China SMIC 7nm Chips: How DUV Beats the EUV Ban — AEI Report 2026

The American Enterprise Institute published a report in April 2026 — authored by Ryan Fedasiuk and Julia Torres — that identifies a fundamental gap in US semiconductor export controls. China's SMIC and its primary customer Huawei are repurposing older-generation ASML NXT:1980Fi-series deep ultraviolet immersion (DUVi) lithography machines via multi-patterning to produce near-frontier AI chips. SMIC has demonstrated 7nm production. It is developing 5nm capability with yields currently at approximately 20%. Huawei's target: 1.6 million high-end logic dies in 2026 for AI accelerators. China acquired approximately 90 ArFi machines in 2024 worth $5-7 billion — all before any restrictions on DUVi tools.

The EUV export ban, widely cited as the cornerstone of US semiconductor containment strategy, has a loophole. China does not need EUV to make near-frontier chips if it has enough DUVi machines and willingness to use multi-patterning — a slower, less yield-efficient, but technically viable alternative route.

What DUVi Multi-Patterning Actually Is

Extreme ultraviolet (EUV) lithography uses 13.5nm wavelength light to print chip features in a single pass. It is the technology ASML makes exclusively and which China cannot legally obtain. Deep ultraviolet immersion (DUVi) lithography uses 193nm wavelength light — an older technology that has been in widespread use since the mid-2000s.

The standard assumption behind export controls is that DUVi cannot produce sub-10nm chips because the light wavelength is too long to print features that small in a single exposure. This assumption is correct for single-patterning. It is wrong for multi-patterning.

Multi-patterning means exposing the same layer of silicon multiple times with slightly offset patterns, then combining the result to achieve features smaller than the light wavelength would normally allow. TSMC and Samsung used multi-patterning to push DUVi to 7nm before EUV was production-ready. The process is slower (3-4x more passes than EUV), yields are lower, and cost per die is higher. But it works.

SMIC is using exactly this technique. The AEI report documents that SMIC's 7nm process (demonstrated in the Huawei Kirin 9000s chip in 2023) used DUVi multi-patterning on ASML NXT:1980Fi machines. The 5nm development work currently at 20% yield is the same approach pushed one node further.

The Machine Fleet: 90 ArFi Machines Worth $5-7 Billion

China acquired approximately 90 advanced ArFi (argon fluoride immersion) DUVi tools in 2024, valued at $5-7 billion. The NXT:1980Fi is ASML's most advanced DUVi machine — not covered by the EUV export ban because it is not EUV equipment. The machines were legally purchased and exported before any restrictions on DUVi tools were implemented.

ASML's total fleet of NXT:1980Fi machines globally is approximately 300-400. China's 90 machines represent roughly 20-25% of the global installed base of the most capable DUVi tool available. That concentration — combined with SMIC's workforce of experienced process engineers who developed multi-patterning techniques over the past decade — creates a meaningful near-frontier production capability.

The 1.6 million high-end logic die target for 2026 is Huawei's production goal for its AI accelerator business. At 20% yield on a 5nm process, achieving 1.6 million usable dies requires printing approximately 8 million dies. That is a large production volume but within the operational capacity of SMIC's DUVi fleet running multi-shift operations.

How Near-Frontier Is Near-Frontier?

The Huawei chips coming out of SMIC's DUVi lines are not competitive with TSMC's N3 (3nm) or N2 (2nm) processes. They are roughly competitive with TSMC N7 (2019 vintage) and approaching N5 (2020). For comparison:

• Nvidia H100 uses TSMC N4 (4nm class, 2022)
• Nvidia B200 uses TSMC N3P (3nm, 2024)
• SMIC current production: 7nm (2023), developing 5nm (targeting 2026-2027)

A SMIC 7nm Huawei AI accelerator chip is approximately 3-4 generations behind current Nvidia in process node. However, at the AI inference workload, process node is not the only variable. Memory bandwidth, interconnect efficiency, and software stack matter enormously. Huawei's Ascend 910B and its successors are designed specifically for inference workloads on Huawei's MindSpore framework, and Chinese state and enterprise customers running those workloads do not need Nvidia H100 performance parity — they need sufficient capability for domestic AI deployment.

The AEI report concludes China's DUVi fleet "will almost certainly" be capable of producing enough near-frontier dies in 2026 to supply Huawei's domestic AI accelerator demand, including for Chinese cloud providers (Alibaba, Baidu, ByteDance) that face US chip restrictions.

