On the evening of February 3, Baidu Smart Cloud announced that the Qianfan Platform of Baidu Smart Cloud has officially launched the DeepSeek-R1 and DeepSeek-V3 models. It has introduced an ultra-low price plan and users can also enjoy a limited-time free service. Logging in to the Qianfan ModelBuilder of Baidu Smart Cloud allows for a quick experience.
Baidu Smart Cloud stated that the models accessed this time have been fully integrated into the Qianfan inference link, integrating Baidu’s exclusive content security operator to achieve enhanced model security and enterprise-level high availability guarantee. At the same time, it supports comprehensive BLS log analysis and BCM alerts, helping users build intelligent applications securely and stably.
In terms of market trends, on February 3rd, the first trading day of the Year of the Snake in the Hong Kong stock market, the three major indices initially plummeted sharply, then rebounded from their lows. By the close of the market, the Hang Seng Index was down by 0.04%, at 20,217.26 points; the Hang Seng China Enterprises Index was up by 0.03%, at 7,384.11 points; and the Hang Seng Technology Index was up by 0.29%, at 4,737.46 points. Boosted by the positive news from DeepSeeK, AI and chip stocks in the Hong Kong market performed strongly throughout the day. Among them, Alibaba’s stock rose by more than 6%, SMIC’s stock by more than 10%, and Kingsoft Cloud’s stock surged by more than 31%, setting a new historical high.
DeepSeeK, which has been unable to challenge the giants since its release of the V3 model in December 2024 and the recent introduction of the R1 model and the multimodal model Janus-Pro, has continued to break through the circle, forming the ‘DeepSeeK phenomenon’ in the global AI community and even the entire technology circle. Lex Fridman, a well-known podcast host who has interviewed AI entrepreneurs such as Musk multiple times, used a term called ‘DeepSeeK Moment’ to describe it, saying, ‘I think it will still be remembered as a key event in the history of technology in five years.’ One of the reasons for DeepSeeK’s breakout is that it uses ‘smarter’ algorithms, reducing AI training costs by nearly 60%, while achieving or even surpassing the performance of similar models. To put it simply, while others spend 100 units to train an AI model, DeepSeeK only needs 40. This ‘cost-saving approach’ directly hits the industry’s pain point – in the past, it was about who could buy more expensive chips; now, it’s about who can use chips more efficiently. DeepSeeK not only provides a new AI technology path, but more importantly, it has torn a hole in the AI narrative wall dominated by Silicon Valley and Wall Street. However, amidst pride and excitement, we must remain clear-headed. Not to mention that in terms of funding, technology, and talent, new AI startups represented by DeepSeeK are still unable to launch a comprehensive challenge to giants like OpenAI and Anthropic. Even in terms of V3 and R1, their algorithm optimization comes at a cost: when dealing with complex scenarios, their performance is significantly inferior to the large models built by throwing money at them. It’s like using a streamlined version of Photoshop for image editing – it’s enough for daily use, but it falls short in professional scenarios. Besides, in a short period, we are still not enough to shake the hardware world dominated by giants like Nvidia. Under the impetus of the ‘DeepSeeK phenomenon,’ the future of computational power is not a single curve drawn.On one hand, products like DeepSeek, with higher traffic and lower development and consumption costs, may lead to a sudden explosion in AI applications, a scenario that all professionals long for. On the other hand, as training costs decrease, more companies enter the market, and consumer applications increase exponentially, fostering a comprehensive boom in the AI ecosystem, and chip demand is expected to grow beyond expectations.
