Cambricon Technologies: The Crucible of China's AI Chip Champion

I. Introduction & Episode Roadmap

On the evening of March 12, 2026, 中科寒武纪科技股份有限公司 Cambricon Technologies Corporation Limited filed an announcement with the 上海证券交易所 Shanghai Stock Exchange that ran to a few hundred words and carried the deliberately bloodless title of a housekeeping notice. Announcement 2026-006 informed the market that, having achieved its first profit, the company's stock abbreviation would change from "寒武纪-U" to "寒武纪" effective March 16, 2026. The securities code, 688256, would not change.¹

To anyone outside China's capital markets, this reads as clerical trivia. Inside them, that single suffix carried five years of accumulated judgment. When the 科创板 STAR Market opened in 2019, it broke with decades of mainland listing orthodoxy by allowing unprofitability at IPO — provided the company wore a marker. The "U" tag designated an issuer that went public while losing money, serving as a persistent reminder to retail investors that they were funding a promise rather than an earnings stream. Cambricon had worn it since July 2020. Removing it required the company to achieve something it had never managed in nine years of existence: turn a profit over a full financial year.

The numbers behind the filing were substantial. For 2025, Cambricon reported revenue of RMB 6.497 billion, a 453% increase year on year, and net profit attributable to shareholders of RMB 2.059 billion — compared with a net loss of RMB 452 million the prior year.² Excluding non-recurring items, net profit reached RMB 1.770 billion, the stricter benchmark Chinese regulators use for the "first profit" test.¹ Cambricon became the first company to exit the STAR Market's unprofitability designation entirely. Growth continued into 2026: first-half revenue reached RMB 5.996 billion, up 108% year on year, while net profit rose approximately 123% to RMB 2.311 billion.³⁴ That marked seven consecutive profitable quarters.³

This episode examines the forces behind that turnaround. From 2016 through 2024, Cambricon stood as a case study in the extreme difficulty of building an independent semiconductor business. The company accumulated roughly RMB 3.8 billion in net losses across its first five post-IPO years, exhausted its IPO proceeds, saw its anchor customer replace its IP with an in-house design, was placed on the U.S. Entity List, and lost its chief technology officer amid ongoing litigation. Yet within eighteen months, it emerged as one of the two central providers of Chinese AI silicon and briefly became the highest-priced stock in China. The core analytical question is how much of this transformation stemmed from internal capability versus state-driven geopolitical shifts.

That distinction — between earned competitive capability and state-directed demand — forms the core framework for analyzing the company across four key themes:

The fragility of pure-play intellectual property. Cambricon's initial business focused on licensing neural-network processor designs to 华为海思 Huawei HiSilicon, whose Kirin 970 chip powered what Huawei marketed in 2017 as the world's first smartphone with dedicated on-device AI hardware.⁵ While providing immediate commercial validation, this model offered little long-term moat. Examining how quickly that advantage evaporated illustrates why the ARM licensing analogy favored by Cambricon's early investors was structurally flawed.

The pivot crucible. Cambricon was forced to rebuild its operating model twice under duress: first moving from IP licensing into government-funded 智能计算中心 intelligent computing center construction — a strategy that expanded top-line revenue while weakening underlying economics — and later, following U.S. export controls, adjusting product design from advanced foreign foundries to the domestic manufacturing limits of 中芯国际 SMIC.

Geopolitics as an exogenous moat creator. The primary driver of Cambricon's revenue expansion was not an internal product milestone, but policy shifts: Washington restricted NVIDIA's access to China, and Beijing subsequently directed major domestic technology firms to halt purchases of NVIDIA hardware.⁶ Policy compressed a decade of enterprise customer acquisition into two years. This shift provided a powerful tailwind, yet it remains a sensitive foundation, as policy mechanisms that generate demand can also alter it.

The software problem. Silicon represents only half the challenge. 寒武纪 Cambricon Neuware — comprising the compiler, runtime environment, libraries, and framework integrations necessary for machine-learning deployments — dictates whether hardware is actively deployed or remains unutilized. Cambricon has made measurable progress on software integration, but a substantial gap remains compared to NVIDIA's CUDA ecosystem, defining a primary risk factor in the bear thesis.

Three prevalent myths require clarification before examining the history. Market consensus around Cambricon formed rapidly, leaving three key assertions that require scrutiny:

First, the characterization of Cambricon as "China's NVIDIA" is inaccurate and misleading. Independent benchmarks evaluate its flagship processor at approximately 80% of NVIDIA's A100 performance in specific workloads, estimating the company's silicon design to be four to five years behind NVIDIA's latest architecture.⁷ Cambricon's gross margins hover in the mid-50% range, whereas NVIDIA's data center segment operates in the 70% range. Furthermore, Cambricon relies on a single main product line served almost entirely within one domestic market. A more precise description is that Cambricon is the primary merchant supplier of domestically fabricated AI inference hardware in a market where foreign competition is restricted.

Second, the view that the 2025 financial inflection resulted solely from technological superiority misinterprets the timeline. While product capabilities advanced, the rate of technological improvement did not change abruptly in 2024. Instead, the customer base shifted. The company recorded its first profitable quarter in Q4 2024, following consecutive rounds of U.S. export controls that removed competitive NVIDIA products from the market. Demand further accelerated when Beijing instructed major technology companies to transition away from NVIDIA silicon in September 2025.⁶ Consequently, macro regulatory changes were the primary catalyst for the revenue shift.

Third, the premise that Cambricon successfully diversified its revenue base is unsupported by reported figures. The company transitioned from relying heavily on a single early client to deriving roughly four-fifths of its revenue from a single primary buyer at significantly larger volume.⁷ Customer concentration remains a structural characteristic of its current business model.

The analysis traces Cambricon's trajectory from its origins at the 中国科学院计算技术研究所 Institute of Computing Technology, Chinese Academy of Sciences, through the end of its Huawei partnership, its STAR Market listing, the impact of U.S. sanctions, its domestic supply chain restructuring, and the revenue expansion of 2025 — evaluating the long-term sustainability of its market position amidst shifting policy environments.

Subscribe on LinkedIn

↑ Back to Top

II. Academic Foundations & The DianNao Breakthrough (2008–2016)

The Institute of Computing Technology sits in Beijing's Haidian district, a short walk from Tsinghua and Peking universities. It holds a distinct place in Chinese technological history as the birthplace of the 龙芯 Loongson processor, the country's first major effort to build a general-purpose CPU outside the Intel and ARM ecosystems. For decades, the institution pursued a core question: what would it take for China to design its own computing architecture from the ground up?

Two brothers from Jiangxi province approached that question from an unconventional direction. 陈云霁 Chen Yunji, born in 1983, and 陈天石 Chen Tianshi, born in 1985, both entered the University of Science and Technology of China's elite youth class and completed computer science doctorates by age 24.⁷ Chen Yunji specialized in computer architecture and joined the Loongson team, while Chen Tianshi focused on artificial intelligence and machine learning theory. Throughout most of the 2000s, these disciplines remained largely separate. Neural networks were an academic niche, and processor architects rarely crossed paths with machine-learning theorists.

At the time, bridging the two fields was a contrarian bet. In traditional computer architecture, designing a single-purpose processor was seen as impractical. The industry favored general-purpose designs to amortize the high fixed costs of chip design and fabrication. Committing to a dedicated neural-network accelerator in the late 2000s meant betting that machine learning would grow from a specialized academic subfield into a mainstream computational workload — a premise that would not be commercially validated for years.

By combining their respective specialties, the brothers began exploring a fundamental architectural question: how would a processor perform if it were designed strictly to execute neural networks?

The value of that approach stems from the limitations of general-purpose hardware. A standard CPU is designed for versatility, running diverse tasks from spreadsheets to operating systems. That flexibility requires significant silicon real estate and energy for control logic — the circuitry responsible for instruction scheduling and flow control. Graphics processing units (GPUs) improved parallel throughput by executing thousands of identical arithmetic operations simultaneously, making them well suited for deep learning. However, GPUs remained architectural descendants of graphics rendering chips, carrying legacy overhead.

