The Stack That China Built

Theme & Background

This chapter explores the underlying logic of China's AI infrastructure investment opportunities. The report argues that the release of the DeepSeek R1 model demonstrates the innovative capacity of China's engineering culture under pressure, while Western markets have asked the wrong question—focusing on "who lost" rather than "what this means." The author points out that China's AI infrastructure is forming a new market domain that had previously been absent from its investment horizon.

Core Thesis

The author's core investment thesis is: Export restrictions and policy-driven domestic construction are creating structural demand for semiconductor equipment, cloud infrastructure, memory architecture, and data transmission technology, yet the market continues to misread these companies as cyclical manufacturers or substitute investment targets. The author believes that China's AI buildout cycle will be longer, larger, and more equipment-intensive than most investors anticipate.

Counter-Intuitive Judgment:

• The market views DeepSeek as bearish (disrupting existing AI giants), but the author sees it as bullish—proof that China's engineering culture can innovate under constraints
• The energy of China's tech sector has shifted from consumer internet (e-commerce, social, payments) to hardware manufacturing and process engineering, yet most international investors have yet to follow

Key Arguments & Data

The author builds the thesis through five key signals ("Five Moments"):

Signal	Time	Event	Author's Interpretation
1	January 2025	DeepSeek releases R1 model	Proves innovative capacity of China's engineering culture under pressure
2	Late 2025	Huawei captures AI chip market share from NVIDIA; Moore Threads, MetaX, Biren Technology, Iluvatar CoreX apply for listings in Hong Kong and Shanghai	Rise of domestic GPU design companies; export restrictions foster local alternatives
3	Early 2026	China approves IPOs for Zhipu AI (Z.ai) and MiniMax	Capital markets provide clear long-term direction; Alibaba's Qwen and ByteDance continue to roll out leading AI models
4	Ongoing	Montage Technology's Hong Kong shares trade at ~30% premium over A-shares	Historical pattern reversal; international investor confidence in China's semiconductor leaders strengthens
5	May 2026	CXMT and YMTC file for listing	Memory is a structural element of AI—no memory means no models, no data centers, no AI autonomy

Policy Support Data:

• China's 15th Five-Year Plan prioritizes advanced manufacturing and AI infrastructure as core objectives through 2030
• A 20-year, $138 billion dedicated AI and robotics fund
• A $295 billion nationwide interconnected data center network plan, requiring at least 80% of technology from domestic suppliers

User Base: China already has 602 million generative AI users, all served by domestic platforms (as overseas alternatives are unavailable)

Companies/Assets Covered

Company/Asset	Role	Key Data	Bullish/Bearish
DeepSeek (under High-Flyer)	Trigger signal	Founded in 2023, no hardware advantage, no legacy infrastructure	Bullish (proves innovation capacity)
Huawei	AI chip market share growth	Captures AI chip market share from NVIDIA in China	Bullish (domestic alternative leader)
Moore Threads, MetaX, Biren Technology, Iluvatar CoreX	Domestic GPU design companies	Apply for listings in Hong Kong and Shanghai	Bullish (beneficiaries of export restrictions)
Zhipu AI (Z.ai), MiniMax	AI labs	Approved for IPO	Bullish (capital market endorsement)
Alibaba (Qwen), ByteDance	AI model development	Continuously roll out leading models	Bullish (industry acceleration)
Montage Technology	Semiconductors	Hong Kong shares trade at ~30% premium over A-shares	Bullish (international investor confidence)
CXMT, YMTC	Memory production	File for listing in May 2026	Bullish (structural AI element)

Investment Implications

The author has already established early positions across the infrastructure stack. For investors, the key takeaways are:

1. Abandon the consumer internet mindset: The center of gravity in China tech investing has shifted from "digital funnel optimization" to "manufacturing and process engineering that solves physical problems"; investors must reposition their analytical frameworks

2. Focus on policy certainty: The $138 billion AI fund and $295 billion data center plan provide early demand certainty rarely seen in other markets

3. Memory is a structural opportunity: The listings of CXMT and YMTC mark the capitalization of "non-peripheral" segments in AI infrastructure; without local memory supply, AI autonomy is unattainable

4. Hong Kong premium as a signal: Montage Technology's 30% premium in Hong Kong is a leading indicator of international capital's confidence in China's semiconductor sector, potentially spreading to other targets

---

Theme and Background

This chapter focuses on the structural opportunities in China's AI infrastructure buildout. The core logic is that U.S. export restrictions have not stifled China's chip production; instead, they have forced China to adopt a more equipment-intensive manufacturing path, thereby amplifying demand for semiconductor equipment, cloud infrastructure, and interconnect technology. Through field research and company analysis, the author argues this contrarian investment thesis that "constraints create markets."