The two sides of the coin constitute the paradox of computational power. Industry insiders cited Tencent’s previously released white paper, stating that for AI Agent applications to achieve leapfrog growth or even explosion, they must pass three thresholds: scene penetration rate greater than 15%, task completion rate greater than 80%, and user trust greater than 60%. Taking trust as an example, a sampling survey by Gartner showed that 64% of people do not want to use artificial intelligence in customer service. Currently, the technical capabilities of AI Agent applications only meet simple scenarios, such as customer service and schedule management. Complex decisions like medical consultations and legal advice still have hard-to-fill gaps. The biggest application scenarios for AI are education, healthcare, and finance, but an AI doctor with a 5% misdiagnosis rate is still hard to accept. It’s like autonomous driving being safer than humans, but every accident involving autonomous driving is scrutinized and magnified. Trust in AI is only at the initial level, and it also faces challenges such as privacy protection regulations in various countries, user habits, energy constraints, technical disagreements, multi-agent collaboration, and ethical dilemmas. The industry previously estimated that it would be around 2026 before a watershed in AI trust (greater than 60%) would occur. Will the ‘DeepSeek phenomenon’ accelerate this timeline? No one can determine that now. Some say that 2025 will be the year of AI Agent applications. DeepSeek, through innovations such as heterogeneous computing architecture, mixed deployment of CPU+FPGA+ASIC, and dynamic load balancing algorithms, has increased unit computational power output by more than double. Does this mean that the technical direction it explores will break the computational power monopoly and lead to an oversupply of computational power? This concerns the current state of computational power, which is characterized by severe imbalance. Firstly, there is regional imbalance. North America, especially the United States, accounts for the largest share of global computational power, followed by China, but high-end computational power is basically concentrated in North America. Secondly, there is supply imbalance. General-purpose computational power chip manufacturers are mainly Nvidia, accounting for more than 70% of the market share, and its GPU sales are expected to reach 7 million units in 2025; the main manufacturers of ASIC chips are Broadcom and Marvell, together holding more than 60% of the market share; in cloud computing, Google, Microsoft, and Amazon account for 65% of the global market share.This is what is commonly referred to as the monopoly of computing power. Thirdly, there is an imbalance among enterprises. Microsoft, Meta, Google, Amazon, and xAI, several major players, currently accumulate computing power equivalent to approximately 3.55 million H100 units, not including AI newcomer OpenAI. Other economic entities cannot match the volume of chips they can obtain. There is also a structural mismatch in the supply and demand of computing power.
Vertically, with the popularization of multimodal applications, the growth rate of computing power demand on the inference side has exceeded that on the training side. However, computing power is still mainly distributed on the training side, and adjustments will take some time. Horizontally, a significant amount of computing power is consumed in non-core processes such as data cleaning and model debugging. Taking the computing power of large models as an example, since the second half of 2024, the computing power of large models has shifted from training to inference, with Nvidia still holding the largest share of the inference computing market. The ‘China Computing Power Development Report (2024)’ cites an IDC report stating that, as of the fourth quarter of 2023, Nvidia’s global GPU market share reached 95. 9%, while Intel and AMD together accounted for 89.2% of the CPU market. Nvidia has built a moat that combines both software and hardware by focusing on both CUDA and GPUs. Some say that DeepSeek, through architectural innovation, bypassed CUDA and broke through its moat, which is actually a misunderstanding. Many experts have read DeepSeek’s publicly (open source) papers and indicated that the underlying architecture of V3 and R1 is still based on the CUDA ecosystem. In this market situation, other regions and their manufacturers want to challenge and attempt to break the monopoly of giants and get rid of the dependence on computing power represented by Nvidia, which is mostly an optimistic outlook. Unless quantum chips achieve large-scale commercial use, which is a matter of 5 or even 10 years later. As for the oversupply of computing power, it does not exist in the short term, and overall, the supply still exceeds demand. The ‘China Computing Power Development Report (2024)’ cites data from the China Academy of Information and Communications Technology, stating that by the end of 2023, the global computing power scale increased by 40% year-on-year, but the annual performance improvement of chips represented by CPUs is less than 15%, which cannot meet the processing needs of unstructured data such as videos and images. The CCID Think Tank calculated at the beginning of 2024 that China’s demand for intelligent computing power in 2023 reached 123.6 EFLOPS, but the supply scale of intelligent computing power was only 57.9 EFLOPS, with a severe shortage. The ‘DeepSeek phenomenon’ currently seems to be increasing rather than decreasing hardware demand. For example, Nvidia’s RTX50 series graphics cards with 32G have seen a sharp rise in the past half month, with the highest price already exceeding 60,000 yuan.Part of the reason lies in consumers purchasing RTX50 graphics cards for the deployment of DeepSeek V3 and R1. Why does the anxiety of AI giants arise when DeepSeek has not yet broken through hardware limitations, and what it has achieved is innovation on the shoulders of giants? On the surface, American AI giants have sensed a crisis, not because technology has been surpassed, but because the technological path to achieving goals has changed, and startups have more options.
