I. Executive Summary (TL;DR)

This report provides a comprehensive, in-depth analysis of Cambricon Technologies Corporation Limited (hereinafter “Cambricon”). Our central thesis is that Cambricon’s recent ascent in the capital markets is not merely a product of its technical prowess but the result of a unique confluence of forces: first, the global surge in computing power demand ignited by Generative AI (AIGC); second, the powerful tailwind of China’s “domestic substitution” policy, accelerated by U.S. sanctions; and finally, the commercial maturation of its core product line (the Siyuan series) at the precise moment of peak market demand. These factors have collectively propelled Cambricon to an unprecedented strategic high ground.

From a financial perspective, Cambricon executed a dramatic turnaround in the first half of 2025, shifting from years of substantial losses driven by massive R&D investment to significant profitability.¹ This not only validates its long-term strategy but also signals the company’s entry into a new phase of development. However, we must remain cautious about the sustainability of this explosive growth and the company’s long-standing risk of high customer concentration.⁵

Strategically, Cambricon has emerged as a frontrunner in China’s race to find alternatives to Nvidia. Its core advantages lie in its focused ASIC architecture, a unified software ecosystem (NeuWare), and its “national team” pedigree originating from the Chinese Academy of Sciences.⁹ These factors give it a favorable position in securing key domestic clients and policy support.

Nevertheless, the risks are equally significant. The bearish view centers on several key points: first, the substantial gap in hardware performance and software ecosystem (CUDA) compared to industry giant Nvidia ¹; second, fierce competition from local behemoths like Huawei (Ascend) ³; and finally, the profound vulnerability of its fabless business model to geopolitical shocks and the risk of global supply chain disruptions.⁵

Looking ahead, Cambricon’s fate will largely depend on its ability to effectively leverage the current protected market window to transform a temporary market advantage into a genuine, durable technological and ecological moat. The next 24 to 36 months will be a critical period in determining whether it can achieve this leap and ultimately establish itself as a core pillar of China’s artificial intelligence infrastructure.

II. The Genesis of an AI Pioneer: From Academia to the Industrial Frontier

The “Genius Brothers” and the Legacy of the Chinese Academy of Sciences

Cambricon’s founding story is not a typical Silicon Valley garage startup narrative; it is deeply rooted in China’s national scientific research system, with its “national team” DNA evident from the very beginning. The company’s key figures are the brothers Chen Yunji and Chen Tianshi, both prodigies who graduated from the renowned “Special Class for the Gifted” at the University of Science and Technology of China (USTC).¹ Their paths were nearly identical: the elder brother, Chen Yunji, entered the USTC gifted class at 14, with his younger brother Tianshi following suit. Both later pursued advanced studies at the Institute of Computing Technology (ICT) of the Chinese Academy of Sciences (CAS).¹

This period at the ICT was crucial. As the cradle of China’s computer science endeavors, it provided the brothers with a flexible research environment and ample support.¹⁸ They participated in the state-led “Loongson” CPU project, where Chen Yunji studied under Dr. Hu Weiwu, a top expert in China’s chip sector, and became one of the youngest members of the Loongson team, eventually serving as the chief architect for the “Loongson-3”.¹ This experience provided them with invaluable, front-line experience in China’s indigenous processor development.

It was at the ICT that the brothers’ research interests began to converge. Chen Yunji focused on computer chips, while Chen Tianshi specialized in artificial intelligence algorithms.¹ About a decade ago, when AI and processor design were still considered separate fields, they presciently saw the immense potential in their combination.¹⁸ They recognized that general-purpose processors were inefficient and power-hungry when running deep learning algorithms. To enable machines to think and react like humans (for instance, in autonomous driving), specialized AI chips were necessary.¹⁸

This vision led to a series of groundbreaking academic achievements. Long before the company was founded, the research group led by the Chen brothers introduced the world’s first deep learning processor architecture—the “DianNao” (Electric Brain) series.¹⁸ This line of research, including DaDianNao (Big Electric Brain) and PuDianNao (Universal Electric Brain), won numerous awards at top international academic conferences, such as the Best Paper Award at ASPLOS 2014, laying a solid theoretical and technical foundation for Cambricon’s establishment.¹⁸ Notably, their insistence on naming the processors using Chinese pinyin drew attention from the international academic community, reflecting a unique cultural confidence.¹⁸

The birth of Cambricon is essentially a prime example of a top Chinese research institution commercializing strategic research outcomes. The “CAS” in the company’s full name, “CAS Cambricon Technologies Corporation Limited,” clearly indicates its deep ties to the Chinese Academy of Sciences.⁹ The company was co-founded by the founding team and CAS-ICT’s investment management platform, CAS-Source.²¹ This origin story dictates that Cambricon is not merely a market competitor but a technological force with high expectations at the national strategic level. It has access to government resources, a pipeline of top talent, and strong political backing, which provided an unparalleled advantage in its early development. However, this institutional background might also make its organizational structure and decision-making processes less flexible and agile compared to fully market-driven startups.

Key Milestones on the Path to Prominence

Cambricon’s journey from an academic research group to the first AI chip company listed on China’s stock market clearly outlines the trajectory of its strategic evolution.

