Source | Bohu Finance (bohuFN)

The main protagonist in the final counterattack against DeepSeek has finally made its move.

On February 18 evening, Musk's xAI made its debut with Grok3. During the launch event, Musk emphasized that over 200,000 H100s were used to train Grok3, with a total training time of 200 million hours. As a result, Grok3 surpassed all currently publicly tested models in terms of scores, including DeepSeek R1.

Musk's achievement led many Twitter viewers to believe that brute force could still work wonders. However, it didn't take long for people to realize that while Grok3 scored well, its actual performance was not as impressive as anticipated. For instance, it still couldn't determine which is greater between 9.11 and 9.90, and many of its answers and suggestions contained obvious errors.

Despite Musk's later attempts to save face by stating that Grok3 improves daily and is still only a beta version, last night's demonstration further underscored DeepSeek's shadow – brute force is not the sole solution for AI, and the key to large model competition is shifting from computing power to high-quality training data.

Unlike their American counterparts who attempted to refute DeepSeek, domestic big tech companies have embraced it.

ByteDance's Volcano Engine officially announced the integration of DeepSeek, while Baidu not only shifted from closed-source to open-source but also integrated its core search business with DeepSeek.

Alibaba has deployed DeepSeek across multiple sectors, including cloud services, 1688, Tmall Genie, and international business. The move that sparked the greatest market reaction was WeChat's integration of DeepSeek, with the addition of an AI search service in the search bar. Upon the announcement, Tencent's share price rose to HK$512, its best performance since October 2021.

The current situation somehow validates Robin Li's famous quote: "When all big tech companies are optimistic, the probability of success for startups is almost zero." As the industry undergoes a significant shift, some of last year's AI darlings have faded into obscurity, but big tech companies can still leverage their own businesses to develop landing scenarios and continue to invest with their profit-making abilities, staying in the game.

01 The Disintegration of the Old World

On February 10, at the Grand Palais in Paris, France, AI scholar and Stanford University professor Fei-Fei Li delivered a speech at the "AI Action Summit" and stated, "There is no doubt that historians will surely refer to this period as 'the true first era of AI'."

One hallmark of the arrival of a new technological wave is that, as it explores uncharted territory, one never knows where the right direction lies. When everyone believed that the only path to AGI (Artificial General Intelligence) was "brute force," DeepSeek demonstrated that every direction is a potential path.

DeepSeek changes the past route in three key areas: cost, open source, and path.

Since the birth of the AI wave, all significant technological breakthroughs have occurred in the United States, so the trend in Silicon Valley can be considered the industry's trend. OpenAI believes that computing power can solve everything, so the key to competition is having enough cards.

Even now, Musk is still trying to prove the past route by using 200,000 H100s and 200 million training hours. However, while Grok3 ranks first in scores on the large model competition leaderboard, the lead is not substantial.

The example of DeepSeek proves that while computing power is important, data quality is even more crucial, and there is significant room for cost optimization.

The trend in the domestic large model industry has also changed accordingly.

Big tech companies' "FOMO" (Fear of Missing Out) sentiment towards this AI opportunity is evident. Alibaba, Tencent, and other big tech companies have invested in half or more of the "AI Six Little Dragons." Not only did they invest, but they also built their own models.

Alibaba launched the open-source large model Qwen series and the closed-source large model Tongyi Qianwen 2.5; Tencent launched the Hunyuan large model; and Baidu, which was the first to invest in AI and the first to follow OpenAI's pace, launched ERNIE Bot.

Compared to the above three, ByteDance reacted slower but invested more fiercely. According to Caijing Magazine, ByteDance's investment in large models is "unlimited." The Doubao team has a size of thousands of people, many of whom were poached from other big tech companies with ByteDance offering 50% or even double their salaries.

According to analysis by China Merchants Securities, ByteDance's capital expenditure in 2024 is approximately RMB 80 billion, close to the combined total of Baidu, Alibaba, and Tencent (approximately RMB 100 billion).

One example of ByteDance's latecomer catching up is that before the emergence of DeepSeek-R1, Doubao's daily active users had already reached the top in China.

But in just one month, without any cost, DeepSeek surpassed it to become the industry leader, achieving the fastest growth to surpass 30 million daily active users in history.

In the blink of an eye, DeepSeek fulfilled Doubao's dream.

According to LatePost, at a recent all-hands meeting at ByteDance, CEO Liang Rubo reflected that after OpenAI released the Chain-of-Thought model in September last year and became an industry hotspot, ByteDance realized significant technological changes but "didn't feel the need to immediately replicate it... Looking back now, if we had rushed to solve major problems from the start, we would have had a chance to achieve it earlier."

