Author: 大yu; Source: X, @BTCdayu

A bird stands on one foot, its eyes turned skyward, and the picture is almost completely blank. Such a painting set a record in the art world and was sold for 62.72 million yuan in 2010. This is the "Lone Bird Picture" by Bada Shanren Zhu Da. Bada Shanren is a descendant of Zhu Quan, King of Ming Ning Xian. He became a monk after the fall of the Ming Dynasty. Many of the birds in his paintings have this attitude since then: they are different from other birds and very distinctive.

Cambrian is also a very special one among chip stocks.

I have never taken a good look at it before. It’s not that I didn’t notice it, but I was scared off by its price-to-earnings ratio. A stock with a market value of nearly 800 billion has been losing money for most of the past dozen quarters. The price-to-earnings ratio shows that it is either negative or thousands of times higher—it will take hundreds of years to earn it back?

Of course, for this kind of growth-oriented enterprise, looking at the price-to-earnings ratio can even be said to be the wrong approach, but we still have to go back to the project itself:What does Cambrian do? Where is it going?

1. From the Chinese Academy of Sciences to "the first AI chip stock"

1. Chen Brothers: You are in the arena, I am in the academy

The story of the Cambrian Period begins with two brothers from Jiangxi.

Brother Chen Yunji was born in 1983. He went to middle school at the age of 9 and was admitted to the Youth Class of the University of Science and Technology of China at the age of 14. In 2002, he entered the Institute of Computing Technology of the Chinese Academy of Sciences to study for a master's and doctoral degree and became one of the core architects of the legendary chip "Loongson".

The younger brother Chen Tianshi, born in 1985, almost copied his elder brother's path to higher education - Junior Class of the University of Science and Technology of China, PhD in Computer Software and Theory of the University of Science and Technology of China, and has been working at the Institute of Computing Technology, Chinese Academy of Sciences since 2010.

The two brothers are two years apart. Their academic trajectories and personality divisions determine what Cambrian will become later.

Chen Yunji prefers architecture and theory, and got into the bottom layer of chip design during the years when he was working on "Loongson"; Dr. Chen Tianshi works on artificial intelligence algorithms and is a person on the software side. One understands hardware, the other understands algorithms, and they are brothers. Being able to come together to make "processor chips specifically for neural networks" is a rare combination in itself.

This idea was first raised in an internal report at the Institute of Computing Technology around 2010. At that time, it was neither favored by the academic community (deep learning was not yet popular, AlexNet detonated vision in 2012, and AlphaGo in 2016), nor was it cared about by the industry (no one in the world bought this kind of chip). They persisted. In 2014, the papers published by the team, DianNao (Computer) and DaDianNao (Big Computer), won the best paper at ASPLOS and MICRO, the top computer architecture conferences. According to public reports, DianNao was the first time that a Chinese scientific research institution won the best paper at the top conference on computer systems and high-performance computing. DaDianNao was the first time in the history of MICRO that a country other than the United States won the best paper.

In 2015, the team successfully taped out the world's first deep learning dedicated processor prototype, and the company was officially established in 2016. The name is taken from the Cambrian era, the "explosion of life" in geology, which means that artificial intelligence is about to bring about an explosion of computing architecture.

After the company was established, his younger brother Chen Tianshi became CEO (holding 28.35% of the shares as of the end of 2025), while his elder brother Chen Yunji continued to stay as a researcher at the Institute of Computing Technology, Chinese Academy of Sciences. In the story of China's technology commercialization, this arrangement of "the elder brother stays in college and the younger brother goes into business" is almost unique. The usual script is that the core team all come out to work together, and the ones who are retained are non-mainstream people.

But the objective result of this arrangement is: Cambrian has had one foot firmly planted in the Chinese Academy of Sciences system from the first day, and papers and patents continue to come from that side; commercial mistakes will not damage the academic roots at the same time.

The origins of the academic school determined many of its subsequent choices. It was more radical in technology, unwilling to take the compatibility route, and had an astonishing number of patents (as of the end of 2025, a total of 2,846 applications had been filed and 1,734 had been authorized).

2. Highlight fall

2017 was the highlight of the Cambrian period.

This year Huawei released the Kirin 970 mobile phone chip, which integrated the Cambrian 1A processor IP. Kirin 970 is installed on popular mobile phones such as Mate 10, P20, and Honor 10, with hundreds of millions shipped worldwide. Cambrian became famous overnight. That year, it successfully received Series A investment from Lenovo, Alibaba, etc. (iFlytek was an earlier angel investor), and its valuation soared to US$1 billion. At that time, more than 98% of its revenue came from Huawei HiSilicon.