Just as in the era of fuel vehicles, the competition was about engines, while in the era of electric vehicles, it is about battery management technology. DeepSeek has proven that Silicon Valley’s approach of piling up hardware and data is not the only choice; efficient use of existing resources can also be effective. When DeepSeek released its cost-effective R1 model, it coincided with the announcement of a $500 billion computing infrastructure plan—Project Stargate—by tech giants like OpenAI, SoftBank, and NVIDIA. In this context, the sense of urgency among American AI giants becomes more intense. A senior AI observer told 21st Century Economic Report’s journalists that DeepSeek’s ‘efficiency revolution’ is a symbolic event in the shift of AI development from technological idealism to engineering pragmatism. It has proven that, under existing hardware and physical constraints, the marginal gains obtained through optimizing computational topology far exceed the linear growth achieved by simply increasing the number of chips. This explains the deep strategic anxiety of American AI giants—when engineering innovation capabilities begin to stand out in AI competition, Silicon Valley’s long-standing technological first-mover advantage faces reassessment, and it will also lead to a reassessment of AI narratives, which is backed by capital and capital markets. Macro trend researcher and economist David Woo recently said in an interview that over the past two years, people have been talking about the ‘exceptionalism’ of the American economy, and AI is a significant factor in shaping this argument. The market value of U.S. stocks accounts for 63% of the global capital market, and in the two years following the emergence of ChatGPT, it has increased by 10 percentage points, with the seven major tech giants accounting for 25% of the U.S. stock market value. These giants have relied on the strong advantages of AI technology to consolidate their positions, thereby indirectly consolidating the advantageous position of the U. S. capital market. Therefore, AI is tied together with American tech giants and capital markets, and it is Silicon Valley and Wall Street that have jointly dominated the global AI narrative. Now, the mysterious force from the East, DeepSeek, has challenged this AI narrative.The ongoing fervent response in the global technology and capital sectors for over ten days can indirectly explain the anxiety of tech giants and the tension in the capital markets. As of February 3rd, before the U.S. stock market opened, Nvidia’s stock price has dropped by 20% since January 24th. Of course, this does not exclude the possibility of short-term fluctuations in investor market (risk-avoidance) sentiment.
The aforementioned senior AI analyst believes that DeepSeek has created two historical values: one is the practice of open sourcing, which is essentially a shared expression of human resources and technological innovation; the other is providing a new technological path beyond the construction of computing power and data. Interviews conducted by journalists in recent days have found that DeepSeek’s aforementioned two values have essentially become a consensus within the industry. On the practical level, DeepSeek also reveals a cruel truth: when innovation enters deep waters, engineering capabilities are more important than academic breakthroughs, cost control is more critical than parameter competition, and social acceptance is more key than algorithmic precision. If we look beyond the geopolitical issue of who wins and who loses and focus solely on the industry itself, the essence of business lies in the fact that profitable companies can survive. When the capital bubble recedes, we may find that while technological strength is important, application and survival are the realities for companies at present, and a must-answer question for all AI companies. Transitioning from ‘who can spend the most money to build the largest model’ to ‘who can do the most with the least money’; the United States still holds the most advanced chips, but China has found a more cost-effective way to play – this is the AI narrative that DeepSeek is rewriting. After all, not all startups can raise $6 billion in funding like OpenAI or xAI. It is in this sense that DeepSeek’s technological approach has been emulated by a large number of AI companies globally, its open-source strategy has been praised by global research institutions, and its pricing strategy has sparked global consumer enthusiasm. High-efficiency, low-cost technological innovation, coupled with the huge traffic of global attention, has allowed DeepSeek to win a precious window of time. This is something that cannot be bought with a pile of dollars, and is envied and even envied by AI startups like OpenAI and Anthropic.
