DeepSeek is only one Tencent away from AGI

The rise of DeepSeek is reshaping the global AI landscape.
Core content:
1. The attention and reactions to DeepSeek at home and abroad
2. The impact of DeepSeek on the US-China technological competition
3. The investment and operational challenges faced by DeepSeek
After DeepSeek went viral, Black Wukong producer Feng Ji immediately commented:
This is a national-level innovation.
This is true. In the past ten days, DeepSeek has subtly changed many things.
Dario Amodei, CEO of Anthropic, who is highly sought after in the AI industry, published an article strongly calling for no American chips to be exported to China in order to maintain the "unipolar world" of AI.
A few months ago, the scientist wrote an article declaring his and his company's good intentions to solve many aspects of human physical and mental health, mental illness, poverty, and peace through powerful artificial intelligence.
This matter makes people feel divisive no matter how they look at it.
Not only that, with the emergence of DeepSeek, foreign capital has begun to re-evaluate the entire Chinese assets.
Linkedin founder Reid Hoffman said on CNBC that "China is in the game ", and Wood also said on BBG TV that " we're looking more closely at China ".
In the first five trading days after the start of the year, the Hang Seng Tech Index rose by 10%. The last time the Hang Seng Tech Index rose so sharply was in September last year.
There is no doubt that DeepSeek has been pushed to the forefront and changed the timeline of Sino-US technological competition. But returning to reality, it also brings up another practical problem:
Since its establishment, DeepSeek has not received any investment. With more and more new users, huge data and computing power requirements, Liang Wenfeng, who is determined to achieve AGI (artificial general intelligence), will undoubtedly become stretched to rely solely on the power of Huanfang Quantitative.
In other words, DeepSeek needs a real strategic investor. But this is not easy because DeepSeek is "special". There are a lot of "non-consensus" about this company:
The organizational structure is bottom-up and advocates natural division of labor instead of pre-emptive division of labor. DeepSeek does not make applications, but only models, and has not even fully considered commercialization.
This places extremely high demands on strategic investors: not only must they have sufficient data and funds, but their culture must also be inclusive enough and they must not have excessive business demands on DeepSeek.
Looking at the entire Internet field, Tencent is undoubtedly the most ideal choice.
There is no need to say much about Tencent's resources. Due to the stability of its social chain, Tencent has no urgent business needs for AI, and the pursuit of an "industrial practicality" strategy is also in line with DeepSeek's idea of industrial division of labor.
DeepSeek needs Tencent, and Tencent also needs DeepSeek. Compared with Alibaba and ByteDance’s fierceness in the field of large models, Tencent has always adopted a follower attitude, and its voice is far lower than the former two.
In the investment circle, there is a classic saying: "When lightning strikes, you'd better be there." This means that when market opportunities come, you must be there and hold the chips.
DeepSeek is exactly what Tencent needs right now. Before the full explosion of AI applications, efficiency will become the key to application implementation. The significance of winning over the most successful company in China in large-scale model computing optimization is self-evident.
Four years ago, Tencent's investment in "Black Myth: Wukong" not only created the most successful investment case in the past few years, but also proved once again that Tencent is the most ideal investor for game developers, with both face and substance.
Now, a similar opportunity has appeared in DeepSeek. From any perspective, Tencent should take action on DeepSeek as soon as possible.
/ 01 /
DeepSeek needs a strategic investor
In just 20 days after its launch, its daily active users exceeded 20 million, making it the fastest growing AI application in the world.
DeepSeek has been unable to cope with the influx of new users. Whether in the official app or website, 8 out of 10 replies are "Server busy, please try again later."
To be honest, this cannot be blamed on DeepSeek. DeepSeek's goal is to achieve AGI, not cloud services, and precious computing resources must be used to explore models instead of ensuring the reasoning needs of hundreds of millions of users.
The huge user service pressure is not the only problem DeepSeek faces. On the road to AGI, DeepSeek needs more data and computing power.