On August 31, 2018, Huawei released the Kirin 980 at the IFA Berlin exhibition. This chip continues to integrate Cambrian's 1H dual-core processor IP. On the surface, everything looks like business as usual, but Huawei's internal "Da Vinci Architecture" NPU is already on the way.

The real twist happened on June 21, 2019.

Huawei released the Kirin 810 mid-range chip, replacing Cambrian IP with its self-developed DaVinci architecture NPU for the first time. This is the first time that the Kirin series no longer uses Cambrian. Three months later, on September 6, the Kirin 990 flagship also switched to Da Vinci, and the entire Kirin series bid farewell to the Cambrian era. Cambrian’s prospectus shows that IP licensing business revenue fell 41.23% year-on-year in 2019 and continued to decline in the second year.

Cambrian's customer concentration problem is not new. It has been following the same trap from the first day: a single large customer feeds the company, and a single large customer suddenly withdraws. This explains why Cambrian has been losing money for the next eight years. Because after this wave of changes, the company was forced to shift from "selling IP" to "selling chips". These are two completely different business models: selling IP is asset-light and pure profit; selling chips requires R&D, tape-out, stocking, and ecology, and every step costs money.

From the peak in 2017 to the cliff-like decline of IP business in 2019, the transition period was only two years. This kind of "boiling frogs in warm water" customer migration is no less lethal to a company than directly cutting off supply.

But there is actually no morality in this matter. Most of the customers in the industry who are capable of making in-house chips will make in-house chips - including the research report on device-side AI that I wrote, which lists many self-developed chips by automobile companies, and they all have similarities.

3. A glimmer of hope

After Huawei began to shift to self-research in 2018, Cambrian accelerated its self-research of chips and successively launched Siyuan 100 (2018), Siyuan 270 (2019), and Siyuan 220 (released in 2019, large-scale shipments in 2020). It was listed on the Science and Technology Innovation Board on July 20, 2020, with an issue price of 64.39 yuan, an opening price of 250 yuan, and a market value of 100 billion, but the company was still losing money at the time.

The next three years were the most difficult period of the Cambrian Period.

On December 15, 2022 local time in the United States (December 16, Beijing time), Cambrian, along with 36 Chinese technology companies such as Yangtze River Storage and Shanghai Microelectronics, were included in the Entity List by the U.S. Department of Commerce. TSMC, which was originally responsible for its advanced process foundry, no longer serves it, and the company was forced to turn to SMIC. The entity list and subsequent supply chain switching, combined with changes in customer project rhythms and demand structure adjustments, will jointly cause a sharp decline in cloud business revenue in 2023. During the same period, the company's annual revenue was only 700-800 million, with losses of 800-1200 million. The market value went from 100 billion when it was listed in 2020 to a minimum of 20 billion in 2022, a drop of nearly 80%.

At that time, many people in the market judged it as astory stock that was going to be cool.

The turning point will occur in the second half of 2023. This year, the company released Siyuan 590, the first truly "usable" cloud training chip. Its performance is 80-90% of NVIDIA A100. At the same time, large domestic models exploded, with DeepSeek, Tongyi Qianwen, Kimi, and Doubao launched one after another, and the demand for computing power exploded. Coupled with the increasing export controls imposed by the United States on Nvidia (A100, H100 banned, H800 banned, H20 restricted), domestic customers were forced to turn to domestically produced chips, and Cambrian just caught up.

In the fourth quarter of 2024, Cambrian achieved profit in a single quarter for the first time since its listing. In 2025, the annual revenue will be 6.5 billion and the net profit attributable to the parent company will be 2.06 billion, achieving annual profit for the first time.

An academic team that came out of the Chinese Academy of Sciences was hit by the Entity List with a market value of 20 billion, and was pushed by the AI wave to a market value of nearly 800 billion.

2. Panoramic view of AI chip track

1. Two roads for AI chips

The AI chip track is divided into two main technical routes globally.

The first is GPGPU (General Purpose Graphics Processing Unit). This is the path Nvidia is taking. The GPU was originally used for graphics rendering. After NVIDIA launched CUDA in 2006, it transformed the GPU into a "universal calculator" that can do any parallel calculation. AI training and scientific computing can be done by GPU as long as they can be parallelized, and AMD is also in this group.

The advantages of this group are high versatility and mature ecology. One CUDA code can be used on all NVIDIA cards. The disadvantage is that it is not extreme enough to specialize in AI. It must support both graphics and AI, and some hardware resources will be allocated.

