China's Chip Independence: SMIC, Huawei Ascend, and What '7nm' Actually Means

On August 29, 2023, Huawei released the Mate 60 Pro in China without a press conference, without a keynote, without even a formal announcement — it simply appeared on Huawei’s website and in stores. The phone sold out within hours. The reason for the quiet launch was immediately obvious to anyone who opened one: the Kirin 9000s chip inside was manufactured on SMIC’s N+2 process, achieving transistor density equivalent to TSMC’s 7nm node — a capability that U.S. government analysts had assessed China would not achieve until at least 2027.

The implications were debated intensely, sometimes hysterically, and mostly without the technical grounding necessary to understand what the achievement actually demonstrated. The Mate 60 Pro launch was simultaneously a genuine engineering triumph and a bounded achievement that does not represent the threat some commentators described. Understanding the distinction requires understanding what “7nm” means, what SMIC’s actual capabilities are, and what the specific constraints on Chinese advanced semiconductor manufacturing remain despite the breakthrough.

What “7nm” Actually Means in 2023

The first thing to understand about the Mate 60 Pro’s chip is that “7nm equivalent” is a description of transistor density, not a description of SMIC’s manufacturing process. SMIC achieved this density using deep ultraviolet (DUV) lithography — the previous generation of lithography technology — rather than the extreme ultraviolet (EUV) lithography that TSMC uses for its N3 process and that produces the same transistor density at higher yield and lower manufacturing cost.

DUV lithography reaching 7nm-equivalent density requires multiple patterning — exposing the same layer of silicon multiple times, with different masks, to achieve finer features than a single DUV exposure can produce. This is technically possible but computationally and operationally complex. Each additional patterning step adds cost, increases process time, and introduces new opportunities for defects. TSMC moved from multi-patterning DUV to EUV precisely because EUV achieves the same density in a single exposure, with dramatically better yield and lower cost.

Related reading

The Kirin 9000s chip manufactured on SMIC N+2 reportedly has a yield (the fraction of manufactured chips that meet specifications) of approximately 30-40%, compared to TSMC’s 80%+ yields on equivalent processes. A chip manufacturer with 35% yield is discarding two-thirds of every silicon wafer produced. The cost per functional chip is therefore roughly 2-3x what equivalent TSMC production would cost — a substantial premium that limits the economics to scenarios where the geopolitical imperative overrides the cost calculation.

The Export Control Logic and Its Limits

The U.S. export controls implemented in October 2022 and significantly expanded in October 2023 were designed around a specific theory: that advanced AI chips and advanced semiconductor manufacturing equipment are so tightly linked that controlling equipment exports would prevent China from building the computing infrastructure needed for advanced AI development.

The theory has partial validity. The specific EUV lithography equipment needed to manufacture chips at TSMC N3 equivalent or better is made exclusively by ASML, a Dutch company, and ASML has complied with Dutch and U.S. government requests not to ship EUV equipment to China. Without EUV, manufacturing leading-edge chips (below 7nm in effective transistor density) is extremely difficult and expensive with current known methods. The Mate 60 Pro demonstrated that 7nm-equivalent chips are possible with DUV; chips below 5nm effective density (where the efficiency gains from advanced nodes become most significant for AI applications) appear to remain beyond DUV’s practical capability.

The specific chips targeted by export controls — Nvidia’s A100 and H100, AMD’s MI250 and MI300 — represent the current generation of AI training hardware. Huawei’s Ascend 910B, manufactured on SMIC’s N+2 process, is a genuine alternative for AI training tasks, though with lower energy efficiency than the controlled chips. The 2023 Bernstein Research analysis estimated the Ascend 910B delivers approximately 60% of the H100’s performance at roughly 2x the power draw — inferior, but not useless.

Chinese AI labs including Baidu, ByteDance, Alibaba, and others have been building AI infrastructure on Ascend 910B clusters since the export controls made H100 procurement impossible through legitimate channels. The consensus from industry sources is that Chinese frontier AI development has been slowed but not stopped by the controls — the additional cost and reduced efficiency of domestic chips creates a capability gap relative to U.S. labs that is real but not insurmountable.

