GuShiio.com Wen| Alphabet List, Author| Ma Shuye, Editor| Zhao Jinjie
Almost at the same time as Liang Wenfeng sat next to Ma Huateng, multiple Tencent products began to embrace DeepSeek.
Recently, according to reports on the emergence of intelligence, Tencent completed another internal product and team adjustment after core apps such as WeChat connected to DeepSeek, officially transferring C-end applications such as QQ browser, Sogou input method, and ima from PCG (Platform and Content Business Group) to CSIG (Cloud and Smart Industry Business Group).
At this point, almost all AI-side products, including Tencent Yuanbao, have gathered in Tencent’s CSIG, which is regarded by the outside world as an important signal for Tencent to increase its AI To C. Prior to this, Tencent was once the one among the major domestic technology companies that was least anxious on AI To C.
However, under the repeated critical attacks of DeepSeek, many major technology companies, including Tencent, had to adjust their strategies to adapt to the new changes in the big model industry.
In the first stage of meeting DeepSeek’s pressure, major technology companies have chosen to deploy DeepSeek on their own cloud services, and each of them will first ride the wave of popularity before talking about it.Today, high-pressure technology companies are entering the second stage of facing DeepSeek’s pressure, that is, in addition to riding on the heat, they truly integrate the changes brought by DeepSeek into their own products and organizations.
Along with DeepSeek’s stressful moment, there is also a new consensus reached by major technology companies on predicting the future. In the development of the big model in the past two years, Baidu has been the most active in C-side applications. Ali did not choose to pay attention to C-side applications until December last year. Tencent once maintained a wait-and-see attitude.
Now, DeepSeek’s catfish effect has enabled major technology companies to reach an unprecedented agreement on AI to C.
A product manager of a business group in Tencent toldAlphabet List (ID: wujicaijing), last year, the business group and AI products were still two distinct lines. By the beginning of the year, the transformation AI product manager had become the new OKR.
Even Byte, who seems to be disinterested in embracing DeepSeek, a person close to Byte told Alphabet thatMany internal businesses have also been connected to DeepSeek, but they have not been opened to the public yet. At the same time, the Byte Big Model team is also working overtime to study an inference model that can surpass DeepSeek-R1. rdquo;
Regarding the above news, the Alphabet List sent it to ByteDance for verification. As of press time, ByteDance had not responded.
When innovation no longer follows the old logic of “resource-driven”, the storm of technological innovation set off by DeepSeek is causing major technology companies to unexpectedly join a new AI stress test.
01
Big technology companies have finally realized DeepSeek’s pain.
On February 15, WeChat started testing AI search internally and announced its access to DeepSeek-R1. In just one day, Tencent realized the same pressure of DeepSeek being unable to serve the needs of more users. On February 16, a crazy influx of users made AI search frequently show that the server was busy. Please try again later. Tencent had to urgently call Tencent Yuanbao resources to support WeChat AI search.
Legend: Tencent feeling traffic pressureSource: Screenshot of the Alphabet List
Previously, major technology companies rushed to connect the full-blooded version of DeepSeek-R1 to their own cloud services. However, because these services are more for professional developers and the latter’s limited user base, the major companies were unable to experience DeepSeek’s traffic in the first place. The pain in front of it.
As DeepSeek has been connected to more of its main C-end products by major manufacturers, for example, Tencent has begun to connect DeepSeek to QQ Music, Tencent Maps, and Tencent Documents in addition to WeChat. In addition to search, Baidu has also begun to connect DeepSeek to Baidu Maps, Xiaodu, etc. Predictably,As the demand for users to call DeepSeek increases, the pain of busy services that major manufacturers will face may be further exacerbated.
However, in the face of DeepSeek’s steep growth curve that has attracted more than 30 million daily users in its 20 days, major technology companies that want to attract more users to stay in their apps have been put on fire. Although they feel painful, they still have to endure it.
Moreover, in the face of the user craze caused by DeepSeek,Big technology companies have also had to re-examine their AI strategies.
Tencent changed its slow attitude and transferred more to C applications, including QQ browser, Sogou input method, ima, etc., to CSIG, showing a strong signal to add AI to C.
In addition to Tencent, Ali, which had a premonition that AI applications were about to explode, rushed to split its AI application Tongyi from Alibaba Cloud and merge it into the Alibaba Intelligent Information Business Group at the end of December last year, demonstrating its new strategy of turning to AI to C.
Among the major technology companies, Baidu is the one who is most concerned about AI to C performance. After the high-profile AII in AI, Baidu has devoted almost the most resources to enter the AI application arena. Li Yanhong emphasized on many occasions that volume applications are more valuable in the era of big models, and proposed that all applications within Baidu need to be reconstructed with AI.