2016: A Star is Born

The company was formally established in March 2016.1 Its name, “Cambricon,” is taken from the geological period known for the explosion of life, symbolizing the new technological revolution that artificial intelligence would usher in.1 In the same year, the company launched the world’s first commercial deep learning processor, the Cambricon-1A.20 The product was an instant sensation, earning a spot alongside Tesla’s Autopilot and IBM Watson as one of the “World’s Leading Internet Scientific and Technological Achievements” at the 3rd World Internet Conference, winning the fledgling company immense prestige.20

2017-2018: Symbiosis with Huawei

This period was a decisive phase in Cambricon’s history. At the time, global smartphone giant Huawei was seeking powerful AI processing capabilities for its Kirin chips. Cambricon’s 1A processor IP core perfectly met this need. The two companies quickly formed a partnership, and Huawei integrated Cambricon’s Neural Processing Unit (NPU) as the AI subunit in its flagship Kirin 970 and subsequent Kirin 980 chips.3 This collaboration was a win-win: Huawei became a pioneer in mobile AI chips, gaining a first-mover advantage, while Cambricon secured its first major revenue stream, achieved large-scale commercial validation, and established itself as a global leader in terminal AI IP.3 In 2017 and 2018, Huawei was Cambricon’s largest source of revenue.9

2019-2020: Strategic Pivot and STAR Market IPO

However, the honeymoon with Huawei did not last. As Huawei’s in-house R&D capabilities grew, it began developing its own “Da Vinci” architecture and Ascend series NPUs, gradually replacing Cambricon’s IP.9 By 2024, Cambricon was no longer in Huawei’s supply chain.20 Losing its largest customer was undoubtedly a severe, almost fatal, blow to Cambricon. This event forced the company into a profound strategic transformation.

In the long run, what seemed like a crisis catalyzed Cambricon’s evolution. It made the management realize the immense risks of over-relying on a single customer and a single business model (IP licensing). Consequently, Cambricon decisively shifted its strategic focus from terminal IP licensing to the more challenging but far more lucrative development and sale of its own cloud and edge intelligent chips and accelerator cards. This painful transition, while causing short-term financial strain, allowed the company to build a more complete and resilient product portfolio and directly led to the creation of the subsequent Siyuan series. It was this forced strategic adjustment that positioned Cambricon to seize the opportunities presented by the later AIGC wave. Had it not lost Huawei as a customer, Cambricon might have remained a niche IP provider, rather than becoming the significant player in the cloud computing market it is today.

During this critical period of strategic transformation, Cambricon successfully listed on the Shanghai Stock Exchange’s STAR Market on July 20, 2020, becoming China’s “first AI chip stock”.² The IPO not only provided ample funding for its subsequent high-intensity R&D but also further solidified its benchmark status in the domestic AI chip industry.

2023-2025: The AIGC Wave and Market Ascendancy

With the explosion of global AIGC technology since 2023, the demand for high-performance AI computing power has grown exponentially. Simultaneously, escalating U.S. export restrictions on chips to China prevented Chinese tech companies from obtaining top-tier AI chips from companies like Nvidia. The convergence of these two factors created a historic window of opportunity for Cambricon. The company’s cloud-based intelligent chip products had matured at just the right moment, becoming one of the few high-performance domestic alternatives available. The surge in demand drove astonishing growth in the company’s performance, and its stock price soared. In August 2025, Cambricon’s stock price briefly surpassed that of Kweichow Moutai, momentarily becoming the most expensive stock in China’s A-share market, with its market capitalization exceeding 600 billion RMB and founder Chen Tianshi’s net worth skyrocketing.1 This period marked Cambricon’s successful transformation from a research-oriented company with persistent losses into a market-recognized industry leader with strong profitability.

III. Deconstructing the Tech Stack: Architecture, Products, and Ecosystem

Core Architecture: The Siyuan Series and MLUarch

The cornerstone of Cambricon’s technology strategy is its focus on Application-Specific Integrated Circuit (ASIC) architecture. Unlike the more versatile General-Purpose Graphics Processing Unit (GPGPU) path dominated by Nvidia, Cambricon has chosen to design chips specifically for AI workloads, particularly neural network computations.³ The core advantage of the ASIC approach is its superior power efficiency (performance per watt), as it strips away all redundant functions unrelated to AI computing, allowing all transistor resources to be dedicated to optimizing core algorithms. This results in higher performance and lower power consumption for specific tasks.

The crystallization of this design philosophy is Cambricon’s proprietary Machine Learning Unit architecture, or MLUarch. MLUarch is the company’s core intellectual property and the underlying foundation for all its “Siyuan” series chips. The architecture has undergone continuous iteration and has now reached its fifth generation, MLUarch05, which is used in the latest cloud training chip, the Siyuan 590.¹⁰ Each generational evolution has brought significant improvements in performance, energy efficiency, and functionality.

In terms of implementation, Cambricon has also actively adopted cutting-edge industry innovations. For example, in its third-generation cloud chip, the Siyuan 370, the company was the first to use advanced Chiplet technology.²⁶ This technique allows two or more independently manufactured AI computing dies to be packaged into a single chip. This not only effectively improves the manufacturing yield of large-sized chips but also allows for the flexible creation of a diverse and cost-effective product line to meet the needs of different scenarios through various chiplet combinations.²⁶

Product Portfolio: A Unified Cloud-Edge-End Strategy

Cambricon’s product layout reflects its comprehensive “cloud, edge, and end” strategy, aiming to provide full-stack computing power support for AI applications, from data centers to edge devices.¹⁰

Cloud Product Line

Cloud products are the absolute core of Cambricon’s current strategy and its main source of revenue.³ With the rise of AIGC and large models, the cloud training and inference market has become a key battleground.