Neural-network workloads, by contrast, consist primarily of matrix multiplication and accumulation repeated billions of times with predictable data movement. Tailoring hardware to these regular operations allows designers to reduce control logic, hard-wire data paths, and place dedicated memory directly adjacent to arithmetic units. Because data transfer consumes far more power than computation, optimizing the memory architecture for neural-network access patterns delivers significant efficiency gains over general-purpose hardware.

This architecture underpinned the DianNao research series, named after 电脑 diànnǎo, the Chinese word for computer (literally "electric brain"). The initial DianNao paper, presented at ASPLOS in 2014 in collaboration with Olivier Temam of Inria, introduced a compact accelerator that achieved marked speed and energy improvements over traditional CPUs on neural-network workloads, earning the conference's best paper award.⁷ Later that year at MICRO, DaDianNao expanded the design into a multi-chip architecture for model training, also securing a best paper award. Subsequent iterations broadened the design family: PuDianNao extended acceleration to non-deep-learning machine-learning algorithms, while ShiDianNao placed the processor adjacent to an image sensor, anticipating edge-inference applications.

The DianNao papers represented a notable academic milestone. In 2014, the team became the first Chinese research group to win back-to-back best paper awards at top computer architecture conferences, establishing domain-specific neural-network accelerators well before Google publicly disclosed its first Tensor Processing Unit in 2016. The research established Cambricon's technical foundation in chip design.

In March 2016, the brothers incorporated Cambricon Technologies, backed by roughly RMB 10 million in initial funding from the Institute's commercial arm and Chinese venture capital investors.⁷ The name Cambricon combines "Cambrian" — referencing the Cambrian explosion of biological diversity — with "silicon," mirroring its Chinese name, 寒武纪. The company's founding strategy envisioned a major expansion in specialized AI computing forms, aiming to position itself like ARM by licensing processor IP and earning royalties across third-party hardware without manufacturing silicon directly.

While strategically compelling on paper, the IP licensing model faced immediate execution challenges. Over the next three years, Cambricon's initial anchor customer demonstrated the structural limits of that approach.

↑ Back to Top

III. The Huawei Mobile NPU Boom & The Near-Death Split (2017–2019)

On September 2, 2017, at the IFA electronics show in Berlin, 余承东 Richard Yu announced the Kirin 970, the system-on-chip powering the Huawei Mate 10. The flagship feature was a dedicated neural processing unit designed specifically to execute AI models directly on the handset, eliminating latency from cloud data centers. Huawei highlighted it as the first mobile chipset featuring a dedicated AI processing block, showcasing real-time on-device image recognition at speeds standard mobile CPUs could not match.⁵

That underlying NPU intellectual property was the Cambricon-1A, licensed from a Beijing startup just eighteen months old.

For Cambricon, securing Huawei was an ideal commercial launchpad. As an aggressive vendor in the premium smartphone market, Huawei shipped massive flagship volumes while using custom silicon to differentiate its products from Qualcomm-powered competitors. Integrating the Cambricon-1A into the Kirin processor validated Cambricon's IP at the highest volume tier of consumer electronics. The relationship deepened in 2018 when the follow-on Cambricon-1H IP was integrated into the Kirin 980 for the Huawei Mate 20. Backed by high-profile validation, Cambricon's valuation swiftly surpassed unicorn status, while domestic media positioned the company as China's national champion in AI silicon.

Financial disclosures from that era revealed severe customer concentration. In 2017 and 2018, IP licensing to Huawei accounted for the overwhelming majority of Cambricon's revenue, reaching up to 98% by some estimates.⁷ The startup essentially relied on a single product design, a single key customer, and a single primary contract. While terminal IP licensing revenue reached approximately RMB 117 million in 2018 — a notable figure for a young startup — it represented a small fraction of Huawei's overall silicon expenditure.⁷

This revenue asymmetry created strategic friction. From Huawei's perspective, the NPU was no longer a peripheral accelerator, but a core marketing centerpiece for flagship smartphones and a foundational block for its broader data-center AI ambitions. Licensing third-party IP granted access to silicon architecture, but lacked control over the long-term technical roadmap. Huawei could not co-optimize an external design against its proprietary camera pipelines, software stacks, and thermal constraints as tightly as an in-house architecture would permit. Furthermore, licensing fees represented ongoing margin leakage to a vendor whose IP HiSilicon engineers had spent two product cycles integrating and mastering.

Testing the moat. Early investors underwrote Cambricon on the thesis that pioneer status in mobile NPU design would yield an ARM-like licensing model — delivering sticky, high-margin revenue that compounded with every smartphone shipped. However, subsequent developments directly contradicted that premise.

In 2019, Huawei unveiled its proprietary 达芬奇 DaVinci NPU architecture, restructuring its AI hardware strategy around in-house designs. Huawei introduced the DaVinci NPU in the mid-range Kirin 810 in June 2019, before expanding the architecture upward into the flagship Kirin 990 and outward into standalone 华为昇腾 Huawei Ascend data center accelerators.[^8] Consequently, Cambricon's terminal IP revenue collapsed from roughly RMB 117 million to the mid-teens of millions of RMB — a drop of over 80% — within approximately two years, effectively eliminating customer retention with its primary buyer.⁷

The loss of Huawei reflected a structural vulnerability rather than a routine commercial dispute or quality issue. The customer driving nearly all of Cambricon's revenue designed an in-house replacement, deployed it across its entire product lineup, and severed the licensing relationship. The switching costs intended to protect Cambricon's model — such as developer familiarity and software toolchain integration — proved minimal for a vertically integrated technology enterprise with extensive software engineering capabilities. The comparison to ARM failed structurally: ARM's competitive advantage rests on a ubiquitous instruction set architecture backed by decades of software integration across hundreds of global licensees. Cambricon, by contrast, relied on a single dominant licensee without a broad third-party software ecosystem or industry-wide standard to defend.

This trajectory demonstrated that Cambricon's original IP licensing moat was largely illusory. Any assessment of the company's long-term customer lock-in must rely on factors beyond its initial mobile IP thesis, particularly given that customer concentration remains a defining vulnerability in its business model.

The Huawei pivot also highlighted the broader challenge of Cambricon's market position. The failure was not driven by technical flaws — the Cambricon-1A functioned effectively in tens of millions of devices — but by value-chain placement. Supplying a core differentiating component to a customer equipped to develop the same technology in-house provided temporary market access rather than durable equity. Cambricon's subsequent strategy focused on escaping that constraint by transitioning from IP licensing to merchant chip manufacturing, constructing a proprietary software ecosystem, and targeting buyers seeking off-the-shelf hardware rather than custom silicon.

By late 2019, Cambricon faced pressing financial realities: its core licensing revenue had fallen sharply, its emerging product lines remained in early development, and net losses continued to widen. Capital requirements grew urgent, just as Chinese regulators launched a new public market venue designed for high-growth, unprofitable technology enterprises.

↑ Back to Top

IV. STAR Market IPO & The Smart Computing Center Stopgap (2020–2022)

The STAR Market had been open for almost exactly one year when Cambricon filed to list, making the company an ideal test case for the new exchange. Beijing established the board in July 2019 with an explicit policy goal: provide domestic hard-technology companies—across semiconductors, biotechnology, and advanced manufacturing—a public listing venue that eliminated traditional profitability requirements, discouraging firms from seeking capital in New York or Hong Kong. Cambricon matched the profile precisely, combining prestigious academic roots with a collapsing revenue base and an urgent requirement for fresh capital.

The offering priced at RMB 64.39 per share for 40.1 million shares, raising RMB 2.582 billion gross and RMB 2.498 billion net, and valuing the company at roughly RMB 25.8 billion at the offer price.⁸ When trading opened on July 20, 2020, the stock price more than tripled. Over thirty institutional investors had participated in pre-IPO funding rounds. The company went public carrying the "U" suffix reserved for unprofitable issuers, finishing 2020 with RMB 459 million in revenue alongside a net loss of RMB 435 million.⁹

Fresh capital did not immediately resolve the core commercial challenge: Cambricon had cash, but lacked a scalable enterprise business.

