so much

Photolithography is the step in chipmaking where ultraviolet light projects circuit patterns onto a silicon wafer. The three most recent generations are dry DUV (in use since the 1990s), DUVi (introduced 2006), and EUV (commercialized 2019). Each generation prints finer features, and finer features are what make AI accelerators fast.

The strategic logic of the current controls is built around EUV denial: the Netherlands blocked ASML’s EUV shipments to China in 2019, and ASML remains the sole EUV supplier on Earth. But the CTS report’s core technical argument is that EUV is not actually the binding constraint on Chinese AI chips today. Competitive AI chips require advanced nodes — 7nm-class logic and modern HBM — and those nodes can be printed, at scale and with acceptable yields, on immersion DUV. Chips built on 28nm-and-above nodes achievable with dry DUV are simply not competitive for AI training or inference; the report’s Appendix A works through the arithmetic of trying to match a B100’s FLOPS and bandwidth with 28nm logic and HBM1, and the numbers collapse.

In other words: Washington has locked the front door (EUV) while the side door (DUVi) stays open — and Huawei’s Ascend 950 series is the proof that the side door works.

The 50x gap, and the decade it could survive

The most policy-relevant finding is the projection model in Appendix F. As of 2026, the authors estimate the US ecosystem holds roughly a 50x advantage over China in AI computing power production — in 2025 terms, US firms produced an estimated 3.67 million B300-equivalent chips versus 40,000–146,000 for Huawei.

Under a “DUVi imports continue” scenario — annual imports modestly above the 2024–2025 pace — China’s lithography-bounded AI chip production could “narrow or even erase” that gap within five to ten years. Under a China-wide DUVi ban starting 2027 (with spare parts and servicing controls), the report projects the US advantage holds through at least 2035.

The asymmetry hinges on China’s inability to build these machines itself. CTS judges that Chinese lithography makers remain “decades behind ASML,” unlikely to commercialize a DUVi capable of supporting 7nm logic and advanced HBM before the early-to-mid 2030s, and unlikely to reach EUV before the second half of the 2030s. Advanced lithography, the report argues, will be the last chipmaking tool China indigenizes — etch and deposition equipment are closing fast, but lithography is the hard one.

Three loopholes, one policy ask

The report identifies three distinct openings in the current control regime:

1. ASML’s NXT:1980i and NXT:1970i are licensed into China despite enabling Ascend production. The Netherlands requires a license for these exports, but public data shows a large number of shipments are approved.
2. Less advanced DUVi tools are entirely uncontrolled, even though they enable earlier generations of AI chips.
3. Dutch servicing rules preserve ASML’s ability to maintain China’s installed DUVi fleet — a stockpile only stays combat-effective if it keeps running.

The remedy the authors propose is a simple one: replace partial DUVi controls with a China-wide ban on all DUVi exports, plus spare-parts and servicing controls, enforced through existing executive-branch authorities. They add a pragmatic sweetener — dry DUV systems can still support China’s large legacy-chip industry, so the ban need not crater the mature-node market.

This is not a fringe position. The MATCH Act, introduced in the US Congress in April 2026, would legislate exactly this kind of DUV ban, and the AEI’s April report on “The Lithography Loophole” reached parallel conclusions. The SCMP’s coverage of the CTS report notes the findings are already “fueling US calls for a complete DUVi export ban to close remaining loopholes.”

The counterweights

Two complications are worth stating plainly. First, ASML’s China business is enormous — 60–70% of its DUVi shipments in recent quarters — so a ban would land directly on one of the West’s most strategic companies, and The Hague would have to sign off. Second, China reported in July that a Shanghai state-backed firm had begun mass-producing its own immersion DUV tool; CTS treats indigenous capability as a mid-2030s problem, but every year of continued imports also feeds the learning curve.

For the AI industry, the stakes reduce to a single question: does the West’s compute moat come from Nvidia’s design lead alone, or from the machines that print the chips? The CTS report’s answer is the latter — and its warning is that the moat is currently being sold, one NXT:1980i at a time.