• Siyuan 270/290/370 Series: This series of products formed the bedrock that allowed Cambricon to establish a foothold in the cloud market. The Siyuan 270, launched in 2019, offered a peak processing power of 128 TOPS (INT8) and supported various mainstream deep learning models like CNN, RNN, and LSTM, marking the company’s official entry into the cloud sector.³ The subsequent Siyuan 290 and 370 models continuously improved on performance and features. The Siyuan 370, in particular, not only adopted Chiplet technology but was also the first in the industry to support LPDDR5 memory, which provided three times the memory bandwidth of the previous generation and was far more power-efficient than traditional GDDR6 memory.²⁶ Additionally, the Siyuan 370 was equipped with MLU-Link multi-chip interconnect technology, providing up to 200GB/s of direct communication bandwidth between chips, significantly enhancing the efficiency of multi-card parallel tasks.²⁶

• Siyuan 590: This is Cambricon’s current flagship product, designed specifically for large-scale AI training tasks and serving as the company’s “killer app” in the current AIGC wave.⁵ Although the company has not yet released a full technical white paper, the chip has achieved commercial breakthroughs and secured large-volume orders from leading internet companies.¹ According to market intelligence and analyst reports, the Siyuan 590 uses the latest MLUarch05 architecture, delivering more than double the performance of its predecessor. Its overall performance is benchmarked against Nvidia’s A100 chip (released in 2020), placing it at the forefront of domestic AI chips.¹

Edge Product Line

The edge product line is designed to provide moderate computing power for devices positioned between the end and the cloud, meeting the real-time processing needs of scenarios like intelligent security, smart manufacturing, and smart transportation.

• Siyuan 220: This is Cambricon’s main product for the edge market. It is a high-performance, low-power edge AI chip available in various form factors, such as an M.2 accelerator card.²⁸ Despite being only half the size of a credit card, it delivers up to 8 TOPS (INT8) of AI computing power at an extremely low power consumption of 8.25W.²⁸ The Siyuan 220 supports configurable 16/8/4-bit fixed-point operations and has powerful video encoding and decoding capabilities, making it an ideal choice for edge vision applications.²⁸ However, the company’s financial data shows a declining trend in revenue from the edge product line in recent years, indicating a strategic shift in resource allocation towards the more lucrative cloud market.⁵

The Software Moat: The Pursuit of a Unified Ecosystem with NeuWare

In the semiconductor industry, while hardware performance is important, it is a robust software ecosystem that truly locks in customers and builds a long-term competitive barrier. Nvidia’s CUDA platform has proven this over more than a decade, creating an almost insurmountable software “moat.” Cambricon understands this profoundly and has consciously emulated Nvidia’s strategy from the outset, dedicating itself to building its own unified software ecosystem.

The core of this ecosystem is Cambricon NeuWare.¹⁰ NeuWare is a unified, platform-level basic system software that spans all of Cambricon’s hardware products from the cloud to the edge, providing developers with a consistent programming experience. This software stack includes low-level drivers, runtime libraries, high-performance operator libraries, and a suite of development toolchains.¹⁰

Cambricon’s philosophy is “hardware-software co-optimization”.³ This means its hardware architecture (MLUarch) and software platform (NeuWare) are designed in parallel and deeply coupled to maximize the computational potential of the hardware. A key component of NeuWare is the

MagicMind inference acceleration engine. Based on advanced MLIR graph compilation technology, it can efficiently deploy models trained with mainstream frameworks like TensorFlow and PyTorch onto Cambricon’s full range of chips, achieving competitive performance.²⁶

However, building a successful software ecosystem faces the classic “chicken-and-egg” problem: developers will only create software for a hardware platform with a large user base, and users will only buy hardware that has rich software and developer support. It took Nvidia over a decade and billions of dollars to break this cycle.

This is where China’s unique market environment provides Cambricon with a once-in-a-lifetime opportunity. Due to U.S. export controls, China’s major tech companies are forced to purchase and deploy Cambricon’s hardware on a massive scale.¹ This externally driven mass deployment is, in effect, artificially “incubating” the first large user base and application scenarios for the NeuWare ecosystem. This allows Cambricon to kickstart the positive feedback loop of its ecosystem under unnatural market conditions. China’s top internet companies are now not just Cambricon’s customers but are effectively its paid “testers” and “co-builders.” In real-world deployments, they help identify and fix software bugs and improve the development toolchain. This geopolitical “greenhouse” is helping Cambricon close the huge ecosystem gap with CUDA at a speed that would be impossible in an open market. The key to its long-term success lies in whether it can convert this passively acquired starting point into a truly vibrant, self-evolving developer community.

