China’s DeepSeek AI Shakes Up the Game

Implication on US Dominance, Nvidia and TSMC By Judy Lin, January 27th, 2025. What’s all the Fuss About? Chinese AI startup Deepseek ignited a round of self-doubt on the US stock markets, sen

**Implication on US Dominance, Nvidia and TSMC** **By Judy Lin, January 27th, 2025.** **What’s all the Fuss About?** Chinese AI startup Deepseek ignited a round of self-doubt on the US stock markets, sending AI-related stocks into a tailspin. While this pessimism may be overblown, it is a good time to take a deep dive on Deepseek. Deepseek unveiled V3 in December and R1 in January. Now at the World Economic Forum (WEF) and all over the world, it is the hottest topic people are talking about. Its app has skyrocketed to the top of the U.S. free app charts just a week after its launch. President Donald Trump called the Chinese company’s rapid rise "a wake-up call" for the U.S. tech industry, as its AI breakthrough sent shockwaves through Wall Street. **The Breakthrough Thesis** Some calculations show OpenAI O1 costs $15 per million input tokens and $60 per million output tokens; DeepSeek Reasoner, based on the R1 model, costs only $0.55 per million input and $2.19 per million output tokens. Deepseek’s open-source reasoning model R1 is on par with the performance of OpenAI’s O1 in several tests. They built their model at the cost of US$5.6 million, which is only a fraction of the cost of OpenAI’s O1. Some say DeepSeek-R1’s reasoning performance marks a significant win for China, especially because the entire work is open-source, including how the company trained the model. Nevertheless, Deepseek’s extremely low cost and efficiency for training AI models invite investigations into how it is possible to spend only US$5.6 million to accomplish what others invested at least 10 times more and still outperform. Deepseek shattered the impression that the US was way ahead of China in AI, largely because China does not have access to the most advanced NVIDIA GPUs. ScaleAI CEO Alexandr Wang told CNBC at the sideline of WEF that Deepseek has at least 50,000 Nvidia H100 chips (though this has not been confirmed), prompting many to question the effectiveness of export controls. Now, who is behind Deepseek? Why is Deepseek able to achieve such great results? What are the possible factors contributing to Deepseek's success? How should we correctly interpret the implications of Deepseek’s success on the AI competition between the US and China? **Introduction to Journalist** Hi, I am Judy Lin, founder of TechSoda, a news platform that provides refreshing insights to the curious mind. Why soda? It is the acronym for “semiconductor,” “optics,” “digital,” and “AI.” I am a senior journalist who covers the macroeconomic and foreign exchange market, banking/insurance/fintech, and technology business news in Taiwan for decades. My research interests in international business strategies and geopolitics led me to examine how industrial and trade policies impact the business of companies and how they should respond or take preemptive measures to navigate uncertainty. My studies in international business strategies and risk communications, along with my network in the semiconductor and AI community here in the Asia Pacific, have been useful for analyzing technological trends and policy twists. Seeing semiconductors become a strategic industry that many countries hold dear for national security, I strive to make my tech articles accessible to people who are not scientists or engineers but are eager to learn more about the semiconductor supply chain. If you find my articles interesting and helpful, please subscribe to my newsletters for the latest updates. In this article, I will discuss the following topics: 1. What and Who: Background of Deepseek and its founder and lead developer. 2. How Deepseek was able to achieve low cost while obtaining similar performance. 3. Does Deepseek’s success mean we don’t need as many Nvidia GPUs? What is the impact of decreased demand for GPUs on the foundry? 4. Why is the Deepseek phenomenon significant: Implications for the US-China AI War going forward? **Background of Deepseek, its Founder and Principal Researcher** **Founder** Deepseek was founded in July 2023 by Liang Wenfeng, a graduate of Zhejiang University's Department of Electrical Engineering and a Master of Science in Communication Engineering. He founded the hedge fund “High-Flyer” with his business partners in 2015, quickly rising to become the first quantitative hedge fund in China to raise more than CNY100 billion. He grew up in the 1980s in a fifth-tier municipality in Guangdong. **Principal Researcher - Talent and Recruitment** Liang’s idealism or curiosity alone cannot make it a success; his recruitment standards and management methods are key, said Feng Xiqian, a Hong Kong commentator. “Liang’s hiring principle is based on ability, not experience, and core positions are filled by fresh graduates and young people who have graduated for one or two years. To measure the ability of newcomers, apart from institutional background (mainly Tsinghua and Peking University students), he also evaluates competition results and does not consider anything below a gold medal – he only recruits 1% of the top geniuses to do what 99% of Chinese companies can't do.” Luo Fuli, Principal Researcher of Deepseek, is one of the 139 employees who have demonstrated exceptional talent at a very young age. A Beijing citizen, she has not turned 30 yet but has already published 41 papers since 2018 and was a recipient of China’s National Scholarship in 2016. **How Deepseek Achieved Low Cost While Obtaining Similar Performance** **About Getting On Par Performance** I tend to hold a critical thinking position on this. Deepseek was established in July 2023, while OpenAI was founded in 2015. While OpenAI and many western AI companies had to build their generative AI from the ground up, Deepseek, as a latecomer, was able to avoid many pitfalls experienced by those predecessors and build on the foundations of open-source contributors. **About Low Cost** 1. **More Efficient Operations** Due to the US export control, Deepseek had to devise a more effective way to train the model. They combined various engineering techniques to improve the model architecture and ultimately succeeded in breaking through technological bottlenecks under the export ban. Using fewer computing resources to perform complex logical reasoning tasks not only saves costs but also eliminates the need for the most advanced chips. 2. **Energy Costs** Some argue that China burns cheaper coal, making energy consumed by Deepseek less expensive. However, this is not entirely accurate. The increased use of renewable energy and innovations in energy efficiency are key factors. 3. **Government Subsidies** In addition to the subsidies provided by the central government, local municipal and provincial governments also have incentives to support AI companies in China. These incentives include tax breaks, investments, affordable rents for offices located in AI clusters operated by local governments, and talent training programs. **Does Deepseek’s Success Mean We Don’t Need As Many Nvidia GPUs?** Will Nvidia be affected in the short term by the drastic reduction in the cost of AI training? “I don't think so, because when AI becomes popularized and generalized at a low cost, it will only increase the world's demand for it,” wrote Sega Cheng, CEO and co-founder of iKala, a Taiwanese AI company. **Why is the Deepseek Phenomenon Significant: US-China AI War Changing Directions?** Export controls have slowed down the catch-up speed of China’s AI technology development, but the win of Deepseek this round is unlikely to make the US give up its policy. The long game for AI supremacy competition is becoming more complex. Chris Miller, author of *Chip War*, revealed at the CommonWealth Economic Forum in early January 2025 how AI is transforming the US-China Chip War into a broader "Cloud War." He noted that scaling laws are faltering and efficiency is overtaking raw scale. Whether through breakthroughs in inference compute, efficient algorithms, or geopolitical maneuvering, the Chip War is evolving into a broader contest for technological and economic supremacy in the age of AI. Miller also believes that tech decoupling is already in place. The rise of DeepSeek AI marks a pivotal moment in the global AI race, proving that innovation can thrive under constraints. While U.S. export controls aimed to slow China’s progress, they may have inadvertently fueled a wave of ingenuity, forcing Chinese engineers to think differently and prioritize efficiency over sheer scale. Yet, this breakthrough is unlikely to prompt Washington to reconsider its policies. If anything, it reinforces the view that the AI rivalry is evolving into a broader "Cloud War," where technological and economic supremacy will be defined not just by hardware, but by intelligence, adaptability, and strategic maneuvering.

By Romain Peter