The Huawei licensing agreement had ended, and Cambricon's proprietary cloud accelerator cards—the 思元 Siyuan series, designated MLU for machine learning unit—struggled to gain commercial traction. In 2020, Chinese cloud operators had little incentive to purchase unproven domestic AI chips over established NVIDIA A100 GPUs. To generate immediate top-line revenue, Cambricon targeted a customer segment where procurement decisions were guided by policy mandates rather than pure price-to-performance metrics: local government.

Building computers for cities. Beginning in 2019, Chinese municipal and provincial governments began funding 智能计算中心 intelligent computing centers—publicly financed AI data centers framed as foundational digital infrastructure and local technology catalysts. Cambricon secured a series of these projects, signing cooperation agreements with the Zhuhai Hengqin New Area and Shaanxi's Xixian New Area that generated roughly RMB 210 million and RMB 81 million, respectively, in 2019.¹⁰ In 2020, the company won an order of approximately RMB 300 million for the first phase of Nanjing's intelligent computing center, followed in 2021 by a contract of roughly RMB 500 million for computing infrastructure in Kunshan, Jiangsu.¹⁰

The project mechanics differed fundamentally from standard chip sales. These were turnkey system integration contracts in which Cambricon procured third-party servers, storage, and networking hardware, integrated its own accelerator cards, and delivered operational compute clusters. Revenue recognition was project-based and volatile. In 2021, the intelligent computing cluster segment generated RMB 455.6 million, representing about 63% of Cambricon's RMB 721 million in total revenue. A single buyer, the Kunshan municipal development vehicle, accounted for roughly RMB 450 million, or 62.5% of total corporate sales that year.¹⁰¹¹

Testing the diversification claim. Management framed this transition as proof of enterprise execution and successful revenue diversification away from mobile IP licensing. Evaluated across gross margins, cash conversion, and revenue repeatability, however, that thesis shows structural weaknesses.

On margins, the revenue mix diluted overall profitability. Cambricon reported blended gross margins of 65.4% in 2020 and 62.4% in 2021, figures that appeared resilient only because legacy chip and IP sales remained in the overall tally.¹¹ Procuring and reselling third-party server hardware compressed underlying economic returns, effectively turning a fabless semiconductor designer into a low-margin system integrator for that segment of its business.

On cash conversion, municipal buyers operated on extended payment cycles. Because Cambricon funded hardware procurement and integration costs upfront against slow-paying government receivables, operating cash flow remained persistently negative throughout this period.

On repeatability—the critical test for enterprise valuation—computing center construction functioned as a series of discrete events rather than a recurring business model. Municipalities typically build a single intelligent computing center without annual hardware upgrade cycles or expanding software subscriptions. Consequently, Cambricon replaced customer concentration with a single smartphone manufacturer for annual customer concentration with individual municipal development entities, accepting lower overall gross margins in the process.

Intelligent computing center projects functioned not as a permanent commercial expansion, but as an operational bridge. The contracts generated immediate revenue to sustain operations while commercial cloud silicon sales remained stagnant, buying the company time to refine its product stack.

The CTO leaves. Strategic disagreement over this computing center focus triggered internal executive turnover. On March 15, 2022, Cambricon disclosed the departure of chief technology officer and vice president 梁军 Liang Jun, who had joined in October 2017 from Huawei HiSilicon to industrialize the company's chip development. The disclosure noted that Liang submitted his resignation on February 10, 2022, following strategic disagreements over commercial direction: leadership prioritized rapid product deployment to capture immediate market opportunities, whereas Liang favored deeper, long-term technical research. During his tenure, Liang was named in 138 invention patent applications.¹²

While the initial announcement noted a standard two-year non-compete agreement, extensive litigation followed. In 2023, entities administering Cambricon's equity incentive plans sued Liang to compel cooperation with a share buyback. Liang counter-sued in Beijing's Haidian District Court in October 2024, seeking confirmation of his employment relationship and RMB 4.287 billion in compensation for unvested equity incentives—a claim Cambricon disclosed on November 1, 2025, confirming it would contest the suit.¹³¹⁴

The executive departure and legal dispute reflected the internal pressures of Cambricon's survival strategy between 2020 and 2022. Choosing immediate deployment over foundational research proved pragmatic when subsequent geopolitical shifts opened an unexpected market window, even if that window differed from what internal debates had envisioned.

In late 2022, U.S. export controls fundamentally altered the competitive landscape for domestic AI hardware.

↑ Back to Top

V. Geopolitical Shockwaves & The Domestic Foundry Crucible (2022–2023)

On December 15, 2022, the U.S. Department of Commerce's Bureau of Industry and Security added 36 technology firms to its Entity List. While initial coverage focused on memory manufacturer Yangtze Memory Technologies, the designation extended to twenty-one entities involved in artificial intelligence chip research, design, manufacturing, and sales—including Cambricon Technologies and a comprehensive sweep of six key subsidiaries across Anhui, Hong Kong, Kunshan, Nanjing, and Xi'an.¹⁵[^17]¹⁶ The listings subjected Cambricon to Foreign Direct Product Rule (FDPR) restrictions, with the formal regulation published in the Federal Register the following day.[^17]

The FDPR fundamentally alters the reach of trade sanctions. Whereas conventional export controls restrict direct sales from American suppliers, the FDPR extends jurisdiction to any foreign-manufactured item produced using U.S.-origin technology or software. Because advanced semiconductor foundries globally rely on American electronic design automation (EDA) tools and manufacturing equipment, the rule prohibited leading software vendors such as Synopsys and Cadence from renewing license agreements, while barring Taiwan Semiconductor Manufacturing Company (TSMC) from fabricating Cambricon's chip designs.

As TSMC had served as Cambricon's primary foundry,¹⁷ the designation immediately severed the company's access to manufacturing lines, software updates, and essential third-party IP blocks—such as licensed memory controllers and high-speed interfaces. Chairman Chen Tianshi later acknowledged that the Entity List inclusion disrupted the company's supply chain and constrained product delivery.¹⁸

What "redesign for a domestic foundry" actually means. Transitioning chip manufacturing to 中芯国际 Semiconductor Manufacturing International Corporation (SMIC) involved far more than transferring design files to a new production line.

Modern semiconductor architectures are built for a specific foundry's process design kit (PDK)—the software library dictating transistor behavior, wiring resistance, clock speeds, and memory density. Migrating to SMIC required recalculating these physical parameters. Furthermore, because SMIC's advanced nodes rely on deep ultraviolet (DUV) lithography rather than extreme ultraviolet (EUV) systems—which Dutch equipment maker ASML cannot ship to China under U.S. export rules—achieving higher transistor density requires multi-patterning techniques. Exposing silicon wafers through multiple lithography passes reduces production throughput, raises manufacturing costs, and depresses yield rates, particularly for complex silicon architectures.

Because AI accelerators require large die surface areas, Cambricon could not simply port existing designs; it had to re-architect its processors around SMIC's manufacturing capabilities. Engineering teams redesigned physical layouts, interconnect structures, and memory interfaces to fit domestic process node limits. In the process, planned tape-outs were cancelled, while completed photomasks and in-process wafers engineered for TSMC lines were rendered unusable.

Financial disclosures reflected the severe operational drag of this transition. Revenue remained stagnant over three years—generating RMB 721 million in 2021, RMB 729 million in 2022, and RMB 709 million in 2023—even as research spending climbed and net losses widened to RMB 1.26 billion in 2022 before narrowing to RMB 848 million in 2023.¹¹[^21] Across the five post-IPO years through 2024, cumulative losses reached approximately RMB 3.8 billion,⁹ fully consuming initial public offering proceeds.

