Recently, Chinese artificial intelligence company DeepSeek has launched a new AI reasoning model, causing ripples across various fields including technology, finance, and policy in the United States, serving as a catalyst for the US-China AI war.
On January 20, the same day as President Trump’s inauguration, Chinese AI company DeepSeek unveiled the AI reasoning model DeepSeek-R1, claiming its capabilities are on par with OpenAI but require less computing power and lower costs.
Founder of DeepSeek, Liang Wenfeng, became the only AI leader to meet with Li Qiang, and was featured prominently on the CCTV (China Central Television) evening news program.
During media interviews last year, Liang Wenfeng stated that his core team consists of local talents rather than returnees from overseas, emphasizing that China’s AI industry should not always be a follower. This narrative resembles the official independent and autonomous storytelling.
Surprisingly, during the Chinese lunar new year, DeepSeek, once unknown two years ago, suddenly emerged as a force stirring the world and made headlines at home and abroad.
By January 26, the model had surpassed ChatGPT, becoming the number one app in the Apple App Store rankings. Billionaire investor Marc Andreessen described the R1 model as the “Sputnik moment for artificial intelligence.”
On January 27, the opening bell saw leading US AI companies facing heavy losses, with Microsoft dropping 4%, Tesla dropping 2%, and Nvidia plunging nearly 17%, causing a market value loss of nearly $600 billion, marking the largest single-day drop in history.
Currently, there are varying perspectives on the claims made by DeepSeek. However, what sets DeepSeek apart is the backing of national forces from the Chinese government.
Founder of Taiwan AI Labs, Du Yijin, stated to the Epoch Times that in the field of AI, whether open-source or closed-source, there have been significant updates every few months in recent years. However, DeepSeek stands out by not just releasing information on the DeepSeek open-source model.
Du Yijin pointed out that there is a high level of state-led influence behind DeepSeek, with a significant presence of fabricated information on platforms like Weibo and Facebook, falsely suggesting the impact of the DeepSeek model release on US tech stocks and the loss of US AI superiority.
Du Yijin highlighted that while the DeepSeek model does have distinct features, the excessive hype created by fabricated accounts, many of which are AI-generated farm accounts with minimal subscribers, is fostering a narrative of heightened competition in the US stock market.
From Du Yijin’s perspective, the portrayal of what could be a simple open-source achievement as a high-stakes national-level technological competition through the dissemination of misleading information is reminiscent of the way the Chinese government handles Russian-related issues with coordinated propaganda efforts involving internet military accounts and official media accounts.
Du Yijin stated that through its maneuvers, DeepSeek demonstrates that China does not necessarily need the highest computing power chips to develop advanced models.
One of the key aspects of interest regarding DeepSeek is its touted low cost and low computing power requirements.
DeepSeek’s original report indicates that during the pre-training phase, training DeepSeek-V3 on every trillion tokens only requires 180K H800 GPU hours. Assuming a GPU rental price of $2 per hour, the total training cost amounts to $5.576 million.
Industry experts believe that DeepSeek employs a technique called “mixture of experts,” which involves delegating tasks to domain-specific experts to alleviate the burden on chips processing all tasks simultaneously. However, this technique is not particularly innovative in itself.
Du Yijin remarked that traditional business companies also utilize a “multi-expert model” to reduce the costs of large models, which is prevalent and not groundbreaking because most AI startups lack GPU resources. The innovation with DeepSeek V3 lies in its multi-head, multi-precision mixed training, effectively lowering model training costs.
However, Du Yijin noted that DeepSeek’s origin of training data and the methods used for training are not entirely clear, raising questions about the transparency of the process.
Du Yijin explained that upon reviewing the DeepSeek source code, it became evident that the model actually utilizes OpenAI’s distillation technique with the all one model, extracting knowledge from large models and training smaller models.
He further highlighted that the OpenAI statement on January 29 indicated that they were reviewing reports of potential “inappropriate” use of their model’s output data by DeepSeek in developing their AI model through distillation, a technique known as “distillation.” He further explained that the OpenAI authorization was originally not allowed, prompting the suspension of their account.
Du Yijin added that the distillation technique offered by OpenAI is not groundbreaking either, having been proposed around 2015. The intent behind OpenAI’s distillation was to reduce the high costs of running large models with the all one model, enabling smaller models to learn from larger models at a lower cost, yielding results similar to specific areas of the larger model.
He likened this to the simplification of hiring a highly experienced and expensive operator to training a less costly assistant with standard work experiences.
Regarding the potential impact of DeepSeek on national security, Caroline Levitt, the White House Press Secretary, stated on January 28 that US officials were investigating the implications of the Chinese AI application DeepSeek on national security.
On January 29, two US congressmen urged the Trump administration to consider restricting the export of Nvidia-produced H20 AI chips, as these chips are relied upon by the Chinese AI company DeepSeek.
President Trump expressed on January 27 that the launch of the DeepSeek AI by Chinese companies should serve as a warning to US industries, signaling the need to focus on competition.
In response to the DeepSeek unveiling, Howard Lutnick, the nominated Secretary of Commerce, harshly criticized DeepSeek during his confirmation hearing, alleging that they acquired a large number of Nvidia chips and found ways to bypass them to advance their DeepSeek model. He insisted that such actions must come to an end.
Reports cite anonymous sources stating that Trump is considering imposing stricter restrictions on Nvidia’s sales of semiconductor chips to China compared to the Biden administration, to limit China’s AI development.
According to Li Guanhua, the head of the Taiwan Industrial Technology Research Institute’s Policy and Regional Research Division, Huawei’s release of 7-nanometer smartphone chips in 2023 further exposed flaws in the original containment net against which the US could reinforce. From this perspective, it brings no benefits to them.
Li Guanhua acknowledged that while DeepSeek’s impact may suggest a failure in US chip control policies, it does not render such bans ineffective. Instead, it signifies that the bans have indeed had an effect, forcing China to explore alternative methods to achieve similar results in AI model training.
Du Yijin emphasized that while the bans may pose inconveniences for AI training, they are effective in some aspects but cannot completely halt AI development.
He noted that many reactions to the situation might be exaggerated, involving non-professionals adding complexity to what was initially a simple technological matter.
On the development of cutting-edge AI technologies and potential military applications, Li Guanhua highlighted that the US-China gap may widen. However, China might excel in applying these technologies in commercial sectors.
Du Yijin stated that as AI has become a national-level strategic focus globally, mastering powerful AI technologies could have revolutionary impacts on countries and entire industries.
He underscored that the US’s past technological superiority was not solely due to chip or computing power dominance but was primarily driven by an innovative environment, talent availability, capital resources, and software ecosystems that attracted top talent to contribute. This is what sustains continuous growth in US artificial intelligence.
He pointed out that current trade barriers, chip restrictions, and computing limitations could potentially lead countries to distance themselves from each other due to heightened competition. As nations gear up for AI technology race, future legislations will likely create pressure between countries, fostering a scenario of alliances based on shared AI technologies.
From the global perspective, the impacts of DeepSeek and the responses to its development indicate the severity of AI competition and its potential implications for international relations and technological advancements.