The second is ASIC/DSA (Application Specific Integrated Circuit/Domain Specific Architecture). To put it simply, it is a "chip specially designed for AI", cutting off all things related to graphics rendering, and using all the silicon area for matrix operations. Google's TPU is the ancestor of this school, and Huawei Ascend, Cambricon, Alibaba Pingtou Ge, and Baidu Kunlun Core are all of this school.

The advantage of ASIC is that it has better performance and energy efficiency. For a chip of the same area, the AI computing power can be much higher than that of GPGPU. The disadvantage is that it has weak versatility and the ecosystem needs to be built from scratch. You cannot take a copy of CUDA code and run it directly, you must re-adapt it.

There is no absolute advantage or disadvantage between these two paths. If the AI ​​algorithm changes drastically every year (CNN ten years ago, RNN five years ago, and now Transformer), then the flexibility of GPGPU is more important; if the AI ​​algorithm stabilizes (now Transformer dominates the world), then the performance advantage of ASIC is more obvious. At present, there are constant voices that AI algorithms are coming to an end. In addition to world models, there are also more cutting-edge ones such as nonlinear dynamics calculations, which allow AI to think like the human brain, rather than the current computer architecture. Of course, we'll just have to wait and see.

2. "Route differentiation" of domestic AI chips

Chinese players have laid out strategies on both roads. Which path each company chooses is often determined by the background of the founding team.

Several companies that take the GPGPU route, including Haiguang, Moore Thread, Muxi, Biren, Tianshu Zhixin, and Suiyuan, are basically the "exit series" of Nvidia and AMD:

• Haiguang Information: Entering the CPU market through AMD’s x86 architecture licensing, the DCU (Deep Computing Unit) is also based on AMD’s CDNA route. And "photosynthetic organization" binding backend x86 ecology. Its biggest advantage is CUDA-like compatibility, and customers will have little code changes when migrating from NVIDIA.

• Moore Thread: The founder Zhang Jianzhong is NVIDIA's "frontier official". He has served as NVIDIA's global vice president and general manager of China for 14 years, increasing NVIDIA's market share in China from 50% to 80%. In 2020, I brought a group of NVIDIA veterans to start a business. The entire company is replicating NVIDIA and making "full-featured GPUs", both for data centers and consumer-grade graphics cards.

• Muxi Shares: The founder Chen Weiliang was the head of AMD’s global GPU SoC design. Both CTOs are AMD corporate academicians. The entire team is "AMD-based" and does something similar to AMD's "general-purpose GPU".

• Biren Technology: The founder Zhang Wen is the former president of SenseTime, the head of hardware architecture comes from Huawei HiSilicon, and the team has a mixed background from NVIDIA, AMD, Qualcomm and other companies. It takes the route of "original architecture + violent computing power".

Several companies that take the ASIC/DSA route, such as Cambrian, Huawei Ascend, Alibaba Pingtouge, and Baidu Kunlun Chip, have completely different backgrounds:

• Cambrian: Originated from the Institute of Computing Technology of the Chinese Academy of Sciences, it makes pure AI chips and does not touch graphics rendering. It is an "independent third party" in this group - not attached to any major Internet manufacturers or hardware giants.

• Huawei Ascend: Huawei's internal project "Da Vinci Architecture". The biggest feature is full-stack self-research. Chips, servers, networking, operating systems, and AI frameworks (MindSpore) are all made by ourselves. This is Huawei's consistent style: if you want to do it, make the whole set.

• Alibaba Pingtou Ge: A subsidiary of Alibaba, it mainly makes chips for Alibaba’s own cloud servers and also sells takeaways. According to people close to Alibaba, the cumulative shipment volume of the new generation Zhenwu PPU in 2025 has reached hundreds of thousands of pieces, which may have exceeded the Cambrian. This matter was directly asked by investors on the Cambrian Interactive platform, and the company neither confirmed nor denied it.

• Baidu Kunlun Core: A subsidiary of Baidu, it is used for Baidu’s own search, cloud, and autonomous driving, and has also begun to be sold externally. IPO coaching has been launched on the Science and Technology Innovation Board in May 2026.

In the above-mentioned route differentiation, we can see several trends:

First, self-research by large manufacturers is the general trend. Alibaba, Byte, Baidu, and Tencent are all making their own chips. The problem is that these large manufacturers are also customers of Cambrian. It's like a supplier. Its largest customers are preparing to build their own factories. The order you get today may be transferred to the customer's internal team tomorrow. This is one of the biggest potential risks in the Cambrian, and in the long run, this risk is actually getting bigger, not smaller, as time goes by. Similar to the Qualcomm research report I wrote, one of his customers has gone from cooperating to leaving, which has created long-term pressure.

When evaluating the AI industry chain, one of my recent feelings is that we need to not only pay attention to short- and medium-term development, but also think from the perspective of first principles, whether its moat will become wider or narrower over time.

