At the end of 2024, OpenAI released the full version of its new model o1, calling it the “smartest model in the world” that thinks “much like a person would.” However, within two months, OpenAI’s dominance faced an unprecedented challenge.
DeepSeek, a Chinese AI startup, rolled out its latest model R1 this year. The model performed remarkably well on reasoning and mathematical benchmarks. Independent testers placed it just behind OpenAI’s o1, surpassing similar closed-source products from Google, Anthropic and Meta. What’s more striking than its performance was its low cost and open-source status. DeepSeek claimed to have trained the model using Nvidia’s A800 chips — a downgraded “alternative product” released for the Chinese market under the U.S. chip ban — for less than $6 million, a fraction of the cost of its American counterparts.
Turning point of Sino-U.S. tech war?
Marc Andreessen, a venture capitalist and advisor to U.S. President Donald Trump, called this “AI’s Sputnik moment”—a reference to the Soviet Union’s surprise launch of the first satellite in 1957 ahead of the United States, which sparked the Cold War Space Race. DeepSeek’s cost-effective approach has rattled the U.S. AI industry, challenging the long-standing “Nvidia spell” and rendering the chip ban useless.
Within a month after its launch, DeepSeek replaced ChatGPT as the top free app on Apple’s App Store. Industry insiders believe that this is a wake-up call for America, escalating the AI race between the U.S. and China to an “AI war.” On last week’s “Black Friday” (January 27), U.S. stocks plummeted and tech stocks tumbled. Nvidia, once Wall Street’s golden child, plunged 17%, though the loss was later recovered. The selloff, however, reflects the growing doubts about America’s ability to maintain its AI edge.
Chinese tech enthusiasts and patriots were thrilled by DeepSeek’s breakthrough. Some likened the domestic model’s rise to a “sneak attack on Pearl Harbor” that “slapped the U.S. in the face.” Zhou Hongyi, founder of cybersecurity firm 360, described DeepSeek’s performance as “unbelievable” and called the company one of the “avengers” taking on America’s dominance in AI. He asserted that China will eventually win despite the increasingly intensive competition.
Both the U.S. and China seem to agree that the two countries are locked in an intense contest for AI supremacy, though the outcome of this technology cold war or AI “arms race” remains unpredictable. With this question in mind, I asked ChatGPT and DeepSeek: “Is there an AI race between China and the U.S.? Who is the winner?”
Surprisingly, yet unsurprisingly, both models “politely” responded: “The question is very complex, as there is no agreed-upon criterion; both countries have their strengths, and the AI race is not a zero-sum game.” I wonder if their developers would agree with this point of view, but at least both models acknowledge that there is indeed a fierce competition between the two countries in the sector.
DeepSeek’s sudden emergence
I also asked DeepSeek: “How did DeepSeek manage to bypass the U.S. chip ban and produce the R1 model?” OpenAI, after all, has spent a decade refining its technology, and this startup produced a comparable model in just two years. This is a David versus Goliath story in the AI world. DeepSeek replied that its advantages in algorithm optimization, distributed computing, software innovation, and hardware collaborative design enable it to circumvent chip restrictions.
In simple terms, constrained by America’s curbs on advanced AI chips, DeepSeek claims to have implemented various engineering enhancements, model architecture improvements, and efficiency optimizations to overcome hardware limitations. This is not DeepSeek’s first attempt to outperform its rivals on cost and performance. Its earlier V2 and V3 models set new benchmarks for affordability, triggering a price war among China’s large language model (LLM) developers. Some in Chinese media have likened DeepSeek to the “Pinduoduo (a Chinese online retailer and e-commerce platform that sells a wide range of products at low prices) of AI,” forcing domestic giants like Tencent, Baidu and Alibaba, who are used to distillation and shelling, to lower their prices. Now, the battle has gone international, putting Silicon Valley’s giants on the defensive. It is worth noting that according to Anthropic CEO Dario Amodei, American firms have spent $1 billion to train the next generation of AI models.
