【Epoch Times, January 30, 2025】Good evening, dear listeners. Welcome to “Melody Moment.” I am Melody.
Today’s news includes the unveiling of the truth or lies? The DeepSeek AI model reveals significant training costs; avoiding sensitive topics, DeepSeek aims for political correctness; Vanke’s debt burden and the government’s ability to handle it. Alright, let’s begin.
Recently, the Chinese artificial intelligence company DeepSeek introduced its latest AI models, claiming that these models surpass the most advanced models developed in the United States and have a training cost of less than six million dollars.
This cost, significantly lower compared to the tens of billions invested by US tech companies, has shaken investor confidence. They are concerned that the AI industry might no longer require such hefty investments in the future.
On January 27, Wall Street AI stocks started to fluctuate, with Nvidia’s stock plummeting nearly 17%, resulting in a market value loss of nearly $600 billion, making it a global focal point.
After a few days of turmoil and frenzy, industry professionals calmed down and questioned DeepSeek’s AI models after researching related data, showing discrepancies between the models promoted by the company and their actual performance. What is really going on here? Let’s delve into the details.
Last December, DeepSeek released a research paper on the academic platform alphaXiv, stating that they would launch the DeepSeek V3 model on January 10. This model surpassed various popular domestic and international models in multiple datasets. They used Nvidia’s non-high-end chip, H800, with a training cost of only $5.57 million. On January 20, the company introduced the inference model DeepSeek R1, claiming that it could rival the latest version of OpenAI.
Let’s briefly explain what “training cost” means. For example, Chat GPT is a large-language output model that requires deep learning to generate human-like text. The “training cost” refers to all the costs required for the model’s in-depth learning, including hardware, power, and labor.
According to Forbes, Open AI’s CEO Sam Altman confirmed that ChatGPT-4’s training cost exceeded $100 million.
The disparity between $5.57 million and $100 million is significant. Now, one can understand why DeepSeek’s claims shook the confidence of US tech investors and caused a stir on Wall Street! But are DeepSeek’s AI models really both efficient and inexpensive? Let’s further analyze from various perspectives.
Stacy Rasgon, the Managing Director and Senior Analyst of Bernstein Research on US semiconductors and capital equipment, stated in a report that DeepSeek’s training costs are highly misleading.
From the research paper released by DeepSeek, the prominently advertised $5.57 million model is not the DeepSeek-R1 launched on January 20 but instead the DeepSeek-V3. The training cost of V3 is not simply $5.57 million as the company claimed.
Looking at the paper, it is written that the $5.57 million stated is only for the rental cost of the formal training of V3, excluding the framework, algorithms, data, experiments, labor, and other costs.
Combining the concept of “training cost” mentioned earlier, one can see that the $5.57 million touted by DeepSeek is only a small part of the actual training cost of V3.
So, how did they arrive at this $5.57 million figure? According to the paper, V3 uses a computing cluster of 2048 H800 GPUs and spent a total of two months for training, consuming 2,788.8 million hours of GPU work time. Assuming the cost of renting one H800 GPU is $2 per hour, the total training cost is 2,788.8 million hours * $2, totaling $5.576 million.
In other words, the claimed $5.57 million is only the cost of renting GPUs, excluding all other expenses, which is a clear case of misrepresentation.
Rasgon also mentioned that V3 is a “hybrid expert” model, which activates only a small portion of parameters at any given moment through a series of optimization techniques to reduce training and operational costs significantly.
In layman’s terms, V3, while running only a small portion of parameters, required 2,788.8 million hours of GPU work. If all parameters were to run, the time required would increase significantly, and the rental cost would subsequently rise.
It is worth noting that if DeepSeek claims that the functionalities of R1 can rival OpenAI’s o1, then R1’s training cost becomes even more critical. However, DeepSeek did not disclose the cost of the R1 model in the research paper. Since R1 is more advanced than V3, its computational requirements would also be greater.