Many people say that DeepSeek’s training cost is very low, and it defeated AI giants with just $6 million.
Although this story is really sexy, we must also be clear that the cost mentioned in the DeepSeek paper is only the training cost of the final version, which does not include the R&D cost, the initial hardware purchase cost, and the initial testing and iterative training cost. The so-called cost of several million is probably just a fraction of the actual training cost.
Considering that DeepSeek is more of a follower, the computing power required by a follower is much less than that of an explorer. Once DeepSeek enters the frontier of exploration, the time and manpower costs will be much higher. This is also the core reason why Altman wants to spend $500 billion on computing power infrastructure.
According to an interview with Liang Wenfeng by Anyong, all the money for DeepSeek now comes from Huanfang:
As one of our investors, Huanfang has sufficient R&D budget. In addition, it has several hundred million yuan in donation budget each year, which was previously given to charitable organizations. If necessary, some adjustments can be made .
Although Huanfang is very profitable in quantitative investment, earning several billion RMB a year, if it distributes a share to investors, the chances of it remaining in its own hands will not exceed 2 billion. For most companies, this is definitely not a small amount.
But compared with existing large-scale model players, the resources that DeepSeek can obtain by relying solely on magic cubes are very limited.
For DeepSeek, which currently only supports question-answering, image reading, and document reading, as the capabilities of large models expand to the fields of image generation, audio generation, and video generation in the future, its demand for computing power and funds will soar exponentially.
For Liang Wenfeng and DeepSeek, finding a real strategic investor has become increasingly urgent.
/ 02 /
Anti-scaling victory and DeepSeek’s financing “difficulties”
DeepSeek did consider raising funds, but ultimately failed.
At that time, Liang Wenfeng explained:
After interacting with them, I feel that many VCs have concerns about doing research. They have an exit need and hope to commercialize products as soon as possible. However, according to our idea of prioritizing research, it is difficult to obtain financing from VCs .
From the perspective of that time, investing in DeepSeek was not an easy decision. The reason was simple. There were many "non-consensus" about DeepSeek:
Looking back, although these "non-consensus" failed to bring investment to DeepSeek, they eventually became the key to DeepSeek's success.
As Steven Sinofsky, former president of Microsoft's Windows division, said, AI is currently in a scale dilemma and has completely turned into a large-scale arms race.
The companies that have joined are either tech giants, such as Google, Meta, and Microsoft, or startups that have completed large rounds of financing, such as OpenAI and Anthropic.
But looking back at the history of computer development, vertical expansion (scale up) will eventually be subverted by horizontal expansion (scale out) , and "faster and stronger" will be replaced by "small but more".
With vertical expansion, computers have evolved from the ENIAC that took up an entire room, to transistor calculators, to integrated circuit computers that can be placed on desks, and then to the microprocessors that are still in use today.
However, the penetration and quantity of more powerful computers are not as high as those of smartphones after horizontal expansion.
When it comes to AI, this vertical development idea is what large companies are good at, but it also makes them fall into the inertia of large companies. They have been making performance improvements without making a qualitative leap.
Back then, the American telecommunications giant AT&T believed that building the Internet should expand the telephone network, add communication equipment, and make the signal more stable and stronger. The reason was simple: they owned the infrastructure of the communication network.
But the facts are completely opposite to what they imagined.
The real Internet is built through inconspicuous technological innovations:
The small company Cisco at that time invented the router, Tim Berners-Lee invented protocols and coding such as HTTP and HTML, and Netscape developed the browser... These individuals and small companies did not have a lot of money and infrastructure at the time, but they built the Internet with limited resources.
In this sense, DeepSeek has escaped the difficulty of scaling and created new possibilities. Back to the present, although many of DeepSeek's "non-consensus" are being verified, it is not easy to find suitable strategic investors.