Second, GPGPU factions are almost all exodus systems from Nvidia/AMD. It is extremely difficult to build a GPGPU, and it is almost impossible for a team without more than ten years of GPU experience to do it. This means that these companies are essentially "copying Nvidia", whether it is good or bad, it is actually difficult to evaluate it from one dimension. But most explosively growing technology companies often need to grow from 0 to 1. In this regard, you can refer to the Xizhi Technology Research Report written by me. Those who are interested can pay attention to it for a long time.

Third, Cambrian chose the loneliest road. It is neither self-developed by a major manufacturer, nor does it have internal customers to back it up; nor is it CUDA-compatible, and has no ecological advantages; nor is it an NVIDIA exodus system, which has no ready-made GPU technology accumulation. It is using self-developed instruction sets and architecture to create a brand new track. The advantage of this is that the barriers are high and the academics have accumulated patents for more than ten years; the disadvantage is that the ecology is weaker.

Looking further into this lonely road, you will see thatthe self-developed instruction set is the biggest bet among all technical decisions in Cambrian.

The instruction set is equivalent to the "language" between the chip and the software. The use of "language" here is a simplified metaphor. Strictly speaking, the instruction set is only the lowest layer. There are several layers above, such as microarchitecture, compiler, operator library, and AI framework.

• NVIDIA's CUDA is a complete set of languages and ecosystems. The finished program can be used by switching to an NVIDIA chip because it is compatible;

• Haiguangzuo's AMD ROCm ecosystem is essentially "CUDA-like translation", and the code that customers moved from NVIDIA has not changed much;

• Huawei Ascend has built the entire stack from the DaVinci microarchitecture to the CANN heterogeneous computing platform to the MindSpore AI framework, but it is endorsed by Huawei's entire ecosystem.

Cambricon also follows the self-research route - self-developed Cambricon ISA instruction set, MLUarch micro-architecture and upper-layer software stack, without the ecological endorsement of major manufacturers, which means that every time it receives a customer, it must translate the model from CUDA, and then adapt the operators of the training framework (PyTorch/TensorFlow).

Regarding leading domestic models such as DeepSeek, terms such as "Day 0 support" and "quick adaptation" often appear in the market. Cambrian is also one of the domestic chip companies that is frequently mentioned. These refer to rapid adaptation at the model level. However, whether the customer is willing to move the entire training task from the production environment depends on the operator coverage, stability, and completeness of the debugging tools - these are engineering problems that require time to settle and cannot be solved at the speed of "Day 0".

The way this bet pays off is: if the AI algorithm stabilizes in the future, the performance/energy efficiency advantages of ASIC can overwhelm the flexibility of GPGPU, and the moat accumulated by the Cambrian patents for twenty years will stand. However, if there is a mainstream architecture that is obviously different from Transformer in the future (the aforementioned world model, state space model SSM, nonlinear dynamics calculation, etc.), the self-developed instruction set itself does not have to be reinvented, but the compiler, operator library, memory scheduling, and debugging tools will all have to go through the engineering test again - this bet will also be loosened.

On this road, you may win thoroughly, or you may lose miserably.

3. Cambrian’s product—Siyuan series

Cambrian's flagship product series is called "Siyuan". It is not the company's entire revenue, but the company's cloud product line (including Siyuan chips, boards, and smart machines) will contribute 99.69% of revenue in 2025, which is the absolute main force.

But Special attention: The specific parameters of Siyuan 590 and Siyuan 690 written below (computing power, memory capacity, unit price, benchmark objects, mass production time) have never been disclosed by the company in the announcement. These figures basically come from brokerage research reports, industry chain calibers and media reports. Cambrian itself also issued a special announcement in 2025 saying that "information spread online about the company's products, customers, supply, production capacity forecasts and other information has the risk of misleading the market." Therefore, what I write below is the product image based on market standards, not the company’s confirmation of facts. I read it with this awareness in mind.

Main generations of products:

• Siyuan 370 (released in 2021): 7nm process, Chiplet architecture, and inference-based. There will still be shipments in 2024, but they have already taken a back seat.

• Siyuan 590 (first disclosed/sampled in 2023, large-scale shipment in 2024): 7nm + Chiplet, 96GB HBM2e video memory, FP16 computing power of about 345 TFLOPS, according to market standards, the performance is close to 80-90% of NVIDIA A100, and the unit price is 60,000-80,000 yuan. The current main driver of the company’s revenue.