DeepSeek’s founder Liang Wenfeng, a graduate of Zhejiang University with a degree in Information and Communication Engineering, mentioned in a recent interview that DeepSeek has always rejected the “big company” model. As Chinese firms have long been hitchhiking on U.S. technological innovations and monetizing their applications, DeepSeek, as a company focused on R&D, aims to be a contributor to the field, starting by optimizing model architecture at its very core, he said.
Under the U.S. R&D model of “more chips, larger models,” DeepSeek clearly falls short in terms of purchasing power, computing capacity, and model-building infrastructure. Rather than following the traditional path of “big companies” toward downstream applications, it has chosen to take a different route — one that focused on improving the foundational structure of AI models to make the most of its limited resources. This approach, if successful, is undoubtedly groundbreaking.
After the R1 model’s sudden success, the mysterious parent company behind DeepSeek, “High-Flyer,” was unveiled. Founded in 2015, High-Flyer operates as a quant fund and AI company. In 2023, Liang, one of its founders, decided to establish DeepSeek as a subsidiary focusing on general artificial intelligence. Before the U.S. chip ban, High-Flyer had already stockpiled over 10,000 A100 chips. DeepSeek released its first large language model, DeepSeek-Coder, in October of the same year of its establishment.
Concerns over DeepSeek
While wowing the industry, DeepSeek also raised several concerns.
● Content censorship
As with all Chinese-made AI, content censorship remains an issue. Currently, DeepSeek employs a “weak censorship” mechanism, allowing users to bypass its “sensitive word” filter or momentarily trigger the display of sensitive content before self-censorship kicks in. However, when directly asked about sensitive topics such as Chinese President Xi Jinping or the Tiananmen incident, DeepSeek refused to answer, asking users to “change the topic.”
Data also show that DeepSeek’s accuracy in providing news and information is only at 17%, trailing behind comparable models like ChatGPT and Gemini.
● Cost transparency
Although DeepSeek claims to be open-source, it has not disclosed key details, such as its training data. Many suspect that the secret to the R1 model’s cost-efficiency lies in these undisclosed data, raising questions about whether DeepSeek has concealed some of its training expenses to present a lower development cost.
The widely circulated claim that “DeepSeek achieves similar results at a far lower cost than ChatGPT” is also misleading. Investigations by professional media outlets found that, before the R1’s launch, DeepSeek had already invested over a billion yuan in building its deep learning training platform — a cost that was not included into the $6 million figure. The $6 million is likely just a fraction of the total expenditure, as the majority of the investment goes into hardware and development.
● Lying about chips?
DeepSeek’s claim of using A800 chips raises suspicions, especially given the company’s reported stockpile of over 10,000 A100 chips. Alexandr Wang, CEO of Scale AI, recently suggested in an interview that China may have access to tens of thousands H100 chips. The stockpiling and potential use of these advanced chips — restricted under U.S. export controls — has become what some in China’s AI sector called an “unspoken secret.” DeepSeek may have sourced these chips through regulatory loopholes or other non-compliant channels, which could trigger an investigation by the U.S.
● Possible unauthorized use of data
Microsoft and OpenAI are recently probing whether DeepSeek “distilled” knowledge out of OpenAI models to train its large language models in order to achieve similar capabilities at a lower cost, and whether DeepSeek obtained data output from OpenAI in an unauthorized manner. Microsoft’s security researchers have alleged that an individual connected to DeepSeek used OpenAI’s application programming interface (API) to extract a large amount of data last year.
While distillation is a common practice in the industry, OpenAI believes that since the API is a paid service, “utilizing outputs from their paid API to develop competing AI models” is explicitly prohibited without its consent as stipulated in its terms of use.
However, some critics argue that this incident exposes potential vulnerabilities in OpenAI’s API. Moreover, some content creators, such as eight major U.S. media outlets including The New York Times, accused OpenAI of illegally using copyrighted internet data to train its systems, suing the company for copyright infringement.