Core Thesis

• Export restrictions are a structural positive: The restrictions force China to use older DUV lithography equipment, compensating with multipatterning techniques. This results in far higher semiconductor manufacturing equipment consumption per unit of AI compute capacity than in the West. Thus, the restrictions have actually expanded the scale of infrastructure supporting China's chip production.
• China's AI infrastructure investment cycle is longer and larger than the market expects: Overlooked areas such as power, cloud platforms, and interconnect technology will become future bottlenecks and investment priorities.
• The market's pricing of Alibaba is severely misaligned: Investors still view it as an e-commerce company, but the author believes Alibaba is the core owner of China's AI infrastructure. Its cloud platform, self-developed chips, and AI model (Qwen) will define value over the next decade.

Key Arguments and Data

1. Supply Gap:

• The U.S. and its allies will consume approximately 12 million AI chips this year.
• China's domestic production is about 2 million chips, while domestic demand may already exceed 5 million.
• This gap will not narrow in the short term, injecting urgency into China's domestic technology buildout.

2. Structural Amplification of Equipment Demand:

• ASML cannot sell the most advanced DUV or any EUV systems to China.
• Chinese fabs use older DUV equipment, compensating with multipatterning (adding more etching, deposition, and cleaning steps) to achieve comparable transistor density.
• Result: To produce equivalent AI computing power, China consumes more semiconductor manufacturing equipment.

3. Key Company Data:

• AMEC (Zhongwei Company): Founded by Gerald Yin (Yin Zhiyao), who previously developed etching platforms at Lam Research. Its most advanced tools target NAND architectures and 3D DRAM applications (a critical path for high-bandwidth memory) that Chinese memory makers are expanding. Over the past five years, Chinese fabs have adopted AMEC's tools on a scale previously unattainable for the company.
• NAURA Technology (North Huachuang): China's largest semiconductor equipment manufacturer, covering multiple process steps including etching, deposition, cleaning, and thermal processing. The author argues that its breadth (rather than single-depth expertise) is crucial in the current strategic position, as rapidly expanding fabs seek qualified local suppliers for every step.
• Alibaba: Operates China's largest cloud infrastructure platform (by revenue), develops self-designed chips, and invests in large language models and enterprise AI services. Its Qwen model family (both open-source and closed-source) is gaining recognition from Western companies.
• Montage Technology (Lanqi Technology): Holds the world's largest market share in memory interface chips (approximately one-third), operating in an oligopolistic industry. Its products sit on the critical path between AI server processors and memory, with high and rising switching costs.

4. Power Infrastructure Comparison:

• China operates the world's largest power grid, generating more than twice the electricity of the U.S.
• If domestic semiconductor production narrows part of the supply gap, China's vast power infrastructure can support data center construction at a pace exceeding external model forecasts.

Companies/Assets Covered

Company	Role	Key Data	View
AMEC (Zhongwei Company)	Semiconductor etching equipment manufacturer	Most advanced tools target NAND and 3D DRAM; achieved large-scale production validation over the past five years	Bullish: Benefits from structural amplification of China's equipment demand
NAURA Technology (North Huachuang)	China's largest semiconductor equipment manufacturer (multi-process coverage)	Covers etching, deposition, cleaning, thermal processing; breadth becomes a strategic advantage	Bullish: Multi-process embedded with customers, benefiting from the next wave of fab expansion
Alibaba	Cloud infrastructure, self-designed chips, AI model (Qwen)	China's largest cloud platform; Qwen adopted by Western companies	Bullish: Market undervalues its position as the core owner of AI infrastructure
Montage Technology (Lanqi Technology)	Global leader in memory interface chips	Approximately one-third global market share; oligopoly; high switching costs	Bullish: Interconnect technology is the most overlooked yet most durable value creation segment in AI infrastructure

Investment Implications

• Semiconductor Equipment: Focus on AMEC and NAURA, which directly benefit from increased equipment consumption by Chinese fabs to compensate for the lack of EUV. AMEC's positioning in 3D DRAM and NAND makes it key to the localization path for high-bandwidth memory (HBM).
• Cloud Infrastructure: Alibaba's valuation should be repriced from e-commerce to an AI infrastructure platform. Its cloud business, self-designed chips, and Qwen model form the core infrastructure for China's AI deployment, a shift the market has yet to fully reflect.
• Interconnect Technology: Montage Technology's oligopolistic position in memory interface chips and high switching costs make it the most defensive investment in the "pipeline" of AI servers. As system scale expands, memory bandwidth and interconnect efficiency will become new bottlenecks.
• Power and Data Centers: The scale advantage of China's power grid is an unpriced long-term catalyst. If the chip supply gap narrows, power infrastructure will support data center construction growth exceeding expectations.