• Siyuan 690 (sent for testing in 2024, mass production in early 2026): Based on SMIC’s N+2 advanced process (the industry has disputes about its process equivalence positioning, the mainstream view is close to 7nm+, quorumthink classifies it as a 5nm node), dual-die chiplet, HBM3, FP16 computing power exceeds 700 TFLOPS and 196GB according to market standards HBM3 has an interconnection bandwidth of over 890Gbps and a unit price of approximately 135,000 yuan. This generation has more than twice the computing power of the 590 and twice the HBM capacity. The performance target is 80-85% of the market benchmark NVIDIA H100.

If you put Cambrian and Nvidia on the timeline:

• A100 is NVIDIA's product in 2020, and Siyuan 590 is aimed at NVIDIA 5-6 years ago

• H100 is a product of 2022, and Siyuan 690 is designed to compete with NVIDIA 3-4 years ago

• Nvidia’s latest Blackwell (B200) will be released in 2024, and this Cambrian is far from keeping up

Cambrian is not fighting a "flat battle" with NVIDIA, but is replacing it within a "generation gap". There is little business to be done in this matter in a normal market environment, but the current global competitive landscape has given Cambrian a good opportunity for development.

4. Timing

In October 2022, the U.S. Department of Commerce officially issued regulations on chip export controls to China (Nvidia had been notified in August to suspend A100/H100 exports to China). Subsequently, the controls were tightened: A100 and H100 were banned, and the special edition "H800" was also banned. Later, a lower level "H20" was launched, and it was tightened again in April 2025.

Nvidia CEO Huang Renxun himself said: Nvidia's market share in mainland China has dropped from 95% in the early days of the Biden administration to about 50% in 2025 (55% based on IDC shipments; at the same time, Huang Renxun also mentioned during the year that in China's high-end AI chip market segment, the share has been close to zero).

What was forcibly released was the 45 percentage points between Nvidia's drop from 95% to 50% - a computing power market of approximately tens to hundreds of billions of yuan.

How much does each domestic AI chip manufacturer ship? There is no truly authoritative statistics on this matter in the Chinese market. Because Cambrian, Huawei Ascend, and Kunlun Core only disclose sales amounts in their annual reports and do not disclose shipments; the statistics from the Ministry of Industry and Information Technology and the Semiconductor Industry Association do not go down to a single company; the rest can only be estimated by third-party research institutions such as IDC, Gartner, Qunzhi Consulting, and Counterpoint. However, they are all private commercial companies and rely on interviews with manufacturers and channel providers to estimate shipments. Each company has different calibers and different numbers, and sometimes the differences are very large.

The most commonly quoted by the media is IDC. The following table, IDC’s 2025 China AI accelerator card caliber, is the most public version currently reported by many media. The tableis based on shipments, not sales amount. The total number is about 4 million. When reading this table, you should pay attention to the fact that whether the statistical caliber includes self-used IDC is not stated, and it cannot be found in the public caliber. This may have a greater impact on the true computing power scale of Kunlun Core (a large amount of self-used within Baidu), Pingtou Ge (self-used by Alibaba Cloud), and Huawei Shengteng (self-used by Huawei Cloud).

According to IDC, the localization rate will reach 41% in 2025 (1.65 million / 4 million), a significant increase from about 30% in 2024.

There are some outliers in the table, such as both Kunlun core and Cambrian are listed as tied for third place with 116,000 pieces. However, referring to IDC's own figures for 2024, Kunlun core has 69,000 pieces and Cambrian has 26,000 pieces. Kunlun core is 2.7 times that of Cambrian; in 2025, Baidu announced that Kunlun core's third generation Wanka cluster will be lit and plans to expand to 3 Wanka; in China Mobile's centralized procurement of inference AI servers from 2025 to 2026, servers based on Kunlun core won 70%, 70%, and 100% of the three bid packages respectively, ranking first. Cambrian is not among the major suppliers.

In addition, Cambrian's "position" in the A-share market does not entirely come from market share, but also from scarcity. It is the earliest, most mature, and first profitable pure AI chip design target in the A-share market. Haiguang Information prefers CPU+DCU dual wheels. Among the "Four Little Dragons of Domestic GPUs", Moore Thread and Muxi Technology will be listed on the Science and Technology Innovation Board at the end of 2025, and Biren Technology and Tianshu Zhixin will be listed on the Hong Kong stock market in early 2026. However, these four are still in the early stages of loss or profit verification.

Organizations want to configure a "profitable pure AI chip target", but the only choice at present is still Cambrian. But this scarcity is unstable. The "Four Little Dragons of Domestic GPUs" have all been listed by the beginning of 2026. Moore Thread and Muxi are on the Science and Technology Innovation Board, Biren and Tianshu Zhixin are on the Hong Kong stock market, and Suiyuan and Kunlun Core are also on the road to IPO. In half a year, Cambrian has changed from "only" to "choose one of five", and the scarcity premium is being diluted.

