On the evening of August 12, China’s “King of Cold” Cambricon answered investors’ questions on an interactive platform and said,The company is concerned about the information spread online about the company's large number of carrier board orders from a certain manufacturer, revenue forecasts, new product status, sample delivery, potential customers, supply chain and other related information. These are all false information that misleads the market.
The company said that investors are requested to improve their ability to discern information and not to disseminate or accept false information from unknown or unverified sources. The company's announcement on the official website of the Shanghai Stock Exchange or the information disclosed on the company's official website shall prevail. In addition, the company will reserve the right to pursue legal liability of relevant personnel for any fabrication, dissemination and dissemination of false information.
This response was related to the inexplicable surge at noon that day. On August 12, the A-share semiconductor sector was in full swing. Cambrian ushered in a 20% daily limit, with a price limit of 848.88 yuan per share, and the stock price hit a record high. The market value once jumped to 355.1 billion yuan, and the whole-day turnover soared to 15.1 billion yuan.
The key driver of the recent increase in market sentiment has been widely targeted by the industry as rumors of the release of the DeepSeek-R2 model. It was previously reported that the model may be officially released from August 15th to 30th, and its reasoning capabilities can compete head-on with GPT-5. Subsequently, details circulated in the market that "a domestic large-scale model team has completed R2 model testing on the Cambrian platform, and its inference efficiency is better than NVIDIA H20", which became the core fuse that directly ignited the upward trend in stock prices.
According to Tencent Technology’s news today, an insider close to DeepSeek responded that the online rumor that DeepSeek’s next-generation large model, DeepSeek-R2, is about to be released is untrue, and that there is no release schedule for the model in August.
However, ifeng.com Technology learned from people familiar with the matter that"DeepSeek will never respond or disclose any release time information. The model will be released once it is completed." This also makes the recent release of DeepSeek's new model full of more uncertainty.
On August 13, after the Cambrian opened down 2%, it rose again at noon. As of press time, the total market value has exceeded 359 billion yuan.
Where is the real growth trajectory of the Cambrian?
Even if the rumors are not true this time, Cambrian's real performance turnaround is its capital core that deserves more attention.
Founded in 2016, Cambrian is the world's leading AI chip design company. At the beginning of its establishment, Cambricon relied on its short-term cooperation with Huawei to integrate the AI processor into the Kirin chip, and became famous in one fell swoop, becoming one of the creators of the world's first AI mobile chip. However, the limitations of the chip IP licensing business and the termination of cooperation with Huawei quickly caused Cambrian to face the dilemma of over-reliance on a single major customer. Coupled with high R&D investment, the company fell into a quagmire of continuous losses.
Until 2024, Cambrian has achieved a total operating income of 1.174 billion yuan, a year-on-year increase of 65.6%, and the loss has narrowed to 452 million yuan, a significant decrease from 848 million yuan in the same period in 2023. It is particularly worth mentioning that the fourth quarter ushered in the first single-quarter profit, which became an inflection point for performance to get out of the quagmire of losses.
Entering the first quarter of 2025, the company has completed a reversal: revenue reached 1.111 billion yuan, a year-on-year increase of 4230.2%, and net profit attributable to the parent company turned to 355 million yuan.
The support behind these financial figures comes from: First, Cambrian is ushering in an explosion in the field of cloud product lines, with related businesses accounting for more than 99% of revenue in 2024, of which cloud smart chip and board business has doubled. Second, the company is accelerating stock preparation and production expansion. As of the end of Q1 2025, inventory has climbed to 2.76 billion yuan from about 1.77 billion yuan at the beginning of the year, and prepayments have also risen to 973 million yuan, which means that the pace of its market orders is accelerating. Third, R&D intensity will not decrease. In 2024, R&D investment will exceed 1.2 billion yuan, accounting for 91% of revenue. In the first quarter of 2025, it will continue to invest 235 million yuan to maintain high-intensity technology accumulation.
However, risks also exist. From 2024 to Q1 of 2025, Cambrian's book cash continued to drop from 1.97 billion yuan at the beginning of the year to 638 million yuan, operating cash flow is still negative, and the capital chain is under obvious pressure. In addition, the company's customers are highly concentrated, with the top five customers accounting for nearly 95% of revenue, and a single customer accounting for 79%.
Is it a bubble or a prelude to an explosion?
This is not the first time that the capital market has been stirred by AI hot spots, nor is it the first time that the Cambrian has been involved in performance myths. Since ChatGPT became popular, the AI chip company founded in 2016 has been repeatedly tied to the narrative of "replacement of domestic computing power".
Especially in the context of Nvidia's high-end chip exports being restricted and H20 performance shrinking, any rumors about domestic chip performance breakthroughs will quickly trigger financial excitement in the secondary market.
This time it was rumored to be bound to DeepSeek-R2, which hit investors’ emotional buttons. DeepSeek's R1 series of large inference models have already triggered a collective surge in computing power stocks in the middle of the year, and the news that R2's "inference performance has doubled and the number of parameters is approaching the limit of Nvidia's flagship GPU" is tantamount to adding another fire to the capital market. Rumor details have even spread to "a large domestic model team has run through inference on the Cambrian MLU390, and the speed is 20% faster than H20." This kind of unconfirmed specific data is easier to believe because it sounds like the truth.
The rumors are not entirely unfounded. Chinese AI chip manufacturers have indeed found some entry points in inference scenarios, such as focusing on power consumption, cost performance, and optimization of specific models. Some practitioners related to AI infra have previously told Phoenix Technology that domestic chips perform relatively well in reasoning and have full cost-effectiveness advantages. However, from R&D to ecology, from software compatibility to developer habits, there is still a double gap between technology and ecology in order to truly replace Nvidia's high-end GPUs. Although Cambrian has accumulated products such as the Siyuan series, it will take several years or even longer for the industry to climb before it can fully shake the status of H100 and H20.