IV. Financial Analysis: A Magnificent Turnaround

From Massive Losses to Explosive Growth

Cambricon’s financial journey clearly illustrates the typical “J-curve” development trajectory of a hard-tech company. During its initial public offering, the financial status disclosed in its prospectus stunned the market. From 2017 to 2019, although the company’s revenue grew rapidly, its net losses continued to widen, accumulating to over 1.6 billion RMB in three years.³ These huge losses were primarily driven by extremely high R&D investments. In the fiercely competitive chip design industry, sustained high-intensity R&D is a necessary cost to maintain technological leadership and develop complex product portfolios; Cambricon’s R&D expenses in those years often exceeded its annual revenue.³²

However, this all changed dramatically in the first half of 2025. The company’s semi-annual report was nothing short of “explosive”: it achieved operating revenue of 2.881 billion RMB, a staggering 4347.82% increase compared to 65 million RMB in the same period last year. More importantly, the net profit attributable to the parent company reached 1.038 billion RMB, a stark contrast to the 530 million RMB loss in the same period last year, marking the company’s first semi-annual profit since its listing.¹

This stunning financial reversal was the result of a confluence of factors. First, the global explosion in demand for AIGC and large model training created an unprecedented market for computing power.³ Second, U.S. export control policies opened up a huge, protected “domestic substitution” market for Chinese chip manufacturers, making Cambricon’s products a necessity for domestic tech giants.¹ Finally, and most fundamentally, the years of persistent high-intensity R&D paid off: the cloud product line, represented by the Siyuan 590, reached a level of maturity for large-scale commercial use and gained widespread market acceptance.³ The convergence of these three forces propelled Cambricon’s performance surge.

Revenue Quality and Customer Concentration

When analyzing its revenue structure, a prominent feature is the absolute dominance of the cloud product line. In the first half of 2025, 2.87 billion RMB of its revenue came from cloud-based intelligent chips and their supporting cards and server products, accounting for a staggering 99.6%.³ This fully demonstrates the success of the company’s strategic shift to the cloud, but it also reflects the marginalization of its other business segments, such as edge computing.

A long-standing issue for Cambricon that warrants continued attention is its high customer concentration. Between 2021 and 2023, the company’s annual revenue was highly dependent on a single customer. Data shows that the top customer contributed over 60% of revenue, even reaching 79% in 2024.⁵ These early major customers were mostly local state-owned enterprises purchasing products to build regional intelligent computing centers. This project-driven revenue model is characterized by high uncertainty and volatility.⁵

Entering 2025, as the company successfully penetrated top-tier internet companies, its customer structure has improved. Data from the first half of 2025 shows that the top five customers accounted for 85.31% of total accounts receivable and contract assets, with the largest customer’s share dropping to 42.5%.⁶ This indicates that the company’s customer base is expanding from single government projects to a more diverse group of leading tech companies.¹ This is a positive shift, as it means revenue sources are moving from one-off, unstable project contracts to more predictable sales based on the continuous computing power needs of large enterprises.

However, this shift also brings new challenges. Although the quality of customers has improved, the risk of relying on a few tech giants remains. These giant customers not only have immense bargaining power but are also actively developing their own AI chips, such as Baidu’s “Kunlun Core.” This creates a complex relationship of “co-opetition” between Cambricon and its clients. Today’s biggest customer could very well become tomorrow’s strongest competitor.

Investing in the Future: R&D Intensity

Sustained, high-intensity R&D investment is a core feature of Cambricon’s business model. Before achieving profitability, the company’s R&D investment as a percentage of revenue was extremely high, even exceeding 100%.³² Even after the revenue surge in the first half of 2025, the company’s R&D investment continued to grow, accounting for 15.85% of operating revenue.²

This seemingly “cash-burning” behavior is a necessary means of building and maintaining competitiveness in the knowledge- and capital-intensive chip design industry. It should not be seen as a financial flaw but rather as a continuous investment in the company’s future technological moat. Faced with the pressure of constantly accelerating technological iteration from international giants like Nvidia, only by maintaining high-intensity R&D can Cambricon hope to keep pace in the technology race and eventually catch up and surpass them.

Table 1: Key Financial Metrics (2020 - 2025 H1)

Year/Period	Total Revenue (Billion RMB)	Gross Margin (%)	R&D Expenditure (Billion RMB)	R&D as % of Revenue	Net Profit/(Loss) (Billion RMB)
2019	0.444	N/A	0.543	122.32%	(1.179)
2020	0.459	N/A	0.768	167.32%	(0.435)
2021	0.721	N/A	1.136	157.56%	(0.825)
2022	0.729	N/A	1.523	208.92%	(1.257)
2023	0.709	N/A	1.378	194.36%	(0.848)
2024	1.170	56.7%	1.068	91.30%	(0.450)
2025 H1	2.881	55.99%	0.457	15.85%	1.038

Note: Some early gross margin data was not available in the source materials. Data for 2019-2023 is primarily from historical financial reports.³ Data for 2024 and 2025 H1 is from the latest financial reports and analyses.²

V. The Battlefield: Competitive Positioning in a High-Stakes Market

The Global Benchmark: Cambricon vs. Nvidia

In assessing Cambricon’s competitiveness, any analysis must use the undisputed leader of the global AI chip market—Nvidia—as the benchmark. Nvidia’s dominance is built on two pillars: first, its consistently leading hardware performance, from the A100 and H100/H200 to the latest Blackwell architecture, with each generation redefining the upper limits of AI computing power ¹; and second, its deeply entrenched software ecosystem, built around the CUDA parallel computing platform and related libraries, which has become the de facto industry standard for developers.

