- 前半段為文章的英文版本 (The first half is the English version)
- 後半段為中文版本 (The second half is the Mandarin version)
Recently, as NVIDIA’s stock price experienced significant volatility, the market’s attention swiftly turned to the catalyst behind it—DeepSeek, an AI startup from China. Founded by 40-year-old Wenfeng Liang, DeepSeek’s AI chatbot has delivered impressive performance. However, what truly shook the market was not just its technology, but its astonishingly low training cost.
DeepSeek reportedly completed its model training for just $6 million, yet achieved performance comparable to OpenAI’s GPT-4o, while OpenAI’s training costs soared to $100 million. This stark contrast in cost, coupled with the strict export controls imposed by the U.S. on high-end AI chips to China, has led to growing speculation: Is America’s lead in the global AI race at risk? Furthermore, it raises a critical question: Are NVIDIA’s high-end chips truly indispensable?
Is DeepSeek’s Performance Overstated?
I do not claim to have the expertise to determine whether DeepSeek’s AI performance has been exaggerated. However, several world-class AI experts, including Sam Altman and Yann LeCun, have given it positive reviews. Additionally, Artificial Analysis’ LLM Leaderboard ranks DeepSeek R1 as the second-best model globally, just behind OpenAI (Artificial Analysis LLM Leaderboards). This strongly suggests that DeepSeek possesses formidable technical and scientific capabilities.
Nonetheless, skepticism remains regarding its low cost claims. DeepSeek states that it trained its model using only 2,048 NVIDIA H800 GPUs, completing the process in under two months at a total cost of $6 million—just 5% of GPT-4o’s cost. However, NVIDIA’s H800 is a restricted version designed for the Chinese market, with lower computational efficiency than the H100. Moreover, in 2023, the U.S. government banned exports of H800 to China. This raises a critical question: How did China manage to train such a powerful model at a fraction of the cost, despite facing chip restrictions? The answer remains unclear and warrants further investigation.
Is This the “Sputnik Moment” of AI?
Regardless of the doubts, DeepSeek’s emergence has undeniably sent shockwaves through America’s AI leadership. This explains why NVIDIA lost $600 billion in market value in a single day—a clear reflection of market anxiety and uncertainty. The escalating AI competition between the U.S. and China inevitably evokes parallels to the “Sputnik Moment” of the Cold War—when the Soviet Union successfully launched the world’s first artificial satellite, stunning the United States and accelerating the space race. However, unlike that era, today’s LLM leaderboards remain dominated by American companies, indicating that U.S. dominance in AI is still firmly intact and far from reaching a true crisis.
That being said, one thing is evident: China’s rapid progress in AI is narrowing the gap at an unprecedented pace—especially considering that DeepSeek achieved its advancements despite stringent U.S. AI chip export controls. Moving forward, the U.S. is expected to further tighten AI-related restrictions on China. Meanwhile, newly elected President Donald Trump has made it clear that he intends to deregulate American tech industries to unleash AI development. With Elon Musk and other tech visionaries playing key roles in his administration, it is foreseeable that America will accelerate its AI ambitions with virtually unlimited investment, resources, and global talent acquisition—ensuring that it maintains a decisive edge in this civilization-defining competition.
How Should We Navigate This AI Revolution?
The controversies surrounding DeepSeek are still unfolding—ranging from the veracity of its claims to the geopolitical implications of U.S.-China AI competition, and even rumors of cyberattacks. However, what concerns me more is the intensifying AI rivalry between the two superpowers, which increasingly mirrors the Cold War’s global tensions and raises concerns about its long-term impact on world affairs.
As the AI revolution reshapes industries, economies, and investments, its effects will be profound for businesses, investors, and professionals alike. How should we position ourselves in this rapidly evolving landscape? How can we not only keep up, but seize the opportunities presented by this accelerating technological shift?
This AI-driven race for the future of human civilization has only just begun.
====
近期，隨著NVIDIA股價劇烈震盪，市場的焦點迅速轉向「始作俑者」——來自中國的AI新創公司DeepSeek。這家由40歲的梁文鋒創立的企業，其AI聊天機器人表現驚人，但真正引發市場動盪的並非其技術，而是它令人難以置信的低成本訓練模式。
DeepSeek據稱僅花費約600萬美元完成訓練，卻能達到媲美OpenAI GPT-4o的水準，相較之下，OpenAI的訓練成本高達1億美元。這巨大的成本落差，加上美國對中國高階AI晶片的嚴格出口管制，讓人不禁開始懷疑：美國在全球AI競賽中的領先地位是否正在受到挑戰？ 更進一步，這也讓市場開始質疑：NVIDIA的高階晶片是否真的不可或缺？
DeepSeek的實力有誇大成分嗎？
對於DeepSeek的AI效能是否「被誇大」，我並沒有足夠的信心去做出判斷。然而，許多世界級AI專家，包括Sam Altman與Yann LeCun，都給予了正面評價。此外，Artificial Analysis的LLM Leaderboard亦將DeepSeek R1評為全球第二，僅次於OpenAI (Artificial Analysis LLM Leaderboards)。由此可見，DeepSeek的技術能力確實不容小覷。
然而，市場對於其「低成本」仍抱持懷疑。DeepSeek聲稱僅使用2,048張NVIDIA H800 GPU，不到兩個月便完成訓練，總成本約600萬美元，僅為GPT-4o的5%。但NVIDIA H800是針對中國市場的受限版本，運算效率不及H100，而美國更已在2023年禁止該晶片出口至中國。因此，在受限條件下，中國如何能以遠低於美國的成本訓練出如此強大的模型？ 這點仍有待進一步驗證。
AI競賽是否迎來「史普尼克時刻」？
無論如何，DeepSeek的崛起無疑敲響了美國AI領導地位的警鐘。這也是為何NVIDIA一天內市值蒸發約6000億美元——市場的不安與焦慮溢於言表。美中AI競爭日益升溫，這不禁讓人聯想到冷戰時期的「史普尼克時刻」——當蘇聯成功發射全球第一顆人造衛星，美國深感震驚，進而加速太空競賽。然而，與當年不同的是，目前全球LLM排行榜仍被美國企業主導，顯示其AI領先地位依然穩固，未至於出現「危機」級別的動搖。
即便如此，我們仍可明顯看到：中國在AI領域的進步速度驚人，差距正在快速縮小——尤其是在美國對中國祭出嚴格AI晶片出口管制的情況下，DeepSeek仍能脫穎而出，足見其競爭力。未來，美國對中國的AI技術限制勢必只增不減，而新任總統川普已明確表態：將放寬美國科技發展的法規，讓企業能更自由發展AI。此外，Elon Musk等科技巨頭亦積極參與AI政策制定，可見美國在這場競賽中的投入將更為無上限。
我們該如何應對這場AI變革？
圍繞DeepSeek的爭議仍在發酵，數據的真偽、美中AI競爭的格局，甚至駭客攻擊的傳聞，都成為熱議焦點。然而，對我而言，更值得關注的是美中之間日益升溫的競爭態勢，這不僅讓人聯想到冷戰時期的全球對抗局勢，更令人擔憂其後續發展對世界格局的影響。
在這場全球AI競爭的浪潮中，無論是事業、投資，還是各行各業的未來，都將受到深遠影響。我們該如何應對？如何不僅跟上，更能抓住這股時代變革的機遇，搭上這台高速行駛的列車？
這場AI作為主角的人類文明賽局，才剛剛開始。