Testing the agility claim. Management had long framed Cambricon as an agile fabless designer capable of adapting across manufacturing partners. The U.S. sanctions provided an empirical test of that positioning. While Cambricon successfully migrated production to a domestic foundry and brought silicon to market—outperforming several domestic peers that struggled to adapt—the transition was neither seamless nor low-cost. Re-architecting products delayed the roadmap by approximately two years, elevated research expenses, and forced asset write-offs. More fundamentally, the migration exchanged a flexible commercial foundry relationship for an operational dependency on a single domestic fabricator whose advanced capacity remains constrained and heavily contested by competing national priorities. That structural bottleneck continues to limit production volume.

Financial reports offer incomplete visibility into the exact cost of the foundry shift. Although Cambricon cited asset impairments as a key factor in its 2022 financial loss, disclosures did not itemize the specific write-downs for stranded TSMC wafers, cancelled mask sets, or abandoned tape-outs.[^21] This lack of detailed reporting obscures the true capital cost of the migration from baseline R&D expenses. It also highlights a historical pattern of capital commitment to silicon that proved unmarketable—a relevant precedent when evaluating the company's expanded inventory assets in subsequent years.

To sustain operations, Cambricon returned to capital markets in 2023 for a private equity placement. Weak market demand forced the company to scale back the offering from a planned RMB 2.47 billion to RMB 1.67 billion at RMB 121.1 per share.[^22] By mid-2023, Cambricon had raised approximately RMB 4.25 billion in cumulative capital across its IPO and secondary placement, absorbing nearly all of it in ongoing operations.[^21] Prior to its 2025 financial inflection, the company relied on equity dilution rather than operating cash flow to fund development—a record that contextually informs evaluations of its balance-sheet stability.

Beyond its immediate operational toll, the Entity List designation created an unexpected long-term strategic advantage. As Beijing increasingly prioritized semiconductor self-reliance as a core national security initiative, Washington's sanctions served as an implicit validation of Cambricon's strategic importance. While the company did not precipitate this geopolitical realignment, being targeted by U.S. export controls positioned it to capture government-backed enterprise demand as market preferences pivoted toward domestic silicon.

↑ Back to Top

VI. The Generative AI Explosion & The 2025 Financial Turnaround (2024–2026)

There is a version of the Cambricon story in which the company's engineers cracked a hard problem and the market rewarded them. There is another version in which two governments, acting from opposite motives, jointly manufactured a customer base. An accurate reading requires both, and the sequencing demonstrates which factor carried greater weight.

First, Washington repeatedly tightened export controls starting in 2022, progressively removing NVIDIA's competitive accelerators from the Chinese market — first the A100 and H100, followed by China-specific workarounds like the H800 and A800, and eventually successive compliance models. Each iteration left domestic buyers with legally importable hardware that was increasingly inadequate for training frontier artificial intelligence models. Simultaneously, the launch of ChatGPT ignited a generative AI boom in China, placing firms such as 字节跳动 ByteDance, 阿里巴巴 Alibaba, 百度 Baidu, and 腾讯 Tencent in an arms race for compute capacity that foreign silicon could no longer supply.

Then Beijing closed the remaining alternative. In September 2025, the Cyberspace Administration of China instructed major technology companies — including ByteDance and Alibaba — to stop purchasing NVIDIA's China-market processors and to cease testing the RTX Pro 6000D, effectively transforming a supply restriction into a state-mandated demand policy.⁶ Publicly funded data center projects were similarly retooled to require domestic chips, with rules applied retroactively to infrastructure already under construction.⁶ Purchasing foreign AI hardware shifted from a commercial procurement choice into a regulatory liability.

For a fabless vendor that had spent six years struggling to sell enterprise cloud accelerators, the impact on order volumes was immediate.

The numbers, and what they actually say. Cambricon's 2024 revenue rose 65.6% to RMB 1.174 billion, despite a net loss of RMB 452 million; however, the fourth quarter of 2024 marked the company's first profitable quarter since listing, establishing the true operational turning point.¹⁹²⁰ Thereafter, growth accelerated rapidly. First-half 2025 revenue surged more than forty-fold year on year to RMB 2.88 billion.²¹ Third-quarter revenue reached approximately RMB 1.7 billion, roughly thirteen times the previous year's figure.²² Full-year 2025 revenue closed at RMB 6.497 billion, generating a net profit of RMB 2.059 billion.²

The composition of that revenue highlights a complete structural shift. By the first half of 2026, cloud processors accounted for more than 99.9% of sales, while the edge chip business generated just RMB 877,000 — not million, but thousand.³ Meanwhile, the municipal intelligent computing center integration business, which once generated nearly two-thirds of total revenue, effectively vanished. Cambricon had transformed into a single-product enterprise selling data center accelerators, rendering every other business segment a rounding error.

Operating leverage compounded the surge. First-half 2026 revenue of RMB 5.996 billion generated a gross margin of 55.3%, flat year on year, while selling, general, administrative, and finance expenses combined totaled approximately RMB 110 million — representing just 1.9% of revenue.³ Although R&D spending grew 30% in absolute terms to roughly RMB 700 million, it fell from 18.8% to 11.7% of revenue.³ Net profit reached RMB 2.311 billion, up approximately 123%, while adjusted net profit rose roughly 137%.²³⁴ Second-quarter revenue of about RMB 3.1 billion slightly surpassed consensus estimates.²³

That 1.9% operating expense ratio warrants close inspection. It is extraordinarily low for a semiconductor design firm — NVIDIA's ratio has never approached such levels — and management attributed the efficiency to extreme customer concentration, which minimizes the need for a sales force.³ While presented as operational lean-ness, it reflects the financial signature of a vendor selling its output to a tiny circle of captive buyers. The same concentration that depresses sales overhead also concentrates customer risk.

Gross margin deserves scrutiny too. Cambricon's blended gross margin exceeded 62% during its loss-making years, when total revenue was modest and mix-dependent. In the high-volume expansion period, gross margin settled near 55% — registering at 54.99% in the first quarter of 2026 and 55.3% for the first half.⁷³ Scaling shipment volumes did not lift unit economics; instead, shifting toward high-volume cloud cards sold to hyperscalers with substantial bargaining power — combined with the cost structure of yield-constrained domestic fabrication — capped profitability. Against NVIDIA's data center gross margins in the 70% range, Cambricon's mid-50% margin indicates that volume expansion has come without pricing power.

A quality-of-earnings note. A gap exists between the headline 2025 net profit of RMB 2.059 billion and the stricter regulatory benchmark used for the STAR Market's first-profit test. Excluding non-recurring items, net profit stood at RMB 1.770 billion — a spread of roughly RMB 289 million.¹ For an issuer that historically relied on government subsidies and non-operating income, this difference measures the share of earnings derived from core silicon sales versus external support. At roughly 14% of reported net profit, the gap is not alarming and represents a far lower proportion than in prior years. Nevertheless, it remains a critical metric: any widening of this spread would signal that core operations are carrying less of the financial load than headline numbers suggest.

What the balance sheet is telling you. The most instructive metrics in the first-half 2026 financial report appeared on the balance sheet rather than the income statement. Inventory reached RMB 8.248 billion — representing approximately 45% of total assets, up 67% year on year and 83% quarter on quarter — composed of RMB 5.75 billion in raw materials and RMB 2.54 billion in work-in-process, against just RMB 85 million in finished goods.³²³ Prepayments climbed to RMB 2.914 billion, nearly quadrupling year to date.³²³ Conversely, accounts receivable dropped from RMB 670 million to RMB 220 million.³

This balance-sheet structure supports two distinct interpretations. The positive view is that finished goods cleared immediately upon production, while shrinking receivables indicated rapid cash collection from enterprise buyers — a stark contrast to the delayed payments of the municipal contract era. Enormous raw-material and prepayment balances reflected aggressive commitments to lock in scarce wafer allocation, high-bandwidth memory, and advanced packaging capacity upfront. From this perspective, management concluded that supply, rather than demand, was the binding operational constraint.