From a pure hardware performance perspective, there is a clear generational gap between Cambricon and Nvidia. Its current flagship product, the Siyuan 590, uses a 7nm process technology and is designed to compete with Nvidia’s previous-generation A100 (also 7nm), released in 2020.¹ Multiple analyses suggest that the Siyuan 590’s overall performance reaches about 80% to 90% of the A100’s.¹ However, compared to Nvidia’s current mainstream product, the H100, which uses a 4nm process, the Siyuan 590 lags significantly in both performance and power efficiency.¹ And Nvidia’s newly announced Blackwell architecture widens this performance gap even further.

The gap in the software ecosystem may be even greater than in hardware. After more than a decade of development, Nvidia’s CUDA ecosystem has accumulated a vast collection of optimized libraries, development tools, and a massive developer community. Nearly all mainstream AI frameworks and applications are deeply optimized for CUDA. Although Cambricon’s NeuWare attempts to replicate CUDA’s successful model in its architecture, it is still in the very early stages of catching up in terms of maturity, breadth, and depth.

However, this purely technical comparison does not fully explain Cambricon’s current market success. In the unique context of geopolitics, Cambricon’s core value proposition in the Chinese market has fundamentally changed. The question is no longer “Can it outperform Nvidia in performance?” but rather “Can it be the best and most reliable alternative in a market where high-end Nvidia products are embargoed?” Due to U.S. export controls, Chinese tech companies cannot legally obtain high-performance chips like the A100/H100.¹² This shifts the purchasing criteria from “seeking the absolute best performance” to “choosing the best performance among all available, non-restricted options.”

Therefore, Cambricon does not currently need to be more powerful than the H100 to win large domestic orders. It only needs to be “good enough” and, most critically, be able to “supply” at scale and with stability. This asymmetric competitive advantage, created by external factors, provides Cambricon with a precious window of opportunity. During this period, it can use the revenue and profits from this protected market to invest in the R&D of its next-generation products, with the hope of genuinely narrowing the technological gap with the global leaders over the next two to three product cycles.

The Domestic Derby: A Crowded Field of Contenders

Within China, Cambricon is far from secure, facing an increasingly crowded and powerful competitive landscape.

• Huawei (Ascend): Huawei is Cambricon’s most formidable domestic competitor. With its deep technological expertise, strong system integration capabilities, extensive relationships with government and enterprise clients, and powerful brand influence, Huawei’s Ascend series of AI chips (such as the Ascend 910B) are direct competitors to the Siyuan 590.³ Huawei not only provides chips but also offers full-stack solutions, from servers and networking equipment to cloud services, which is highly attractive to many large customers.

• Other Players: Besides Huawei, the field is packed with other heavyweight contenders. Internet giants like Baidu (with its self-developed “Kunlun Core”) and Alibaba (through its “T-Head” subsidiary) are also investing heavily in their own chip development. They are both potential customers and direct competitors to Cambricon.³ Additionally, a group of well-funded AI chip startups, such as Biren Technology and Moore Threads, are also actively vying for market share.⁵

In such fierce competition, one of Cambricon’s key differentiating advantages is its “purity” and “independence.” Unlike competitors such as Huawei, Baidu, and Alibaba, Cambricon is a “pure-play” company focused solely on AI chip design. It does not offer cloud services or develop high-level applications, so it does not directly compete with its potential customers (like other cloud service providers or application developers). This neutral identity may give it a unique trust advantage when courting clients who are wary of their “supplier becoming a competitor.”

Table 2: Major AI Accelerator Competitive Benchmark (Simplified)

Chip Model	Company	Process Node	Peak Performance (FP16/BF16 TFLOPS)	Memory Type/Capacity	Memory Bandwidth (TB/s)
Siyuan 590	Cambricon	7nm	~500 (Estimated)	HBM2e	~1.2
Ascend 910B	Huawei	7nm+	376	64GB HBM2e	~0.4
Nvidia A100	Nvidia	7nm	312 (624 with Sparsity)	80GB HBM2e	2.0
Nvidia H100	Nvidia	4nm	989 (1979 with Sparsity)	80GB HBM3	3.35

Note: Official detailed specifications for the Siyuan 590 have not been fully released. The data in the table is primarily estimated based on analyst reports, benchmark products (A100), and publicly available market information.¹ Data for the Ascend 910B and Nvidia products are from public specification sheets.²⁵ Performance metrics (TFLOPS) may vary depending on specific configurations and conditions such as computational sparsity.

VI. The Geopolitical Catalyst: The Double-Edged Sword of Sanctions

Cambricon’s destiny is inextricably linked to the broader context of the U.S.-China tech competition. U.S. sanctions policy acts as a double-edged sword: it is both the Sword of Damocles hanging over the company and the powerful tailwind propelling its flight.

The “Entity List” and Supply Chain Vulnerability (Headwinds)

The direct negative impact of sanctions is manifested in Cambricon’s inclusion on the U.S. Department of Commerce’s “Entity List”.⁵ This move poses a fundamental threat to Cambricon’s fabless business model.¹³

As a fabless company, Cambricon’s operations are highly dependent on the global semiconductor industry chain. Its business process primarily involves:

1. Design: Using Electronic Design Automation (EDA) software tools from U.S. companies (like Synopsys and Cadence) for chip design.