5. There are counterexamples to the position of "neutral third party"

Cambrian is an independent third party that does not have "internal customers" in the camp of major domestic AI chip manufacturers.

Put the players aside:

• Huawei Shengteng: Huawei mainly uses it internally, but also sells it outside

• Alibaba Pingtou Ge: Alibaba mainly uses it internally and is expanding external sales

• Baidu Kunlun Core: Baidu mainly uses it internally and is expanding external sales

• Byte SeedChip: purely for personal use, mass production in 2026

• Cambrian: Purely a third party, without any "internal business"

This matter can be viewed in two parts: the bad side is that others have their own business as a stable source of orders for chips, and Cambrian can only find customers outside; but the good side is that neutral third parties are one of the few suppliers that can be accepted by multiple major manufacturers at the same time, because there is competition among major Internet manufacturers. Byte is not willing to buy Alibaba's chips, and Alibaba is not willing to purchase Baidu's solution - in our last in-depth article about TSMC, we mentioned that TSMC's neutrality is a key factor in achieving its dominance.

It does not belong to any one family, so it can serve all families at the same time.

In person, there is a difference in the long run. For example, NVIDIA is neutral. Although everyone buys NVIDIA, in the end, all the major Internet companies that have the ability want to develop their own. For example, Google is developing TPU? Because the core driving force of self-research is not "getting rid of dependence", but cost optimization and scenario customization. Therefore, "neutrality" is a short-term window period asset. It can slow down the speed of procurement transfers, but it cannot prevent large manufacturers from self-research. Whether Cambrian can build ecological barriers thick enough before large manufacturers' self-developed chips mature is its most critical race in the next three to five years.

6. Who is the customer

Cambrian has a characteristic that gives all investors a headache: It never publicly discloses the names of its major customers.

In the company’s financial report, only the largest customers were disclosed in 2021 and 2022-state-owned intelligent computing center project companies in Kunshan and Nanjing, Jiangsu respectively. In other years, there are vague expressions such as "computing infrastructure investment entities with local state-owned assets background" or "leading companies in operators, finance, and Internet fields."

But the customer concentration figures are clearly disclosed:

Cambrian has never really diversified its customers - the top five customers accounted for 94.6% in 2024 and dropped to 88.7% in 2025, which is still high for a chip design company that mainly focuses on server and cloud customers.

Changes have occurred in 2025: the proportion of the largest customer dropped from 79% to just over 26%, and the top five customers only dropped by 6 percentage points, which means that the revenue is not spread across dozens of customers, but from "one very large customer" to "several still large customers". Among the top five customers disclosed in the 2025 annual report, the third largest is "long-term partner", and the other four are all new in 2025. The top five sales are 1.703 billion, 1.401 billion, 1.237 billion (estimated), 764 million, and 655 million respectively.

What needs to be explained separately is the 85.31% in the 2025 interim report. This number is not in terms of sales revenue, but the proportion of the closing balance of accounts receivable and contract assets of the top five customers - the specific proportions of the five companies are 42.5%, 18.0%, 8.9%, 8.4%, and 7.6%. This is not the camp collection concentration, but the collection concentration, and the two cannot be directly compared. What this number really tells us is not that "the largest customer accounted for 42.5% of revenue in the first half of the year", but that "at the time of the interim report, money is still mainly in the hands of these five companies." It is common for a company to have diversified sales revenue and concentrated receivables to coexist, but if receivables have been concentrated in a few companies, the risk of collection must be looked at separately.

There are rumors in the market about who these big customers are. Bloomberg reported in 2025 that ByteDance is currently the largest customer, accounting for more than 50% of orders; China Mobile’s Harbin Intelligent Computing Center will be put into production in August 2024, deploying more than 18,000 AI accelerator cards (including a considerable number of Cambrian 590s according to public reports); Alibaba and Tencent also have cooperation, but the scale is unknown. None of this has been officially confirmed by Cambrian itself. In addition, Now the largest customer (market speculation is Byte) is developing its own chips. According to public reports, the team has more than 1,000 people, and mass production is planned in 2026 - This will remind people of the sweetness of the honeymoon and the decision to break up with Huawei.

In short, the company is unwilling to tell shareholders where the money comes from. It depends on how everyone interprets this signal.

3. Why the sudden outbreak in 2025

A company that had been losing money for eight years suddenly made 2 billion in 2025, and 1 billion in one quarter in 2026 - what happened?

1. Ten-year financial panorama

Cambrian's financial data in the past few years and its financial data in recent quarters are not at the same level at all.