The YMTC Memory Achievement

The discussion of Chinese semiconductor independence focuses primarily on logic chips — the CPUs, GPUs, and AI accelerators that dominate the AI hardware conversation. The equally important and somewhat less-covered story is in memory chips, specifically NAND flash storage.

YMTC (Yangtze Memory Technologies Corporation) was founded in 2016 as a state-sponsored effort to create domestic NAND flash manufacturing capability. By 2022, YMTC had achieved 128-layer 3D NAND — competitive with Samsung, SK Hynix, and Micron’s leading products — and was beginning to sample 232-layer NAND with performance comparable to Western products.

Related reading

The U.S. added YMTC to the Entity List in December 2022, restricting U.S. companies from supplying equipment and software to the company. This slowed YMTC’s technology roadmap — U.S.-origin EDA software is critical for chip design, and U.S. equipment is critical for certain manufacturing steps — but did not stop production of existing products. Apple evaluated YMTC chips for iPhone NAND flash in 2022 and decided against them following U.S. government pressure, depriving YMTC of what would have been a transformational design win.

The NAND situation illustrates a pattern that repeats throughout the semiconductor geopolitics: U.S. export controls impose costs and slow Chinese development, but rarely create absolute barriers. The timeline stretches by years, not decades.

The Semiconductor Equipment Counter-Campaign

China’s response to U.S. equipment export controls has included domestic investment in semiconductor equipment manufacturing that has surprised Western analysts with its speed. NAURA Technology Group and Advanced Micro-Fabrication Equipment (AMEC) are the two most notable Chinese equipment makers, and both have made meaningful progress in the post-2022 period.

AMEC’s ICP (inductively coupled plasma) etching equipment, used for chip manufacturing, has been adopted by SMIC for its domestic production. The tools do not match the performance of Applied Materials or Lam Research equipment for the most advanced applications, but they are functional for current Chinese production nodes. NAURA has made progress in physical vapor deposition and atomic layer deposition equipment.

The one domain where domestic substitution has genuinely failed — where there is no credible Chinese alternative in development — is the photolithography space. EUV lithography requires extreme ultraviolet light sources based on tin plasma, collecting optics with surface accuracy measured in angstroms, and integration engineering that took ASML twenty years and over $1 billion in annual R&D to develop. SMEE (Shanghai Micro Electronics Equipment) has announced plans for DUV equipment, but its most advanced confirmed product remains at 90nm resolution — roughly two generations behind ASML’s most advanced DUV equipment.

The lithography constraint is real and will remain constraining for at least five years. The question is what happens at year seven or ten.

Huawei’s Kirin Comeback and What It Signals

The Kirin 9000s’s commercial success — Huawei sold approximately 40 million Mate 60 Pro and related devices using the chip in 2024, recovering significant smartphone market share in China — has validated SMIC’s investment and emboldened further investment in domestic chip capability.

More significantly, Huawei’s Ascend AI chip series (910, 910B, 910C) demonstrates that China has genuine ambition and partial capability to develop AI training infrastructure independent of Nvidia. The Ascend 910C, expected in volume production in 2025, reportedly achieves performance comparable to the H100 through architectural improvements that partially compensate for the process node disadvantage. If this claim proves true in production, it represents a meaningful narrowing of the capability gap.

Related reading

The trajectory suggests that the U.S. export control strategy will succeed in maintaining a technology gap — Chinese AI hardware will remain a generation or two behind the frontier for as long as the EUV bottleneck holds — but will not prevent China from developing genuine AI capabilities at scale. The gap may be enough to matter for the most advanced applications; it is clearly not enough to prevent commercially significant AI development.

This is what “7nm” actually means in 2023: not parity, but persistence. China cannot match TSMC’s most advanced processes. It can manufacture chips advanced enough for most commercially relevant AI applications, at higher cost and lower yield, on a timeline that U.S. export controls have compressed but not eliminated.

The export control bet is not lost. It is also not won.