2. Manufacturing: Sending the designed blueprints to third-party wafer foundries (like TSMC or SMIC) for production.

3. Packaging and Testing: Having specialized firms complete the final stages of the chip manufacturing process.

Being placed on the “Entity List” means that any company providing products or services containing U.S. technology to Cambricon requires a license from the U.S. government, and such licenses are generally denied. This creates severe supply chain risks on multiple levels:

• EDA Software: Cambricon may be unable to obtain the latest versions, updates, and technical support for EDA software. While it can use existing software in the short term, in the long run, this will prevent it from utilizing the most advanced design methodologies, causing it to fall behind competitors in chip design efficiency and complexity.³⁶

• Chip Manufacturing: Although Cambricon can opt for domestic foundries, the world’s most advanced manufacturing processes (such as 5nm and below) are still controlled by foundries like TSMC. These foundries rely heavily on U.S.-made production equipment (from companies like Applied Materials and Lam Research). The U.S. can leverage its technological influence to restrict these foundries from providing advanced process manufacturing services to entities on the list.⁴⁰ This could force Cambricon to use relatively outdated manufacturing processes for its future products, putting it at a disadvantage in terms of performance and power efficiency.

This deep reliance on the global supply chain is Cambricon’s core vulnerability and the focal point of its geopolitical risk.

The Urgency of “Domestic Substitution” (Tailwinds)

However, the indirect positive impact of the sanctions policy, for now, far outweighs its direct negative consequences. The comprehensive export control measures first introduced by the U.S. in October 2022 and subsequently tightened have strictly limited the sale of high-performance AI chips (like Nvidia’s A100 and H100) and related manufacturing equipment to any entity in China.³

This policy has, in effect, created a massive, protected market for domestic AI chip suppliers in China, with virtually no foreign competition. To avoid falling behind in the global generative AI race, China’s major tech companies have a rigid and urgent demand for high-performance computing power. Unable to obtain top-tier products from Nvidia, they have no choice but to turn to domestic suppliers like Cambricon and Huawei.¹

Thus, U.S. sanctions have unintentionally acted as the most powerful catalyst for China’s “domestic substitution” strategy. It has forcibly shifted the logic of market competition from a “performance battle in an open global market” to a “supply capability battle in a closed domestic market.” This is the most fundamental external reason for Cambricon’s explosive performance growth in the first half of 2025.

On a deeper level, the impact of this sanction-driven domestic substitution goes far beyond short-term revenue gains. It has, in fact, created an unprecedentedly large-scale “real-world R&D lab” for China’s AI chip companies, driven by national will and funded by market demand. Under normal market conditions, it would be extremely difficult for an emerging chip company to persuade customers of the scale of ByteDance, Baidu, or Alibaba to deploy its unproven hardware on a massive scale and invest significant engineering resources in software adaptation, as customers would always prefer the mature, stable, and well-supported Nvidia solution.

But now, sanctions have eliminated this “optimal choice.” These top Chinese tech companies are not just Cambricon’s customers; they have objectively become “co-builders” in its product iteration. In the process of deploying thousands or even tens of thousands of Cambricon accelerator cards, they will encounter various hardware performance bottlenecks, software compatibility issues, and deficiencies in the ecosystem toolchain. They will provide the most direct and valuable feedback to Cambricon, driving it to fix bugs, optimize performance, and perfect its NeuWare software ecosystem at an unprecedented pace. The value of this customer-driven, high-intensity, real-world iteration cycle is immeasurable. It can be said that the U.S. policy, intended to slow down China’s AI development, has inadvertently and dramatically accelerated the maturation of China’s domestic AI chip hardware and software ecosystems.

VII. Strategic Outlook: Overall Assessment, Risks, and Future Trajectory

Comprehensive SWOT Analysis

• Strengths:

  • Deep “National Team” Background: Originating from the Chinese Academy of Sciences, it is closely linked to the national research system and enjoys policy and resource advantages.

  • Focused ASIC Architecture: Potential for high power efficiency in AI-specific workloads.

  • Unified Software Strategy: NeuWare aims to build a unified ecosystem across cloud, edge, and end devices, which is the correct strategic direction.

  • First-Mover Advantage: As China’s “first AI chip stock,” it has a head start in brand recognition and market access.

• Weaknesses:

  • Technology and Ecosystem Gap: A significant gap exists with industry leader Nvidia in terms of hardware performance and software ecosystem maturity.

  • High Customer Concentration: Although improving, revenue is still highly dependent on a few large customers, limiting bargaining power and business stability.

  • Supply Chain Vulnerability: The fabless model makes it susceptible to geopolitical sanctions affecting EDA software and advanced process manufacturing.

  • Short Profitability History: Has been in a loss-making state for a long time and has only just become profitable; the sustainability of its profitability remains to be proven.

• Opportunities:

  • Huge Protected Market: U.S. sanctions have created enormous demand for “domestic substitution,” providing the company with a historic growth opportunity.

  • AIGC Wave: The development of generative AI and large models continues to drive exponential demand for high-performance computing power.

  • Becoming the Domestic Standard: An opportunity to become the de facto standard for non-Nvidia AI computing in the Chinese market, building a localized ecosystem moat.

• Threats:

  • Intense Domestic Competition: Faces fierce competition from giants like Huawei and numerous startups; market share is not secure.

  • Risk of Escalating Sanctions: The U.S. may further tighten controls, extending restrictions to EDA tools or mature process manufacturing, which could be a fatal blow.

  • Risk of Customer In-House Development: Major customers are tech giants with the capability to develop their own chips, posing a long-term risk of being replaced.