However, in the face of the explosive growth of DeepSeek, whether it is Baidu, Alibaba, and Tencent, theirAI to C needs to face a new question, that is, how to catch up with DeepSeek, at the very least not to be left behind too far by DeepSeek.
02
Experiencing the pain of DeepSeek, there are still more new problems to be solved.
Recently, DeepSeek was once again exposed to financing news by the media, saying that in order to meet the access needs of more users, DeepSeek hopes to expand server resources through financing.
For major technology companies that connect to DeepSeek, they also face new challenges that require more resources to meet user call needs.
Why are major technology companies backed by cloud services still unable to completely solve the problem of busy servers? This may be related to the commercialization of current large-scale model products,Including ChatGPT and DeepSeek, all C-oriented AI applications are currently losing money and are encountering growth traps, that is, the more users flock in, the more money the platform will lose.
Although Tencent does not lack cloud resources, they also need to calculate an ROI (input-output ratio) account. As of the end of 2024, WeChat’s monthly active users have exceeded 1.3 billion. Even if only 1/10 of the users are open for internal testing, it is still more than 100 million users, almost three times the current DeepSeek monthly active users. It is not surprising that DeepSeek’s service is busy on WeChat.
More importantly, without seeing a clear commercial prospect yet, Tencent is unable to make up its mind to mobilize more resources to meet users ‘DeepSeek needs.
As the number of users increases and the number of inquiries increases, the cost of maintenance only increases and does not decrease.
The unclear prospects for commercialization were once the core reason why some major technology companies had little interest in developing AI to C in the past year.
Even if DeepSeek becomes popular in the circle, there are still no new changes in the commercialization problems of C-side applications. Big technology companies that seek new traffic by connecting to DeepSeek will inevitably be caught between long-term high investment and expected returns, struggling to balance while making a lot of excitement.“”
Some big model star startups that are not as powerful as big technology companies have also made some new changes in C-side applications.
According to interface news reports, the dark side of one of the six AI dragons has recently decided to significantly reduce its product launch budget, including suspending the launch of multiple Android channels and cooperation with third-party advertising platforms, shifting more focus to technology research and development and innovation.
The person responsible for the launch of a large head model productTell the Alphabet List that for toC’s AI products, the first new stage of Paula is actually over, and in 2025, retention will become a new focus. DeepSeek squeezed out the bubble, and large model manufacturers began to explore new solutions with higher delivery efficiency.
03
As more and more products are connected to DeepSeek, while experiencing the pain of DeepSeek’s happiness, major technology companies are bound to face the dilemma of competing left and right.
For the large-scale model research and development team within the big factory, it turned from a biological son to a working person overnight. In addition to my own research and development projects, a lot of resources are also allocated to maintain channels connected to DeepSeek. same time,The deeper they embrace DeepSeek, the more major technology companies will face doubts about their self-research capabilities.
For the large-scale model self-research team within major technology companies, the most powerful way to dispel the above doubts is to launch a new model that can surpass DeepSeek as soon as possible. In this regard, Baidu has announced that it will release the Wenxin 4.5 model in the next few months, and Tencent has also begun to internally test the T1 inference model.
Illustration: Tencent’s internal test T1 reasoning modelSource: Screenshot of the Alphabet List
But the greater pressure lies in the fact that as major technology companies update their models,DeepSeek is also improving.
It took Liang Wenfeng’s team 7 months from the release of DeepSeek-V2 in May 2024 to the release of V3 in December of the same year. From V3 to the release of R1 in January 2025, the model update time was shortened to less than one month. Recently, Liang Wenfeng also led his team to update a technical paper to demonstrate DeepSeek’s thinking on the new architecture NSA (Native Sparse Attention).
According to DeepSeek’s model update speed, perhaps in the first half of 2025, Liang Wenfeng is expected to lead the team to launch a new generation of models.
Although many major technology companies have expressed their determination to persist in self-research, recentlyMusk ‘s rollover after the release of the new model Grok 3 undoubtedly reminded major manufacturers that if the model’s performance cannot be surpassed,There are too many DeepSeek, making it difficult to attract users to jump to their own products.
Before the release of Grok-3, when Musk commented on DeepSeek R1 on the live broadcast, he said that xAI would launch a better AI model and described Grok-3 as the smartest AI on earth. However, according to the actual experience of some media, which is the greater the classic 9.11 or 9.9? rdquo; When Grok-3 was tested, Grok-3 could not give the correct answer without adding any annotations.
At the same time, Grok-3 has not set off a DeepSeek-like download craze around the world.Musk’s embarrassment may also be something that domestic technology companies will soon face.
If the new models subsequently launched by major technology companies fail to significantly catch up with DeepSeek and rush to provide users with a semi-finished product like Musk, I am afraid that the pain left to the major technology companies will only be more.