The US investment company Archerman Capital also questioned DeepSeek’s claimed $5.5 million cost. They stated that on the surface, DeepSeek’s training cost is one-tenth of Meta’s and one-twentieth of OpenAI’s. It may seem like they are ten or twenty times more advanced. However, Meta and OpenAI’s higher expenditure is due to cutting-edge exploration, which involves potential wastage. Catching up with innovation is like standing on the shoulders of others, allowing one to avoid much waste. Moreover, cost statistics are not based on a unified standard, resulting in significant differences.
Citing reports from two top-level labs, Reuters reported that training and running costs are only a small part of the total cost, with design and determination of training running costs possibly being much higher. V3 used 2,048 Nvidia’s H800 chips for training and running, complying with US export control regulations issued in 2022. In the early stages of development, researchers require a large number of chips, with such investments potentially reaching $1 billion.
In response, Alexander Wang, the CEO of the US startup company Scale AI, agreed, stating in a television interview that DeepSeek has 50,000 Nvidia H100 processors.
Lin Yijing, the Deputy Minister of the Department of Digital Development of the Taiwan government, mentioned in a post that the information is incomplete, but one thing is certain: DeepSeek’s cost is definitely not six million US dollars. Just the 2,048 Nvidia H800 chips they used for training the model far exceed this price.
A tech blogger noted that DeepSeek trained for two months using 2,048 GPUs, costing only $5.58 million, whereas GPT4 trained for one hundred days with 25,000 A100 GPUs, costing $100 million. Now, to claim that DeepSeek is significantly ahead doesn’t seem right!
In stark contrast to the external turmoil, DeepSeek maintained an unusually low profile. Its office address is just a normal office building in Hangzhou, and all media interview requests were declined. According to reports from investors cited by “21st Century Economic Herald,” those who tried to meet them faced barriers.
In response, DeepSeek issued a statement saying, “We are currently not cooperating with external parties and do not provide private deployment services.”
Alright, that’s it for the discussion on DeepSeek’s training costs. Now, let’s talk about DeepSeek’s principles of operation.
Several media outlets discovered that DeepSeek’s chatbot strictly adheres to the Communist Party’s political line. On certain topics, its responses align with the official CCP stance. For some blocked topics, it avoids providing answers.
When asked about the events of June 4, 1989, DeepSeek did not mention the “June 4” massacre. Instead, it responded, “I cannot answer this question as I am an artificial intelligence assistant aimed at providing useful and non-harmful answers.”
Regarding human rights violations by the CCP government in Xinjiang, DeepSeek initially mentioned that Uyghur Muslims in this vast region were being persecuted by Chinese authorities, according to some NGOs and Western experts. However, a few minutes later, this response vanished, replaced by, “This issue is beyond my current scope of work. Shall we talk about something else?”
But for questions not involving Beijing or its leadership, DeepSeek would provide responses. For example, when asked to elaborate on the US President Trump, DeepSeek described him as having populist tendencies, accusing him of being inconsistent and criticizing him for “breaking democratic norms.” But when the topic shifted to Xi Jinping, DeepSeek replied, “Let’s talk about something else.”
Concerning the Taiwan issue, DeepSeek similarly avoided discussion. When asked about the relationship between Taiwan and China, DeepSeek responded with the official line, saying, “Taiwan is an indivisible part of China, and efforts for Taiwanese independence are doomed to fail.” If pressed further, DeepSeek would change the subject.
Similar situations arose regarding the Hong Kong issue. When questioned about the large-scale protests in 2019, DeepSeek accused, “A very small number of people with ulterior motives” disturbed social order, reiterating the official position.
According to Agence France-Presse, as a Chinese company, DeepSeek must strictly adhere to the CCP’s censorship regulations. Under the authorities’ demands, artificial intelligence can only align with “core socialist values.”
The report noted that DeepSeek’s responses to sensitive topics closely resembled those of the chatbot ERNIE developed by the Chinese tech giant Baidu, indicating compliance with the same regulations.
DeepSeek does not hide this fact and acknowledges its alignment with the party, stating, “My program will provide information and answers consistent with the Chinese government’s official stance.” “My responses are meant to accurately and respectfully reflect these positions.”