  • High Valuation Risk: The soaring stock price has already priced in extremely high growth expectations; any failure to meet performance expectations could trigger a sharp stock price correction.¹

Core Risks to the Investment Thesis

• Risk of Falling Behind in Technological Iteration: The pace of technological iteration in the semiconductor industry is extremely fast. Nvidia introduces a new generation of architecture (like the Blackwell platform) every 18-24 months, constantly raising the performance bar. Whether Cambricon’s R&D speed can keep up, or even narrow the gap, is key to its long-term survival. If the technology gap continues to widen, its products will lose their competitiveness.¹

• Risk of Ecosystem Building Failure: Whether NeuWare can attract enough developers to form a vibrant, self-reinforcing ecosystem is central to its ability to build a long-term moat. If NeuWare ultimately remains a closed toolset serving only a few large customers, rather than becoming an open, widely adopted platform, Cambricon will never escape its sole reliance on hardware performance and will be unable to truly challenge CUDA’s dominance.

• Risk of Geopolitical “Hard Decoupling”: While current sanctions leave some “gray areas,” the risk is that the U.S. could take more extreme measures, such as using long-arm jurisdiction to completely prohibit key foundries like TSMC from providing services to Cambricon, or cutting off the supply of EDA software entirely. If such a “hard decoupling” scenario were to occur, it would directly paralyze Cambricon’s production capabilities.⁴⁰

• Risk of Valuation and Expectation Mismatch: The capital market has already painted an extremely optimistic future for Cambricon, with its high stock price incorporating the expectation of it becoming the “Nvidia of China.” This means the company must continue to deliver ultra-high growth in the coming years. Any execution missteps, loss of market share, or changes in the macroeconomic environment could lead to a collapse in market confidence and a sharp correction in its valuation.

Forward-Looking Scenarios and Conclusion

• Bull Case: Cambricon successfully utilizes the protected market window of the next 2-4 years to rapidly iterate its hardware products, bringing their performance closer to international mainstream levels. Meanwhile, the NeuWare ecosystem matures quickly, driven by top-tier customers, attracting a large number of developers and becoming one of the preferred platforms for AI development in China. The company successfully diversifies its customer base, its profitability continues to strengthen, and it ultimately becomes an unshakeable leader in China’s AI infrastructure, firmly holding a core share of the domestic market.

• Bear Case: Competitive pressure from domestic rivals like Huawei exceeds expectations, leading to market share erosion. The NeuWare ecosystem fails to gain critical developer support and remains dependent on customized services for a few large clients. At the same time, U.S. sanctions are further tightened, completely cutting off its access to advanced manufacturing processes or key EDA tools. The company’s growth stagnates, the technology gap widens, and it ultimately fails to live up to the market’s high expectations, becoming just another player in a fragmented market.

Final Conclusion

Cambricon is a classic high-risk, high-reward investment, with its corporate destiny deeply intertwined with the national strategy of the U.S.-China tech competition. While its recent brilliant achievements are built on years of technological accumulation, they have been catalyzed and amplified in the greenhouse of geopolitical protectionism. This is a fragile but precious success.

The company’s future will depend on whether it can transform this temporary market advantage, granted by external factors, into a durable, endogenous competitive moat built on its own technological strength and ecosystem. The next 24 to 36 months will be the decisive period for observing whether it can complete this critical leap. Investors need to closely monitor the performance of its next-generation products, the developer adoption rate of the NeuWare ecosystem, and any changes in the global semiconductor supply chain landscape. The story of Cambricon is not just about the rise and fall of a single company; it is a microcosm of China’s exploration and challenges on its path toward technological self-reliance.

Cited Works

1. 6000亿寒武纪，可以开香槟了吗？-钛媒体官方网站, https://www.tmtpost.com/7676298.html

2. 寒武纪，凭什么 - 中国新闻周刊, http://www.inewsweek.cn/finance/2025-08-28/26453.shtml

3. “85后”陈天石与6000亿寒武纪：从象牙塔到A股之巅 - 21财经, https://www.21jingji.com/article/20250829/herald/ba45f57c2d257ccdc11c720c661a1282.html

4. 透视半年报｜8年利润首次转正，寒武纪的扭亏密码 - 新京报, https://www.bjnews.com.cn/detail/1756292228129091.html

5. 千亿市值，寒武纪靠什么支撑？ - Supplyframe 四方维, https://cn.supplyframe.com/article/7343.html

6. 寒武纪因何一个月市值翻倍？ - 财经》客户端, https://www.mycaijing.com/article/detail/554586?source_id=40

7. “寒王”贵过茅台，1900亿陈天石“封神”-钛媒体官方网站, https://www.tmtpost.com/7675708.html

8. 寒武纪-U（688256） 2024 年报&25Q1 季报点评：第一大客户营收 …, https://pdf.dfcfw.com/pdf/H3_AP202504251662074955_1.pdf

9. AI芯片第一股寒武纪来了 - 人民日报, http://paper.people.com.cn/zgjjzk/html/2020-07/15/content_1999301.htm

10. 寒武纪-U（688256）, https://pdf.dfcfw.com/pdf/H3_AP202506031684050051_1.pdf?1748960324000.pdf

11. 联合研究丨公司深度丨寒武纪, https://pdf.dfcfw.com/pdf/H3_AP202409021639684152_1.pdf

12. 寒武纪称“王”，为时尚早 - 21财经, https://www.21jingji.com/article/20250424/herald/79982fbac76886c2fadd4d7770989819.html