/ 03 /
Tencent and DeepSeek, a two-way choice
Looking at China's technology companies, there are not many that can provide sufficient resources to DeepSeek, except for Alibaba, ByteDance and Tencent.
It seems that Alibaba and ByteDance are a perfect match. They were the companies that invested the most in AI in the domestic Internet field last year.
Alibaba has invested in almost all of the six AI unicorns (Intelligent, Dark Side of the Moon, Baichuan Intelligence, Zero One Everything, MiniMax, and Step Star) .
Not to mention ByteDance, which has concentrated its resources on AI. According to statistics from Zheshang Securities, ByteDance's capital expenditure on AI will reach 80 billion yuan in 2024, close to the total of Baidu, Alibaba, and Tencent (about 100 billion yuan) .
Although they are rich, Alibaba, ByteDance and DeepSeek are very different in their understanding of the development of AI.
First, both Alibaba and ByteDance have a strong imprint of mobile Internet and firmly believe that great efforts can bring miracles. Alibaba invests in a scattershot manner, while ByteDance directly brings the logic of the App Factory to the AI field. DeepSeek firmly believes that the business logic of AI is different from that of the Internet.
Second, both Alibaba and ByteDance have strong business needs for AI. For Alibaba, its main e-commerce business has been impacted, so it can only work hard on AI and make up for it in cloud computing. ByteDance hopes to find a new traffic entrance in the AI era, as the moat of the content platform is too fragile. DeepSeek prefers to conduct innovative research on basic models rather than meet business needs.
Compared with the first two, Tencent is undoubtedly a more realistic and ideal choice.
Due to the stability of the social chain, Tencent has no urgent business needs for the development of AI, and adopts a follow-up attitude with great patience. Ma Huateng once made the following analogy for this strategy:
From the perspective of implementation strategy, Tencent and DeepSeek are highly complementary. DeepSeek only wants to make models, while Tencent pays more attention to scene implementation. In Tencent's own words, the core strategy of developing large models is "industrial practicality." In simple terms, it emphasizes industrial scene coverage and application implementation.
More importantly, Tencent's investment style is "Buddhist", and its investment and business remain relatively independent, which can maintain the independence of DeepSeek to the greatest extent. The success of Game Science is the best example.
DeepSeek needs Tencent, and Tencent also needs DeepSeek.
Looking back at previous technological revolutions, when technology is fully applied, the industry will quickly enter a period of expansion driven by efficiency. In this stage, efficiency will determine everything, and the industry window will be fleeting.
In 1913, Ford introduced assembly line production, which reduced the assembly time of the Model T from 12 hours to 1 hour and 33 minutes, and the price from $850 to $360. This was not just a simple price reduction promotion, but a thorough supply-side revolution.
In this supply-side revolution, only those players who followed up with large-scale production in a timely manner survived: General Motors created a full matrix product line through the acquisition of Cadillac, Chevrolet and other brands, and Chrysler completed its expansion by acquiring Dodge.
By the 1930s, the “Detroit Three” had more than 80% of the market, establishing an oligopoly. A century later, the field of AI may be experiencing a similar turning point.
In the process of AI industry implementation, a significant reduction in the cost of inference models is crucial. The reason is simple: competitive differentiation with low cost as the priority factor means that application scenarios can be found in a shorter time, which also means finding a path to return on capital expenditure.
For Tencent, which aims to seize the opportunity of AI landing, winning over the most successful company in China in large-model computing power optimization will also bring more opportunities for its subsequent application-side layout.
Four years ago, Tencent adhered to the "Three No's" principle ("no interference in operational decisions, no seizing the leading position in projects, and no seeking distribution and operation") and invested in Game Science. In the end, Tencent not only reaped the most successful investment case in the past few years, but also once again proved that it is the most ideal investor for game developers.
Now, such an opportunity is once again in front of Tencent. With the success of "Black Myth: Wukong", Tencent should take action on DeepSeek as soon as possible.