The alternative view is that holding RMB 8.2 billion in inventory represents a concentrated bet on a specific hardware generation. If customer requirements shift or a domestic rival introduces a superior architecture, inventory impairments could follow — a pattern Cambricon experienced in earlier product cycles.

Capacity is the binding constraint. Demand significantly outstripped available supply. ByteDance alone pre-ordered approximately 200,000 Siyuan 590 accelerators, whereas Cambricon's total 2025 shipment capacity was estimated at around 80,000 units.²² For 2026, the company targeted approximately 500,000 units, including up to 300,000 Siyuan 590 and next-generation Siyuan 690 cards — more than triple the roughly 142,000 units expected in 2025.²⁴²⁵ However, reported yield rates for large dies on SMIC's N+2 process node remained near 20% as of April 2026, meaning four out of five processed dies were defective.²³⁷ Furthermore, the Siyuan 690 remained in customer testing, with mass production potentially slipping into the second half of 2026.²⁵

Consequently, Cambricon's commercial position in 2026 was defined not by winning market share in open competition, but by rationing scarce silicon to enterprise customers restricted from buying foreign alternatives. Revenue became a function of foundry wafer allocations rather than competitive sales execution.

Who is actually buying, and what management now emphasises. Enterprise demand spanned multiple sectors. Internet platforms represented the largest purchasing block, complemented by state-owned telecommunications operators and municipal computing initiatives where data center builds were retooled to mandate domestic processors retroactively.⁶ Together, domestic suppliers led by Huawei and Cambricon were projected to capture roughly 56% of China's AI server chip market in 2026, up from 46% in 2025.²⁶

Management's strategic communication shifted accordingly. During its loss-making period, public messaging emphasized research investment and architectural innovation. By the first half of 2026, corporate disclosures centered on order fulfillment and delivery execution.³ This operational focus aligned with the company's binding constraint, replacing abstract architectural claims with verifiable shipment metrics.

The "U" comes off — and the market had long since moved. By the time the stock abbreviation suffix was formally removed in March 2026, public equity markets had already priced in an extraordinary valuation. On August 27, 2025, Cambricon shares hit an intraday high of RMB 1,465, briefly surpassing 贵州茅台 Kweichow Moutai as the highest-priced stock in China and trading at approximately 4,500 times trailing earnings.²⁷²⁸ Market capitalization peaked near RMB 664.3 billion.²² Recognizing the extremity of the valuation, management publicly noted that market prices might exceed underlying fundamentals.²² In October 2025, the company capitalized on investor demand by completing a private placement of nearly RMB 4 billion, led by GF Fund Management with participation from UBS AG and New China Asset Management.²²

The broader trajectory reflects a dramatic transformation: from roughly RMB 450 million in revenue at its 2020 IPO, to lingering around RMB 700 million during sanction-impacted years, reaching RMB 1.17 billion in 2024, RMB 6.5 billion in 2025, and expanding toward an annualized run-rate of RMB 12 billion in 2026. Cumulative net losses spanned every year from 2020 through 2024 before yielding over RMB 2 billion in net profit. The fundamental catalyst behind this financial reversal was not a sudden acceleration in product innovation, but a geopolitical realignment of the addressable market.

↑ Back to Top

VII. Core Business Deep Dive: Hardware Architecture & The Neuware Moat

Walk into a Chinese AI data center in 2026, and the physical product at the core of Cambricon's business is unglamorous: a PCI Express card roughly the size of a hardcover book, drawing several hundred watts, featuring a custom processor beneath a heatsink alongside stacked high-bandwidth memory. Racks of these cards connected by cabling power the infrastructure. Financially, this single card represents virtually the entire company.

Segment economics. Corporate disclosures highlight extreme revenue concentration. In the first half of 2026, the cloud product line accounted for more than 99.9% of total revenue.³ Meanwhile, the edge and terminal line—comprising MLU220 and MLU370-S chips targeted at smart vision, robotics, and industrial equipment—contributed less than RMB 1 million during the six-month period.³ Intelligent computing cluster integration, which previously generated the majority of corporate sales, no longer represents a meaningful business line, while initial IP licensing revenue has become immaterial.

Despite marketing descriptions framing Cambricon around a diversified "cloud-edge-device" portfolio, reported figures indicate a single-product enterprise. The edge business functions as an unexercised option rather than an operational growth driver, meaning any investment thesis relying on edge AI rests on a segment with minimal commercial contribution over the company's decade-long history.

Architectural trade-offs. NVIDIA's GPU architecture originated in graphics rendering. Its organizing principle—single instruction, multiple threads—functions like a flexible manufacturing floor where thousands of parallel execution units apply identical instructions across distinct data streams. This architectural versatility allows the same hardware to handle fluid dynamics, molecular modeling, ray tracing, or deep learning, establishing the foundation for CUDA's platform adoption.

Cambricon's machine learning unit (MLU) architecture operates on a contrasting design premise. Rather than managing thousands of general-purpose programmable threads, the design organizes silicon around dedicated matrix execution engines, positioning on-chip memory directly adjacent to arithmetic processing units so neural network weights and activations sit near the circuits consuming them. Because deep-learning mathematical structures are largely predictable, the compiler statically schedules data movement rather than relying on dynamic hardware scheduling at runtime. This configuration yields higher computational throughput per transistor and per watt for transformer workloads.

The tradeoff lies in architectural flexibility. When frontier artificial intelligence models adopt novel operational primitives—such as custom attention mechanisms, non-standard sparsity patterns, or new numeric precision formats—general-purpose GPUs adapt through software updates. By contrast, domain-specific accelerators often require specialized software kernels, compiler re-engineering, or revised silicon iterations. This architectural bet carries inherent risk in a market where model architectures evolve rapidly.

Competitive position and hardware constraints. The flagship 思元 590 Siyuan 590 pairs the MLU processor with 80 gigabytes of high-bandwidth memory, designed for deployments in clusters ranging from one thousand to three thousand cards, primarily serving AI inference rather than frontier-scale model training.⁷ Fabricated on SMIC's N+2 node—a 7-nanometer-class deep ultraviolet lithography process—the accelerator reflects domestic fabrication limits.²⁵ Independent evaluations benchmark the processor at approximately 80% of an NVIDIA A100's performance in specific workloads, estimating Cambricon's silicon design to be four to five years behind NVIDIA's latest architecture.⁷ Recent performance enhancements have relied on memory architecture and advanced packaging integration rather than semiconductor process advances.⁷

Two primary structural conclusions emerge from these technical parameters:

First, Cambricon does not compete on peak technical performance; its value proposition rests on operational adequacy combined with guaranteed domestic availability.

Second, memory supply chains represent a critical vulnerability. High-bandwidth memory chips are sourced from South Korean manufacturers due to the absence of commercially viable domestic alternatives, leaving a vital component of China's artificial intelligence hardware supply chain exposed to external trade restrictions.⁷

Software ecosystem and Neuware. Silicon capability remains ineffective without an accessible software toolchain. NVIDIA's primary competitive advantage stems not from raw silicon, but from two decades of accumulated CUDA developer adoption, optimized libraries, and framework integrations across industry and academia.

Cambricon's software platform, 寒武纪 Neuware, provides a comprehensive software stack spanning drivers to application frameworks.²⁹ Its core architecture mirrors NVIDIA's software environment: BANG C serves as the low-level kernel programming language compiled via CNCC; CNNL provides accelerated neural-network primitives analogous to cuDNN; CNCL handles multi-card and multi-node collective communication similar to NCCL; and CNCV accelerates computer vision algorithms.³⁰ At the framework level, Cambricon publicly maintains torch_mlu as its primary PyTorch execution backend,³¹ while offering compatibility with domestic frameworks MindSpore and PaddlePaddle, the Triton kernel language, and vLLM, a leading open-source engine for large language model inference.³⁰

Software execution versus ecosystem defensibility. The primary benchmark for Neuware's functionality lies in model adaptation timeline and scale. On April 24, 2026, Cambricon completed optimization for the 285-billion-parameter DeepSeek-V4-Flash and 1.6-trillion-parameter DeepSeek-V4-Pro models using vLLM, releasing the underlying adaptation code publicly. To support the architecture, engineering teams built a custom fused operator library, Torch-MLU-Ops, utilizing BANG C kernels.³²³³ Delivering immediate compatibility for a trillion-parameter model demonstrates technical execution and toolchain viability.