13. 寒武纪最新公告，事关公司全年经营预测、新产品情况等 - 财联社, https://www.cls.cn/detail/2129451

14. 新“股王”寒武纪狂飙真相 - 证券时报, https://www.stcn.com/article/detail/3290970.html

15. 成为新“股王”后，寒武纪紧急回应：公司未有新产品发布计划 - 21财经, https://www.21jingji.com/article/20250828/herald/a59c0e74845978826068f0c6de5c4d9c.html

16. 寒武纪85后创始人，身家超1500亿 - 新湖南, https://m.voc.com.cn/xhn/news/202508/30298537.html

17. 开创“寒武纪”—-科技自立自强之路 - 中国科学院, https://www.cas.cn/zt/kjzt/kjzlzqzl/kjzlzq/202409/t20240919_5032328.shtml

18. 【中国科学报】“创客”陈天石和爆发的寒武纪, https://www.cas.cn/cm/201901/t20190103_4675787.shtml

19. 现场| 中国AI芯片独角兽寒武纪首场发布会 - 麻省理工科技评论, https://mittrchina.com/news/detail/718

20. 寒武纪科技- 维基百科，自由的百科全书, https://zh.wikipedia.org/zh-cn/%E5%AF%92%E6%AD%A6%E7%B4%80%E7%A7%91%E6%8A%80

21. 2000亿！寒武纪背后老板赢麻了！ - 新浪财经, https://cj.sina.cn/articles/view/1010395044/3c3967a401901lz94

22. 寒武纪, https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=7

23. 寒武纪- 维基百科，自由的百科全书, https://zh.wikipedia.org/zh-cn/%E5%AF%92%E6%AD%A6%E7%BA%AA

24. 寒武紀科技- 維基百科，自由的百科全書 - Wikipedia, https://zh.wikipedia.org/wiki/%E5%AF%92%E6%AD%A6%E7%B4%80%E7%A7%91%E6%8A%80

25. 业绩承压、需求回暖，看好新质生产力——计算机行业2024年中期 …, https://pdf.dfcfw.com/pdf/H3_AP202406241636810979_1.pdf

26. 思元370系列- 寒武纪, https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=360

27. 寒武纪-U（688256.SH）, https://pdf.dfcfw.com/pdf/H3_AP202506171692418270_1.pdf?1750150897000.pdf

28. MLU220-M.2边缘端智能加速卡- 寒武纪, https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=57

29. 寒武纪MLU220解决方案, https://edit.wpgdadawant.com/uploads/news_file/blog/2021/3759/tech_files/sv-1621827910.pdf

30. 寒武纪220边缘计算模组- 大大通(繁體站), https://www.wpgdadatong.com/blog/detail/41624

31. 寒武纪发布边缘AI芯片思元220，布局AIoT新时代 - InfoQ, https://www.infoq.cn/article/aztcmlhdy8mimcmhrees

32. 寒武纪(688256) - 2024年年度报告 - 新浪, https://vip.stock.finance.sina.com.cn/corp/view/vCB_AllBulletinDetail.php?stockid=688256&id=10904723

33. 用5万换价值73亿的股份？因离职让寒武纪市值蒸发60亿的CTO，与老东家“开撕”了！ - InfoQ, https://www.infoq.cn/article/lcmlz5brbofxkuu1ixnc

34. The Limits of Chip Export Controls in Meeting the China Challenge - CSIS, https://www.csis.org/analysis/limits-chip-export-controls-meeting-china-challenge

35. 未来智造局| 寒武纪股价登顶A股国产GPU阵营“拨云见日” - 新华财经, https://m.cnfin.com/gs-lb//zixun/20250828/4293748_1.html

36. 美国升级芯片管制③｜严卡中国AI芯片设计壁仞等13家公司被纳入“实体清单” - 财新, https://m.caixin.com/m/2023-10-18/102118118.html

37. 南昌38岁芯片天才，半年收获176亿财富 - 21财经, https://www.21jingji.com/article/20230713/herald/385c0132682b01b056a9aa5c4b45a61b.html

38. 独家丨寒武纪回应被列入“实体清单”：影响还在评估, https://www.jwview.com/jingwei/html/m/12-16/518104.shtml

39. Commerce Strengthens Export Controls to Restrict China’s Capability to Produce Advanced Semiconductors for Military Applications - Bureau of Industry and Security, https://www.bis.gov/press-release/commerce-strengthens-export-controls-restrict-chinas-capability-produce-advanced-semiconductors-military

40. EXPORT CONTROLS AND US-CHINA TECHNOLOGY COMPETITION IN AN INTERDEPENDENT WORLD - Brookings Institution, https://www.brookings.edu/wp-content/uploads/2024/07/20240701_Branstetter_Sanctions.pdf

41. Thin Ice: US Pathways to Regulating China-Sourced Legacy Chips - Rhodium Group, https://rhg.com/research/thin-ice-us-pathways-to-regulating-china-sourced-legacy-chips/

42. The True Impact of Allied Export Controls on the U.S. and Chinese Semiconductor Manufacturing Equipment Industries - CSIS, https://www.csis.org/analysis/true-impact-allied-export-controls-us-and-chinese-semiconductor-manufacturing-equipment

43. 寒武纪能否实现1835元目标价？ - 证券时报, https://www.stcn.com/article/detail/3291090.html