From 2018 to 2020, the company's annual revenue increased from more than 100 million to more than 400 million; in the three years from 2021 to 2023, it really "stalled" around 700 million. Then it will start to turn upward in 2024, reaching 6.5 billion in a year in 2025, and 2.9 billion in a quarter of 2026. That $2.9 billion is roughly 4 times revenue for any full year from 2021-2023, and 2.5 times revenue for all of 2024.

2. The last five quarters

Gross profit margin remains stable between 54-56%. Cambrian did not engage in a price war during the rapid increase in volume, which is a hard indicator of high-quality growth. The scale has increased but the gross profit has not decreased, which shows that the pricing power is still there.

Revenue jumped. Q2-Q4 revenue was stable at 1.7-1.9 billion, and it seemed that growth had peaked; but Q1 suddenly jumped to 2.885 billion, +53% quarter-on-quarter. The so-called "slowdown in the second half of the year" is actually a matter of delivery pace, not a trend peak.

A rollercoaster of cash flow. Huge losses in Q1, a sharp turn to positive in Q2, negative again in Q3-Q4, and then positive again in Q1. This kind of fluctuation is a typical feature of chip companies: spending a lot of money when preparing goods, and collecting a lot of money when delivering goods. It only makes sense to look at the accumulation over a year. 2026Q1 has accumulated +834 million, and the direction is improving.

Dual signals of contract liabilities + prepayments. Contract liabilities jumped from 00.6 million (610,000) at the end of 2025 to 396 million at the end of 2026Q1. Customers have prepaid nearly 400 million, but Cambrian has not yet delivered. During the same period, advance payments rose from 745 million to 1.897 billion, +155% month-on-month. Cambrian is also paying a large amount of money to upstream foundries to lock in production capacity. The two data surged simultaneously, indicating that both customers and Cambrian themselves are making the same judgment: there will be large orders to be delivered in the next one or two quarters.

But there is another side to the same story.

The timing of inventory should be clearly distinguished. Inventory at the end of 2025 rose from 1.774 billion at the end of 2024 to 4.944 billion, +178.67% year-on-year. The company's annual report stated: "If the market environment changes in the future, the risk of the company's inventory depreciation may increase, which will have an adverse impact on the company's profitability." This risk has been partially realized in 2026Q1: an asset impairment loss of 2.46 billion, mainly due to the decline in inventory prices caused by the increase in the age of raw material warehouses. But at the same time, inventories are also being digested - The inventory at the end of 2026Q1 fell back to 4.497 billion, from 49.44 to 44.97, a decrease of 450 million from the previous quarter, and the direction is good.

The 4.9 billion figure has been quoted repeatedly over the past few months, but it belongs to late 2025, and those looking at 2026Q1 need to look at the 44.97 figure. The impairment provision of 246 million in Q1 was equivalent to nearly a quarter of the net profit of 1.013 billion in the quarter.

Trace of short-term capital allocation. In the 2026Q1 cash flow statement, "obtaining loans" and "cash paid to repay debts" are both 81.5 million yuan. At the same time, the balance of short-term borrowings in the balance sheet at the end of the period is 0, which means that the money was borrowed and repaid, which is a short-term turnover within the quarter and does not form an ongoing debt. Interest expenses for the current period rose from 158,800 yuan to 2.2911 million yuan, a relative increase of about 13 times. The absolute value is not large, but it shows that the frequency of bank credit utilization increased significantly during the quarter. The pressure to stock up puts pressure on financial expenses.

Accounts receivable are growing faster than revenue. Accounts receivable at the end of 2026Q1 was 1.219 billion, an increase of 81.79% from 671 million at the end of the previous year, and revenue increased by 159.56% year-on-year. The proportion is acceptable, but the absolute number is getting bigger.

Looking at these together: customers paid 400 million in advance, there was still 4.5 billion in inventory at the end of Q1, and 80 million short-term funds were used during the quarter. These three things point to the same judgment. The company is placing great emphasis on the delivery pace in 2026. If the volume of Siyuan 690 increases smoothly, the inventory will continue to be digested; if the volume increases slower than expected, the provision for inventory depreciation will continue to rise.

3. Four reasons for the outbreak

Reason 1: The market gap left by Nvidia.

Reason 2: Siyuan 590 has finally reached a playable level. The performance can reach 80-90% of A100, and the price is about one-third of similar products from NVIDIA. For Chinese customers, it is in a state of "enough to use, available to buy, and in stock".