However, rapid model adaptation does not establish a proprietary software moat. Competitors including 华为昇腾 Huawei Ascend and 海光信息 Hygon achieved day-zero DeepSeek-V4 compatibility within the same timeframe.³³ In China's domestic semiconductor market, rapid optimization for major model releases has become a baseline requirement, executed through coordinated state and corporate efforts. While technical execution confirms software functionality, it does not generate developer lock-in.

Neuware remains capable of supporting hyperscale production inference, but trails CUDA substantially in developer maturity. Rather than benefiting from self-service developer adoption, deployment relies on intensive hands-on field engineering. Software developers operating across both environments report extended compiler optimization cycles and higher integration friction on domestic platforms. Consequently, Cambricon's software capability functions as a resource-intensive engineering service rather than a compounding platform ecosystem, requiring proportional headcount expansion as customer deployments grow.

Customer concentration and pricing power. Evaluating whether Neuware generates genuine switching costs requires examining customer concentration figures. Over the three most recent reported financial years, Cambricon's top five clients generated 92.4%, 94.6%, and 88.7% of total revenue, respectively.⁷ In the first half of 2025, the top five buyers accounted for 94% of sales, with the largest customer—identified by Caixin as ByteDance—contributing approximately 80%.⁷ Media reports indicate that ByteDance accounts for more than half of all cumulative hardware orders.⁷

This revenue structure mirrors the customer concentration of 2018, when Huawei accounted for nearly all company sales, albeit at a significantly larger financial scale. Operational distinctions exist: ByteDance does not operate an internal chip design unit capable of replacing vendor silicon immediately, while Alibaba and state telecommunications operators represent expanding secondary buyers during a period defined by supply constraints. Nevertheless, deriving four-fifths of revenue from a single client limits vendor pricing power, a structural dynamic reflected in Cambricon's gross margins remaining near 55%.

The evidence regarding developer switching costs remains inconclusive. While hyperscalers that have adapted inference pipelines to Neuware and committed to large-scale hardware pre-orders face immediate switching friction, broader platform lock-in remains unproven. Demonstrating a durable ecosystem advantage will require expanding sales across multiple major enterprise buyers without margin concessions, whereas continued revenue concentration alongside declining gross margins would signal persistent pricing vulnerability.

↑ Back to Top

VIII. Competitive Dynamics & China's AI Chip War

Consider an enterprise procurement meeting in 2026: an infrastructure executive at a major Chinese internet platform needs tens of thousands of inference accelerators. NVIDIA is no longer an option—partly because Washington restricts sales of high-end chips, and partly because Beijing has instructed domestic firms to stop purchasing lower-tier compliant processors.⁶ While domestic vendors number roughly a dozen, only two can deliver silicon at volume, and one of those two operates a public cloud that competes directly with the buyer's business.

That single scenario encapsulates Cambricon's entire commercial positioning.

NVIDIA: the benchmark that left the room. NVIDIA held roughly 66% of China's AI accelerator market in 2024, but that share dropped to about 40% in 2025 and is trending toward single digits in 2026.²⁶ The market share collapse was driven by regulatory intervention from both Washington and Beijing rather than market competition. Yet NVIDIA retains an asset that regulation cannot easily erase: CUDA, the programming ecosystem on which a generation of Chinese machine-learning engineers trained. If policy restrictions ever ease, that latent developer preference could reassert itself, leaving the long-term durability of Cambricon's market position uncertain.

华为昇腾 Huawei Ascend: the rival that is also the ceiling. Huawei's Ascend 910B and 910C, paired with its CANN software stack and MindSpore framework, serve as China's reference domestic platform. Huawei plans to ship roughly 600,000 units of the 910C in 2026, approximately double its 2025 output.²⁶ Structurally, Huawei offers a fully integrated model: it designs the silicon, builds the servers and network fabrics, delivers full compute clusters, operates a public cloud, and maintains one of China's largest enterprise support organizations. For state-owned enterprises or local governments seeking complete turnkey platforms, Huawei is the primary choice—especially since Cambricon does not manufacture servers or compute clusters.⁷

Beyond customer acquisition, Huawei competes with Cambricon for an even scarcer resource: advanced wafer capacity at SMIC. Both chip designers rely on the same constrained N+2 process node, where Huawei commands greater strategic priority in allocation.⁷²⁵ In a supply-constrained environment, the primary competitive struggle takes place upstream at the foundry.

The second tier. A second tier of domestic designers fills out the market. 摩尔线程 Moore Threads pursues a GPGPU architecture focused on CUDA compatibility; 壁仞科技 Biren Technology targets high-performance training; 海光信息 Hygon Information Technology pairs x86-compatible CPUs with its DCU accelerator line, benefiting from an established footprint in government and financial computing; and 天数智芯 Iluvatar CoreX alongside 燧原科技 Enflame Technology round out the field.

Collectively, these firms pose less of an immediate threat to market share than a challenge for raw materials, competing for the same scarce foundry capacity and high-bandwidth memory. Their presence underscores a common dynamic in policy-driven markets, where government support sustains more entrants than market demand alone would justify. Analysts project that domestic suppliers, led by Huawei and Cambricon, will capture about 56% of China's AI server chip market in 2026, up from roughly 46% in 2025, while foreign vendors drop from 34% to about 21%.²⁶ When including hyperscalers' proprietary in-house ASIC designs, domestic silicon is projected to command nearly 80% of the market.³⁴ Notably, both Huawei and Cambricon are included on government-approved vendor lists, whereas NVIDIA is excluded.²⁶

Counter-positioning, and its limits. Cambricon's principal strategic advantage lies in its non-compete positioning. The company operates no public cloud, sells no consumer devices, and offers no downstream enterprise applications. For hyperscalers like ByteDance and Alibaba—both of which compete directly with Huawei Cloud—purchasing chips from a neutral merchant vendor rather than a direct competitor is a clear strategic choice. This preference largely explains why Cambricon's order book is dominated by internet platforms, whereas Huawei's client base skews toward local governments, telecommunications operators, and state-owned enterprises.

This represents a classic counter-positioning strategy: Huawei cannot match Cambricon's neutrality without dismantling the vertical integration that forms its primary competitive advantage. However, this positioning has clear limits. It appeals primarily to internet companies that compete with Huawei, offers little advantage with state-backed buyers, and protects Cambricon against only one competitor. Should another independent domestic merchant supplier achieve commercial scale in silicon, Cambricon's counter-positioning advantage would diminish.

Running the frameworks. Applying Porter's Five Forces framework reveals a distinctly unbalanced competitive dynamic. Supplier power is critical: Cambricon relies on a single domestic foundry for advanced process nodes, foreign suppliers for high-bandwidth memory, and domestic packaging houses for 2.5D integration. Any bottleneck across these providers can constrain shipment volumes regardless of demand. Buyer power is also high in substance: despite limited alternative suppliers, a customer base concentrated in four or five entities—with one buyer generating four-fifths of revenue—exerts substantial bargaining leverage that caps gross margins in the mid-50% range. Threat of substitutes is moderate and rising, as Huawei's Ascend offers a viable alternative with lower software switching costs than transitioning from CUDA. Threat of new entrants remains constrained by high capital intensity and limited foundry allocation, while competitive rivalry is temporarily muted by market-wide supply shortages and will likely intensify once supply aligns with demand.