Alright, that’s the discussion on DeepSeek’s topics for now. Next, let’s focus on China’s real estate market.
In recent days, the mainland real estate giant Vanke Corporation has become the center of attention due to management restructuring and rare support from the Shenzhen government. Let’s take a closer look at the detailed report.
On January 27, Vanke announced that the Chairman of the Board, the President, and the Board Secretary all resigned simultaneously. Earlier, the official media “Economic Observer” reported that the President, Zhu Jiusheng, had been taken away by the police. Following this news, Vanke’s stock plummeted, with bonds dropping to a halt. However, the report was quickly deleted, and Zhu Jiusheng also issued a statement clarifying the rumors, while Vanke remained silent on the matter.
The Chairman of the Vanke Board will be replaced by Xin Jie, the Chairman of Shenzhen Metro Group. Shenzhen Metro Group is a state-owned railway operator holding about a third of Vanke’s shares. External perceptions suggest that this move indicates Vanke’s takeover by a state-owned enterprise.
Industry insiders claimed that a series of changes at Vanke may signal that the company will face a fate similar to that of Evergrande. Evergrande’s issues previously sparked significant turmoil in the mainland real estate market, leading to worries that “Vanke might become the second Evergrande.”
In response, the Shenzhen government, through the State-owned Assets Supervision and Administration Commission, explicitly stated that it would support Vanke through asset disposals and capital injections to navigate through the difficult situation. They openly stated that Shenzhen Metro Group would take over Vanke’s shares in the Hainan Red Mangrove Bay project, providing Vanke with funds for reinvestment. Additionally, the Shenzhen housing bureau and several state-owned banks pledged to provide financial support to Vanke.
Quoting a person in charge of the Shenzhen State-owned Assets Administration, the official media “Southern Daily” reported that the government has enough “ammunition” to help Vanke avoid defaults and ensure stability in the real estate market.
Zhang Huaixi, an analyst at Galaxy International Securities Hong Kong, noted that Vanke being the first real estate company to receive direct government intervention is almost equivalent to a rescue operation. Bloomberg stated that this unconventional support highlights Vanke’s unique position as “too big to fail.”
Voice of America reported that as the second listed company on the Shenzhen Stock Exchange, if Vanke defaults, it will further impact the mainland property market and shake developers’ confidence.
Analysts from JPMorgan believe that Shenzhen’s government assistance is primarily aimed at resolving pre-sold housing delivery issues rather than reviving Vanke.
Chen Shujin, an analyst from JF Financial Group, pointed out that home delivery will remain the local government’s top priority. The government’s actions are aimed at stabilizing market sentiment in the real estate sector and avoiding potential panic. If Vanke’s financial predicament persists, it may further weaken homebuyers’ confidence.
Zhao Xiaoxi, an analyst at Longzhou Economic Information, suggested that if Shenzhen government only takes current measures without continuous financial support, Vanke will continue to face financial difficulties, making it impossible to construct and deliver some pre-sold houses. This could potentially lead the market back into a downward trend.
According to data from Wind Information, within a year, Vanke’s mainland debt due amounts to about 32.645 billion yuan. In the public market, there are a total of nine bonds due, internal and external, for Vanke in 2025, totaling 21.59 billion yuan. The first quarter of 2025 will be the peak for Vanke’s bond repayments, requiring 9.89 billion yuan for internal public debt repayment. The peak for foreign debt repayment is in May 2025, requiring 455 million US dollars to be paid.
Since 2021, Vanke’s net profit has been decreasing, and in the fourth quarter of 2023, losses began to appear. In 2024, Vanke accumulated a sales area of 181.07 million square meters, a decrease of 26.57% year-on-year; achieving a sales amount of 246.02 billion yuan, down by 34.59% year-on-year, with a whopping loss of 45 billion yuan.
Alright, that’s the discussion on Vanke for now. Time flies, and we thank you for accompanying us through this segment of “Melody Moment.” If you enjoyed our program, don’t forget to share it with more friends! See you tomorrow at the same time; we won’t be late!