Reason 3: Once the scale effect starts, profit elasticity is extremely high.芯片设计公司的成本结构有个特点：研发投入是固定大头，不会随营收线性增长。一旦营收上规模，研发占比迅速下降，利润率非线性往上跳。

2024 年寒武纪研发投入占营收的比例是 91.30%，花掉的钱比赚的还多。 2025 年全年这个比例骤降到 17.99%，2026Q1 进一步降到 11.23%。从 91% 到 18% 再到 11%，这就是规模效应启动后的典型形态，寒武纪近几年研发投入和研发费用基本一致。从季度看，研发费用率从 2025Q1 的 24.5% 降到 2026Q1 的 11.2%，营收增长 160%，研发费用只增约 18%，差距全部体现到利润上。

这就是为什么营收增长 160%、净利润增长 185%、扣非净利润增长 239%。

原因四：合同负债 + 预付款的同步好转。订单和备货同步加速，业绩的可见度比单看营收更强。

四个原因叠加，造就了"一年赚出过去十年的钱"。

这不是泡沫，是真实发生的产业变化。

4、思元 690——已经量产，但战役才刚开始

思元 690 在 2026 年初已经实现量产，接下来还有三关。

良率。中芯国际 N+2 工艺生产寒武纪这类大芯片的实际良率，市场上有几个差异巨大的口径：彭博社报道约 20%（寒武纪官方否认），第三方咨询机构群智咨询估计中芯国际 5nm 工艺 2025Q1 约 34%、2025Q4 可能提升到 40% 以上。两个数字相差一倍以上，谁更接近真实情况现在没有公开口径能验证。如果按 20% 这个偏悲观的口径，每生产五颗芯片只有一颗合格，会严重影响出货量和毛利率；按 40% 这个偏乐观的口径，影响相对可控。作为参照，台积电先进制程成熟产品良率一般 70-90%。

产能。寒武纪是 Fabless 模式，自己不生产芯片。被列入实体清单后，台积电不再为它代工先进制程，所以全部依赖中芯国际。中芯国际 N+2 先进制程的总产能，2025 年底约月产 3.5 万片晶圆，2026 年扩产到约 6 万片。其中据行业估算，华为预计锁定 1.5 万片以上，剩下的留给寒武纪、海光、摩尔线程等所有国产 AI 芯片厂商一起抢。

先进封装。思元 690 采用双 Die Chiplet+HBM3 同封装方案，封装由长电科技与通富微电承担。但 88×88mm 以上的 ABF 基板国内基本无法本土供应，需要从日本揖斐电、台湾欣兴、景硕等海外厂商进口。

思元 690 能不能放量、什么时候放量、能放出多少量，是寒武纪未来两年业绩的关键变量。所有看好寒武纪的，本质上都在押注这件事；所有看空它的，也都在质疑这件事。

四、几组数字

同行业上市公司之间的相对位置如下，供读者综合判断。

注：英伟达营收和净利润均为 FY2026 口径（财年截至 2026 年 1 月），与其他公司 2025 自然年口径不严格可比；毛利率为 FY2026 全年口径（Q4 单季高至 75%）。华为昇腾营收来自 FT/Reuters 转述华为内部口径（约 75 亿美元），Reuters 注明未独立核实，且华为不在财报中单独披露 AI 芯片业务营收。市值数据截至 2026 年 5 月 8 日，正处剧烈波动期——5 月 6 日海光单日涨 18%、市值一度超 8000 亿，与寒武纪互有高低；发表前请以当日实际为准。

把寒武纪和海光信息放在一起看：两家市值都在 7000-8000 亿区间，过去一周内反复互有高低。但 2025 年寒武纪营收是海光的 45%，净利润是海光的 81%。 摩尔线程的存在本身就是一个参考点。市值 3000 亿、营收只有 15 亿、还在亏损，但 2025 年增速 243%、2026Q1 已实现单季盈利（主要靠政府补助）。一家刚刚跨过盈亏平衡的公司市值能撑到这个量级，告诉我们当前国产 AI 芯片的估值锚不是利润，是预期。

市场的乐观，公司的现实

• 基本面好转：2025 年首次年度盈利、2026Q1 单季净利逾 10 亿。

• 产业链变好：思元 690 已量产、合同负债 +3.96 亿、预付款 +18.97 亿。

• 规模在加大：存货 Q1 末仍压 44.97 亿、Q1 计提存货跌价 2.46 亿、季度内动用过 0.815 亿短期资金。

• 客户集中风险仍在：第一大客户占比仍在 26% 以上、市场推测是字节、字节自研团队 2026 量产。

• 与华为差距大：国产阵营内排第三、与华为差距 5 倍以上。

写到这，想起那只单脚伫立、白眼朝天的鸟：它接下来是要飞、还是继续站着，交给时间吧。