Why Export Controls Have Not Closed This Gap

The current export control framework, administered by the US Bureau of Industry and Security (BIS), restricts:

• EUV lithography tools (complete ban, no exceptions)
• Certain advanced DUV tools below a specific numerical aperture and overlay accuracy threshold
• Specific Nvidia AI chip models (A100, H100, A800, H800, and successors above certain thresholds)
• Memory chips above specified capacity and bandwidth thresholds

The NXT:1980Fi is not caught by the current DUV controls. Its specifications fall in the grey zone between restricted and permitted — it is more capable than the restricted threshold but not EUV. The AEI report explicitly identifies this as the loophole: the controls are drawn at specific technical thresholds that ASML's most capable DUVi tool sits just outside.

The servicing problem compounds this. Approximately 200+ DUVi tools are already in China, installed and operational. Even if new exports are restricted, existing machines continue to operate as long as ASML and its service partners can maintain them. The AEI report recommends restrictions on servicing and maintaining DUVi tools in China as a critical second-order control — without maintenance, high-utilisation multi-patterning processes degrade rapidly as optics drift out of calibration.

Congressional Response

A bill was introduced in the House in April 2026 to ban the sale of key AI chipmaking equipment to China with broader coverage than current BIS rules. The bill — targeting aperture size and overlay accuracy rather than specific product names — would close the capability-based gap the AEI report identifies. Whether it passes and how quickly BIS implements revised rules is the operative policy question.

The policy timeline matters for chip companies and AI hardware teams: if capability-based controls are implemented in 2026, the Chinese AI chip supply landscape changes significantly. If the legislative path stalls into 2027, SMIC completes 5nm yield ramp and Huawei has a domestic AI accelerator generation competitive with 2022-era Nvidia hardware.

Developer and AI Hardware Implications

Fabless chip design: Any chip design house considering SMIC as a manufacturing partner for AI accelerator products faces increasing US compliance overhead. SMIC is on the Entity List; working with SMIC requires US Commerce Department licences for US-origin technology. The DUVi report makes clear that SMIC's capability is higher than assumed — raising the urgency of existing controls on US-origin EDA tools and IP used in designs headed to SMIC.

Chinese AI cloud costs: Alibaba Cloud, Huawei Cloud, and Baidu AI Platform customers in China will have access to domestically produced AI inference hardware in 2026. The Huawei Ascend accelerator stack running on SMIC 7nm is not H100-class performance, but it is sufficient for inference deployment on Chinese state-approved LLMs. Pricing will reflect domestic supply rather than import pricing.

Export control compliance: Companies selling AI software, developer tools, or cloud infrastructure to Chinese enterprise customers should model the scenario where domestic Chinese AI hardware becomes sufficient for their customers' workloads. A customer that can run your software stack on Ascend accelerators without importing Nvidia hardware has a different compliance risk profile than one dependent on controlled exports.

Key Takeaways

• AEI "Lithography Loophole" report April 2026: authors Ryan Fedasiuk and Julia Torres; SMIC and Huawei using ASML NXT:1980Fi DUVi multi-patterning to produce near-frontier chips despite EUV ban
• SMIC capability: 7nm production demonstrated (Kirin 9000s); 5nm in development at ~20% yield; 1.6 million high-end dies targeted in 2026 for Huawei AI accelerators
• Machine fleet: ~90 ArFi machines acquired by China in 2024 for $5-7B, representing ~20-25% of the global NXT:1980Fi installed base; all legally purchased before DUVi restrictions
• The gap: current BIS controls are drawn at technical thresholds that NXT:1980Fi sits just outside; servicing restrictions on existing machines are the recommended second-order control
• Congressional response: House bill introduced targeting capability-based controls (aperture, overlay accuracy) rather than product names; timeline determines when SMIC 5nm ramp completes vs when controls close
• Practical implication: Chinese domestic AI inference hardware in 2026 will be competitive with 2022-era Nvidia — sufficient for domestic AI deployment, not competitive with current H200/B200

For the Terafab Intel 14A US chip independence angle, read Elon Musk Terafab Intel 14A Chips. For SK Hynix HBM supply context relevant to the broader chip stack, read SK Hynix Record Earnings: HBM Shortage Until 2030. For the NEO Semiconductor 3D X-DRAM alternative memory angle, read NEO Semiconductor 3D X-DRAM vs HBM.