---

Theme and Background

This chapter focuses on the third pillar of AI infrastructure—interconnectivity—and provides an in-depth analysis of Montage Technology's strategic positioning in this domain. The report argues that as AI training and inference workloads expand, signal integrity, latency, bandwidth, and power consumption have become system bottlenecks, with the interconnect layer emerging as a constraint on AI system performance. Additionally, by contrasting the market's general perception of four Chinese AI infrastructure companies with the author's actual observations, this chapter reveals undervalued structural opportunities.

Core Views

• Montage Technology is not a narrow-domain component supplier but a core bottleneck solver in the AI system interconnect layer. The market views it as a leader in memory interface chips, but the report argues that it is expanding into PCIe, optical, and Ethernet interconnects, with its value increasing as AI system scale grows.
• The Chinese AI infrastructure investment story is not just about "substitution" but also about the formation of "moats." The competitive advantages of AMEC and NAURA stem from engineering accumulation and customer dependency, rather than mere policy-driven substitution.
• Alibaba's long-term value is severely misread by the market. The market discounts it due to e-commerce and regulatory risks but underestimates the compounding effect of Alibaba Cloud as an infrastructure platform—its data relationships, enterprise integration, and AI model capabilities are building switching costs that are difficult to replace.

Key Arguments and Data

• Montage Technology: The company is not only a JEDEC board member with deep ecosystem ties to Samsung, SK Hynix, and Micron but is also expanding from memory interfaces to PCIe, optical, and Ethernet interconnects. The report argues that memory bandwidth issues will not disappear but will intensify with each new model generation, context length expansion, and increased inference complexity. The current valuation does not reflect the potential multiples under an AI-scale scenario five years from now.
• AMEC: Founder Yin Zhiyao's goal is not to fill gaps but to manufacture world-class etching tools. Every wafer run accumulates process knowledge, making the next tool better—this learning curve cannot be purchased or reverse-engineered by new entrants and deepens as the installed base expands.
• NAURA: Its value lies not in single-process optimization but in the breadth of certifications across multiple processes such as etching, deposition, cleaning, and thermal treatment. Once a fab embeds its multi-step tools, switching costs increase with process complexity.
• Alibaba: Alibaba Cloud sells not just computing power but also accumulated data relationships, enterprise integration, developer tools, and AI model capabilities (e.g., Qwen). Qwen is not just a model but a distribution mechanism, extending Alibaba Cloud's infrastructure reach to organizations that may not have been cloud customers a few years ago. The report expects that after this cycle, Alibaba will resemble more of an "indispensable infrastructure layer" for China's digital economy than an e-commerce company.

Companies/Assets Involved

Company	Role	Key Data/Arguments	Bullish/Bearish
Montage Technology	Core supplier of AI interconnect layer	Expanding from memory interfaces to PCIe/optical/Ethernet; JEDEC board member; deep collaboration with Samsung, SK Hynix, Micron	Bullish: The interconnect layer is becoming a binding constraint on AI system performance, and the company occupies a key bottleneck position
AMEC	Semiconductor etching equipment manufacturer	Each wafer run accumulates process knowledge, forming an irreplicable learning curve; not a substitute but a world-class tool	Bullish: A moat is forming, and competitive position deepens over time
NAURA	Multi-process semiconductor equipment supplier	Breadth of certifications across etching/deposition/cleaning/thermal treatment; multi-step switching costs rise with process complexity	Bullish: Breadth brings pricing power and customer retention; the market misreads it as a "generalist"
Alibaba	Cloud infrastructure + AI model distribution	Alibaba Cloud's accumulated data relationships, enterprise integration, developer tools, and Qwen model; market discounts due to e-commerce and regulation	Bullish: The market is still pricing an old narrative, underestimating the compounding effect of cloud infrastructure

Investment Implications

• Focus on the interconnect layer: Montage Technology represents the "quiet but value-accumulating" part of AI infrastructure, where memory bandwidth issues intensify with model scale. The current valuation does not reflect long-term potential. Investors should monitor its expansion progress in PCIe, optical, and Ethernet interconnects.
• Go beyond the "substitution" narrative: The moats of AMEC and NAURA come from engineering accumulation and customer dependency, not policy protection. Investors should assess the pace of their process knowledge accumulation and customer switching costs, rather than just policy continuity.
• Re-price Alibaba: The market's discount on Alibaba for e-commerce and regulatory risks may be excessive. The compounding effect of Alibaba Cloud as an infrastructure platform (data relationships, enterprise integration, AI model distribution) is building switching costs that are hard to replace. Investors should focus on its transition from e-commerce to an infrastructure layer.
• Incorporate geopolitical risks into position management: Chinese AI infrastructure is already viewed as a strategic asset, and the policy environment may affect position size and execution difficulty. Investors need to assess which companies are closer to national priorities and which may face constraints.