Evaluating the company against Hamilton Helmer's Seven Powers framework yields a similarly restrained assessment. Cambricon possesses a potential cornered resource in its CAS-derived patents and architecture, though intellectual property historically offers limited defense in chip hardware. It holds genuine counter-positioning against Huawei, alongside localized switching costs for clients integrated into the Neuware toolchain. However, key structural powers remain absent: there are no network effects, minimal brand power among performance-focused enterprise buyers, limited scale economies while manufacturing costs depend on external foundries, and no process power, given that Cambricon has demonstrated resilience under external shocks rather than an unmatchable operational pace.

Ultimately, Cambricon's market position rests on two core factors: it is one of only two domestic suppliers capable of delivering AI chips at scale, and it is the sole option that does not compete with its customers' core businesses. Both factors provide immediate commercial traction, but neither constitutes a durable technological moat.

↑ Back to Top

IX. Management Audit, Governance & Capital Allocation

A widely circulated photograph in Chinese financial media in the autumn of 2025 captured Chen Tianshi at age forty in an understated jacket, presenting the demeanor of an academic who found himself leading one of China's most strategically sensitive technology enterprises. Unaligned with the public persona of a typical Silicon Valley founder, Chen rarely grants media interviews. When he addressed the U.S. Entity List designation in 2023, he noted understatedly that the supply chain had been blocked and product deliveries affected to a certain extent — reflecting his characteristically restrained public communication.¹⁸

Control. Chen holds a 28.35% direct stake as the largest shareholder, while Beijing Zhongke Suanyuan Asset Management — the commercial arm of the Institute of Computing Technology — holds 15.57%.³⁵ Although the STAR Market permits dual-class voting structures, Cambricon chose a single-class framework; however, Chen additionally manages an employee partnership vehicle that expands his voting influence beyond his direct equity.³⁵ Together, the founder and the state-backed research institution command a stable controlling bloc of over 40%, ensuring tight alignment with national strategic objectives.

For public investors, this governance structure presents a double-edged reality. It eliminates control contests and ensures predictable alignment with Beijing's technology priorities. Conversely, it leaves minority shareholders with virtually no leverage to influence capital allocation, while signaling that corporate decision-making may prioritize national strategic goals over pure equity returns.

Credibility, judged by behavior over time. Cambricon's primary operational strength lies in its persistence under severe external pressure. After losing its anchor client, foundry access, and chief technology officer, the company absorbed roughly RMB 3.8 billion in net losses across five post-IPO years while maintaining its research commitment. Cumulative research and development spending reached approximately RMB 5.85 billion by early 2025, dwarfing its total revenue over the same period.³⁶ Management also displayed rare candor during the stock's dramatic rally, explicitly warning investors in late 2025 that market prices might exceed underlying business fundamentals.²²

However, three documented governance issues warrant scrutiny:

First, corporate narrative diverged from operational reality during the municipal computing center period. Management presented turnkey municipal integration projects as commercial progress while the core enterprise cloud silicon business lacked private-sector traction, obscuring the revenue concentration and non-recurring nature of those integration contracts.

Second, equity monetization by early backers highlights shifts in investor sentiment. By September 2023, six original pre-IPO institutional investors had fully liquidated their holdings, extracting roughly RMB 6.8 billion while the company remained unprofitable, prompting Chen to issue a public pledge not to sell his own shares.³⁷ In the first half of 2026, prominent retail investor Zhang Jianping — long listed among the top ten shareholders — fully exited a position estimated at RMB 10 billion.³⁸ Insiders and early investors cashing out during policy-driven market reratings remains a critical data point for evaluating long-term valuation sustainability.

Third, ongoing litigation involving former CTO Liang Jun represents a material governance overhang. Liang's lawsuit seeking RMB 4.287 billion in compensation for unvested equity incentives, alongside a parallel share repurchase action brought by the company's incentive entities, creates a substantial contingent liability.¹³¹⁴ With claimed damages exceeding double the company's 2025 net profit, the unresolved dispute highlights lingering internal friction from the 2022 executive departure.

Additionally, government grants and non-operating support functioned as a persistent feature of the financial accounts during loss-making years. While common for state-supported semiconductor firms, government subsidies provide an unreliable foundation for equity valuation. The narrowing gap between reported net profit and adjusted earnings in 2025 indicates diminishing reliance on state subsidies,¹ though the absence of granular disclosure regarding program conditions leaves investors unable to model how support might adjust if policy priorities shift.

Capital allocation: what the record actually shows. Cambricon has eschewed corporate acquisitions, concentrating capital strictly on internal research and development and, starting in 2025, working capital commitments to secure scarce wafer, memory, and packaging capacity. While logical for a fabless designer facing manufacturing bottlenecks, this capital deployment relied entirely on equity dilution rather than operating cash flow or debt financing. The company executed three successive equity offerings: a RMB 2.582 billion IPO in 2020, a 2023 private placement scaled down to RMB 1.67 billion from an intended RMB 2.47 billion, and a private placement of nearly RMB 4 billion in October 2025.⁸[^22]²² Raising equity near peak market valuations reflects opportunistic balance-sheet management, though future returns depend on converting funded capacity into delivered product.

Looking ahead, capital allocation is effectively concentrated in a single balance-sheet item: RMB 8.2 billion in inventory, dominated by raw materials and work-in-process. This represents the largest financial commitment in Cambricon's history — an unhedged bet on near-term demand and hardware architecture whose ultimate return will be determined over upcoming quarters.

Guidance discipline is about to be tested for the first time. Historically, Cambricon provided minimal forward guidance. That posture shifted with the 2026 restricted stock incentive plan, which granted 5 million restricted shares — representing roughly 0.8% of equity at RMB 750 per share — to over 85% of its 1,107 employees.³⁹ Vesting requires achieving steep annual revenue milestones: RMB 13.5 billion in 2026, RMB 27 billion in 2027, and RMB 59.5 billion in 2028, representing a cumulative total of more than RMB 100 billion over three years.⁴⁰

Reaching the 2026 benchmark requires generating approximately RMB 7.5 billion in the second half of the year, compared to RMB 5.996 billion delivered in the first half — a target dependent on the production ramp of the Siyuan 690.³ The 2028 target implies a ninefold expansion over 2025 revenue within three years, an ambition constrained more by domestic foundry yield rates and wafer allocations than market demand. While tying executive compensation to explicit public targets improves management accountability, it also intensifies incentives to accelerate shipments into a highly concentrated customer base.

Finally, the company's 2025 performance saw 1,107 employees generate RMB 6.5 billion in revenue, or nearly RMB 6 million per employee. While this suggests high operating leverage typical of a fabless semiconductor model, it primarily reflects the absence of a traditional enterprise sales force — reinforcing how tightly revenue performance remains tied to a small group of captive hyperscale buyers.

↑ Back to Top

X. Playbook: Strategic & Investing Lessons

Every business story leaves behind transferable principles, and Cambricon's are unusually clean because the company ran the same experiment twice with opposite results.

A component supplier's moat is measured by the customer's incentive to build it themselves. The most durable version of the Huawei lesson is not "don't depend on one customer" — that is a truism. It is that the strength of a supplier's position is inversely proportional to how strategically important its component is to a customer capable of internalizing it. Cambricon's NPU was too important to Huawei to be safely outsourced. A supplier of something genuinely peripheral might have survived indefinitely; a supplier of the marketing centerpiece was always going to be replaced. The corollary for investors examining any component or IP business: ask what fraction of the customer's differentiation runs through the product, and then ask how many engineers the customer employs. High marks on both counts represent a warning rather than a moat.

Geopolitics can substitute for a decade of go-to-market — and the substitution is rented, not owned. Cambricon spent from 2019 to 2024 unable to sell cloud accelerators to private Chinese enterprises, not because the product was unusable, but because a superior alternative was legally available. When that ceased to be true, the customer acquisition problem that had defeated the company for five years dissolved in about eighteen months. The generalizable principle is that in industries where a regulator can remove an incumbent, traditional competitive analysis is subordinate to policy analysis. The uncomfortable corollary is symmetry: a demand base created by policy can be altered by policy, and a company whose growth is a second-order effect of an export-control regime remains exposed to changes in that regime it can neither hedge nor control.