Baidu, which had previously adhered to the closed-source route, also announced free and open-source access.

02 The Privilege of Big Tech Companies: Joining While Catching Up

Almost without hesitation, internet giants have chosen to join DeepSeek.

It is understood that during and after the new year, many big tech companies held online conference calls, requiring the completion of DeepSeek's adaptation and integration work. On the one hand, DeepSeek is thoroughly open-source, making it very convenient to download, deploy, and use; on the other hand, as the most discussed large model currently, DeepSeek represents a huge amount of traffic, and its founder Liang Wenfeng is committed to achieving AGI and has no intention of using daily active users for publicity.

It can be said that whoever can seize this opportunity of DeepSeek has at least gained the first-mover advantage in AI to C. This is also an opportunity for this round of adjustments in big tech companies' to C businesses.

Unexpectedly yet reasonably, Tencent moved the fastest. From a product perspective, Tencent integrated DeepSeek into its core WeChat product and conducted a gray-scale test of the AI search function. Additionally, Tencent Maps, QQ Music, and Tencent Docs have already integrated DeepSeek.

From an organizational perspective, according to Intelligent Emergence, after Tencent Yuanbao (an all-round AI product) was transferred from TEG (Technology and Engineering Group) to CSIG in January this year, more products and applications, including QQ Browser, Sogou Input Method, and ima, will officially be transferred to CSIG (Cloud and Smart Industries Group).

Although being the first to integrate DeepSeek into core products does not align with Tencent's style of "daring to be the last," the reason is not hard to guess.

Previously, the industry generally believed that the reasoning ability of large models would eventually emerge, but the process might take a long time. The emergence of DeepSeek has advanced this node, and application scenarios are gradually becoming clearer.

Take search as an example. In the past, Tencent had a traffic entrance but always struggled in search. The industry consensus is that once AI appears, it will be the first to disrupt search because AI's inherent ability allows you to get the most desired answer in the shortest amount of time. Tencent's gaming and other businesses are also core scenarios for AI implementation.

Since DeepSeek currently has no intention of doing to C or external financing plans, why not use its traffic and capabilities to cultivate its own products?

Other big tech companies have also chosen to integrate DeepSeek into their businesses, but they haven't taken as big of a step. For example, Alibaba's to B product DingTalk was the first to integrate DeepSeek, while on the to C side, Alibaba considers Kuake as its core and hopes to acquire more users and traffic through its cooperation with Apple.

Big tech companies have not given up on foundational model research.

According to a report by LatePost, ByteDance re-emphasizes pragmatic romanticism. Regarding AI, Liang Rubo put forward three key goals: first, to pursue intelligence online without neglecting key technological nodes; second, to explore new interactions while strengthening scale effects. At the same time, the Seed Edge project, which explores cutting-edge research on general artificial intelligence (AGI), was established, with the project employees' evaluation cycle and goals being more relaxed than those of other employees.

Shortly after Tencent integrated DeepSeek into WeChat Search, it launched its self-developed Hunyuan T1 deep-thinking model on Tencent Yuanbao, building independent model capabilities around AI search scenarios.

The surprise of DeepSeek is not groundbreaking innovation but rather exacerbates the industry's "FOMO" sentiment. Big tech companies with resources will certainly not give up the opportunity to obtain an AGI ticket.

03 Written at the End

It is understood that the former popular AI darling Moon's Dark Side has also proactively adjusted its strategy. According to Interface News, Moon's Dark Side recently decided to significantly reduce its product placement budget, including suspending placements on multiple Android channels and cooperation with third-party advertising platforms, to focus its main efforts on enhancing model capabilities.

Among the previous AI Six Little Dragons, Stepwise Stars and MiniMax have integrated the DeepSeek-R1 model, ZeroOne Everything clearly stated that it would not develop foundational large models and instead shift to application layer development. Baichuan Intelligence launched the "AI Pediatrician" based on the Baichuan-M1 platform, while Zhipu AI chose to increase its investment in Agent intelligence.

We often say that big tech companies do not have a soil for entrepreneurship, but the fact is that after startups receive venture capital funding, they also need to consider shareholder returns, thereby having to focus on commercialization.

As for DeepSeek? It doesn't lack money.

The cover image and accompanying images are the property of their respective copyright owners. If the copyright owner believes that their work is not suitable for public browsing or should not be used for free, please contact us promptly, and this platform will immediately make corrections.