---

Theme and Background

This chapter focuses on a "factory floor" perspective of China's AI infrastructure buildout. The author argues that the market generally interprets China's AI investment opportunities through a simplistic "substitution narrative" (i.e., export restrictions driving domestic substitution), but this view is too narrow. The real opportunity lies in the independent and controllable full-stack infrastructure that China is constructing, whose scale, duration, and ultimate value far exceed the substitution logic.

Core Thesis

The author's central judgment is: The investment opportunity in China's AI infrastructure buildout is not driven by "what is being restricted," but by "what is being built." This is a longer-cycle, larger-scale, and more structurally profound story. The market has yet to properly price the strategic value and competitive moats of these infrastructure companies.

Contrarian Judgment:

• The market focuses on short-term comparisons of model performance between China and the US, but the author believes the more critical question is: Can China build the infrastructure required to support large-scale AI deployment without relying on an uncontrollable supply chain?
• The author believes that the core driver for investing in these companies is not geopolitical narratives or policy tailwinds, but the quality of the companies themselves, the depth of the moats they are building, and the significant opportunity for long-term investors who understand them early.

Key Arguments and Data

1. Continuity of Engineering Culture: The engineering culture exemplified by the R1 model is not an isolated event but is reflected across the entire industry chain:

• AMEC engineers' slow iteration on tool reliability.
• NAURA's ambitious platform strategy.
• Montage Technology's founder solving a problem the market had not yet named through over a decade of technological iteration.
• Alibaba's strategic transformation from a consumer platform to cloud and model infrastructure.

2. Lessons from Historical Precedents: The author cites the astonishing returns from early investments in Chinese tech giants, suggesting that current infrastructure investments may have similar potential:

• SoftBank's $20 million investment in Alibaba in 2000 was worth approximately $58 billion at the time of its 2014 IPO.
• Naspers' $32 million investment in Tencent in 2001 was worth $175 billion at the time of a partial divestment in 2018.
• Sequoia Capital's $60 million investment in WhatsApp returned approximately $3 billion (50x) when it was acquired by Facebook in 2014.

3. Policy and Market Signals:

• China plans to invest 1 trillion RMB in robotics and high-tech industries (IFR, March 2025).
• China is preparing a $295 billion plan to fund nationwide AI construction (Bloomberg, June 2026).
• China has 515 million generative AI users, leading the world.
• NVIDIA's market share in China's AI chip market has fallen to below 60%, with domestic chipmakers delivering 1.65 million AI GPUs.
• China mandates that chipmakers use 50% domestic equipment (Reuters, December 2025).
• Alibaba's cloud business grew 26% year-over-year in Q3 2025, with AI-related revenue achieving triple-digit growth for the eighth consecutive quarter, accounting for approximately 20% of external cloud revenue. Models derived from the Qwen model exceed 180,000, more than double the second-largest provider. China's AI cloud market share stands at 35.8% (Omdia).

Companies/Assets Covered

Company	Role/Key Data	Bullish/Bearish
AMEC (中微公司)	Semiconductor etching equipment manufacturer; engineering culture reflected in iterative tool reliability	Bullish
NAURA (北方华创)	Semiconductor equipment platform company with platform strategy ambitions	Bullish
Montage Technology (澜起科技)	Memory interface chip design company; founder solved a problem through over a decade of technological iteration	Bullish
Alibaba (阿里巴巴)	Transitioning from consumer platform to cloud and model infrastructure; 35.8% AI cloud market share; triple-digit AI revenue growth	Bullish
NVIDIA	Market share in China's AI chip market fell to below 60%	Bearish (relative)

Investment Implications

Investors should strategically overweight the full stack of China's AI infrastructure, including semiconductor equipment, cloud infrastructure, memory architecture, and data transmission technology. The core logic is not short-term substitution but long-term structural construction. Specific directions:

• Semiconductor Equipment: Companies like AMEC and NAURA will benefit from mandatory increases in domestic content rates and continuous technological iteration.
• Cloud Infrastructure: Cloud providers like Alibaba will benefit from the explosion in enterprise AI demand, with their AI cloud businesses showing strong growth and leading market share.
• Memory and Interconnect: Companies like Montage Technology, with their long-term technological accumulation in memory interfaces and data transmission, will become critical in AI data center construction.
• Caution: Avoid short-term trading based solely on the "substitution narrative"; focus on the companies' own moats and long-term strategic value.