In concentrated markets, neutrality is a product feature. Cambricon's most durable strategic asset is a negative — the businesses it has chosen not to enter. Not operating a public cloud is worth real money when prospective clients compete directly with the only other viable domestic supplier's cloud. This pattern recurs whenever an industry consolidates around vertically integrated giants: contract manufacturers, merchant foundries, white-label suppliers, and neutral infrastructure providers all capture customers whose primary consideration is avoiding the funding of a competitor. It is a genuine strategic advantage. However, it is an advantage that evaporates if another credible neutral supplier emerges, meaning it should be evaluated as a temporary structural benefit rather than a permanent moat.

Certification is not commercialization, and adaptation is not adoption. Cambricon's trajectory provides a compact case study in the gap between technical achievement and repeatable revenue. Best-paper awards at ASPLOS and MICRO in 2014 yielded no revenue for three years. Integration across tens of millions of Huawei smartphones produced revenue that quickly vanished. Winning municipal computing center tenders generated revenue that failed to repeat. Day-zero support for a frontier model in 2026 demonstrated impressive engineering capability, but three domestic competitors accomplished the same milestone simultaneously. The pattern offers a consistent rule for evaluating the company: a milestone announcement signals technical capability, but provides no evidence of durable economics until it translates into repeat orders from buyers with alternative choices.

A final lesson, on reading a turnaround. The 2025 financial results — revenue surging more than fivefold to RMB 6.5 billion alongside a swing from net losses to over RMB 2 billion in net profit — represent the kind of headline figures that risk ending analysis rather than starting it. Yet segment disclosures, flat gross margins near 55%, a 1.9% operating expense ratio, and an RMB 8.2 billion inventory balance all describe the underlying reality far more precisely: a supply-constrained, single-product company rationing scarce output to a queue. Analytical rigor requires refusing to let a dramatic income statement mask the mechanism that produced it. Cambricon's financial reversal is legible, and it remains primarily the result of external geopolitical shifts rather than internal execution alone.

↑ Back to Top

XI. Investment Thesis: Bull vs Bear Case & Key KPIs

Stripping away market sentiment leaves a single core question: has Cambricon built a durable position in the world's second-largest artificial intelligence compute market, or has policy temporarily granted it one?

The bull case. The positive investment thesis does not require Cambricon to outperform NVIDIA in head-to-head technical benchmarks; framing the opportunity in those terms misinterprets its commercial positioning. Instead, the case rests on four testable propositions:

First, China's domestic AI compute market is expanding rapidly and remains structurally protected from foreign suppliers. Market projections indicate that domestic vendors will capture roughly 56% of China's AI server chip market in 2026, up from about 46% in 2025, while the share held by foreign suppliers drops toward 21% — a figure that reaches nearly 80% domestic control when including hyperscalers' custom in-house accelerators.²⁶³⁴ U.S. export controls and Chinese self-sufficiency mandates reinforce this partition from opposite directions.

Second, within this protected environment, only two domestic merchant vendors currently deliver cloud AI silicon at commercial scale. Cambricon is the sole provider that does not operate a competing public cloud or consumer ecosystem, giving it a structural neutrality advantage among internet platforms — the market's largest and fastest-growing buyers.

Third, the business exhibits substantial operating leverage. Operating expenses fell to just 1.9% of revenue in the first half of 2026 while research expenditures declined from 18.8% to 11.7% of sales, allowing incremental gross profit to flow directly to net income once fixed costs are absorbed.³

Fourth, commercial volume is driven primarily by model inference rather than model training. Because inference workloads expand alongside end-user adoption rather than fixed research budgets, Cambricon's hardware architecture can capture high-volume deployment growth without needing to close the architectural gap with NVIDIA's frontier training platforms.

The bear case. The skeptical thesis does not question reported revenue or earnings, which reflect verified customer deliveries. Instead, it highlights that almost every catalyst driving the turnaround depends on external factors beyond management's control:

First, manufacturing bottlenecks present an immediate operational boundary. Cambricon relies on SMIC's N+2 process node for flagship processors, foreign suppliers for high-bandwidth memory, and domestic packaging facilities for 2.5D integration.⁷²⁵ With large-die yield rates reported around 20% in April 2026,²³ meeting management's 2026 target of approximately 500,000 total units — including up to 300,000 flagship processors, up from roughly 142,000 units in 2025 — depends on constrained foundry capacity where competitor Huawei commands greater allocation priority.²⁴²⁵ Consequently, long-term incentive targets reflect manufacturing access rather than pure sales execution.

Second, Huawei represents a formidable, vertically integrated competitor featuring larger research budgets, turnkey cluster manufacturing, an extensive field-engineering network, and broader software ecosystem backing through CANN and MindSpore.⁷ While Cambricon's neutrality protects its footprint among commercial cloud providers, it offers minimal defense across government, telecommunications, and state-owned enterprise accounts, which constitute the majority of total procurement clients.

Third, severe customer concentration limits vendor pricing power. With top five clients accounting for 89% to 95% of total sales across recent years and a single buyer generating roughly 80% of revenue, Cambricon functions with minimal bargaining leverage — a structural dynamic reflected in gross margins remaining flat in the mid-50% range despite surging shipment volumes.⁷

Fourth, regulatory demand creation carries bilateral policy risk. Present order volumes stem from trade restrictions that bar buyers from purchasing foreign hardware. Any regulatory shift permitting foreign access would re-introduce preferred commercial alternatives, whereas stricter export limits on memory components or fabrication tools would further constrain domestic chip production.

Finally, governance and balance-sheet risks warrant scrutiny: - Holding RMB 8.2 billion in inventory — representing 45% of total assets — exposes the balance sheet to impairment risks if raw materials committed to current processors become obsolete before next-generation chips complete customer qualification.³²³ - The unresolved RMB 4.287 billion lawsuit brought by a former chief technology officer creates a contingent liability exceeding double 2025 net profit.¹³ - Raising nearly RMB 4 billion in secondary equity in October 2025 near peak share valuations aligns with historical capital dilution patterns, while employee incentive structures reward top-line revenue volume rather than capital efficiency.²²³⁹

Evaluating the investment thesis. Historical disclosures contradict claims of a durable technological moat: Cambricon has not demonstrated sustained pricing power or customer lock-in. However, the evidence supports a narrower thesis: the company holds a scarce, policy-protected supply position in an expanding market, combined with strong operating leverage and a genuine neutrality advantage over its primary domestic rival. Evaluating its long-term valuation requires assessing the stability of these external regulatory and supply conditions rather than assuming an inherent architectural advantage.

Key performance indicators. Assessing Cambricon's trajectory requires monitoring three critical metrics over time:

Cloud product gross margin. Cloud processor gross margin serves as the direct measure of vendor pricing power versus captive buyer leverage. Margin has remained flat near 55% across a fivefold expansion in revenue.³⁷ Margin expansion as shipment volumes scale would signal strengthening market position and software lock-in, whereas declining margins alongside volume growth would confirm that hyperscale buyers retain pricing control.

Customer concentration metrics. Annual disclosures of top five clients and single-largest buyer concentration test the durability of developer switching costs. Broadening the revenue base across additional hyperscalers without margin concessions would validate commercial demand, while continued concentration would indicate persistent client dependence.

Inventory and prepayment turnover. Converting RMB 8.2 billion in inventory and RMB 2.9 billion in prepayments into delivered revenue tests management's working-capital strategy.³ Rapid conversion into cash sales with minimal receivables growth would confirm that foundry allocation was the primary growth bottleneck, whereas inventory accumulation outpacing sales or resulting in asset write-downs would indicate overcommitment to legacy product generations.

These metrics cannot be evaluated from headline earnings figures alone; their resolution will emerge from disclosures in periodic financial filings.