Earlier reports suggested that DeepSeek may have trained on data outputs from ChatGPT. When asked, “What kind of model are you?” DeepSeek identified itself as ChatGPT, saying “I am ChatGPT,” and even claimed itself being developed based on ChatGPT-4 released in 2023. While ChatGPT-4 has some publicly available datasets, it remains unclear what specific data DeepSeek might have used, or whether this constitutes a breach of intellectual property. Replies like this could also stem from database limitations or training errors. DeepSeek is not the first AI model to misidentify itself. Google’s Gemini has also claimed to be Baidu’s Ernie Bot in Chinese-language interactions.
● Cybersecurity risks
DeepSeek’s English-language privacy policy states that it stores user data “in secure servers located in the People’s Republic of China.” This has prompted fears that the company could face privacy breaches like TikTok. Reports indicate that the U.S. Navy has advised its personnel to avoid using DeepSeek’s models, citing “potential security and ethical concerns.”
Meanwhile, DeepSeek experienced multiple outages within a week. The company announced it would temporarily limit registrations due to being targeted by “large-scale malicious attacks.”
The U.S. tech community’s growing mistrust in DeepSeek has only intensified the government’s concern and hostility towards the company, and vice versa. Some U.S. politicians believe that DeepSeek’s low pricing is a result of intellectual property theft, urging the U.S. national security officials to launch an investigation. After initially praising DeepSeek’s achievements, Trump quickly switched his attitude and sided with his Silicon Valley allies, highlighting the potential national security threats posed by the company.
A victory for open source or China?
All of DeepSeek’s models are currently open-source, allowing anyone, including tech professionals and companies worldwide, to freely access their reports and code.
Yann LeCun, Meta’s chief AI scientist and the 2018 Turing Award winner, recently wrote on his social media platform that DeepSeek’s performance does not indicate that “China is surpassing the U.S. in AI,” but rather that “open source models are surpassing proprietary ones.” Yann argued that DeepSeek R1 is built on open-source models, including Meta’s PyTorch and Llama, and its results are, in turn, fed back into the community through open access. This, he said, is the power of the open-source model — something proprietary systems cannot easily match.
There is a long-standing debate between open-source and closed-source AI models in the industry. Open-source models, such as Google’s TensorFlow, Meta’s PyTorch, and Hugging Face’s Transformers, provide code, models, and tools that anyone can modify and use. By contrast, closed-source models are proprietary technologies, internal assets with their technical details and code only available for internal use or authorization. OpenAI, once an advocate of open-source models, has since shifted to proprietary technology to keep its edge, a move that has drawn criticism from peers and former employees.
DeepSeek’s rise is being hailed as a victory for open-source models, and many believe that this success will boost the development of the open-source community. This idealistic vision of collaborative technological advancement is certainly a case of many hands making light work. Nvidia’s top AI researcher and former OpenAI employee, Jim Fan, praised the company on X: “We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive - truly open, frontier research that empowers all.”
However, this ideological battle does little to ease tensions between the U.S. and China. For leaders in both countries, AI is not just a technological breakthrough but an innovation capable of reshaping geopolitical power. Historically, each wave of technological transformation has shifted the balance of international geopolitical and military influence, allowing specific countries to rise to global leadership. AI is not just a tool for convenience or economic productivity through labor automation; it is also a potential force multiplier in military and defense strategy, with the power to enhance national strength. In its latest strategic blueprint, OpenAI has warned U.S. policymakers and the public to be wary of the Chinese Communist Party leveraging artificial intelligence to expand its global influence.
What unsettles the U.S. is DeepSeek’s free of charge. If the Chinese company offers high-quality open-source services at no cost, who would still pay for OpenAI or other paid models? Many researchers, startups, and even mid-tier U.S. companies may turn to DeepSeek’s transparent and accessible models. In this case, China’s AI models have the potential to rival or even surpass their U.S. counterparts, positioning themselves as the global standard in the sector.
Tech control does not hinder innovation
In 2017, China released “A New Generation Artificial Intelligence Development Plan,” setting a goal of achieving world-leading AI capabilities by 2030. But in 2022, the Biden administration imposed chip export controls aimed at cutting off its access to the advanced semiconductor chips needed to train large language models. Washington hoped that without cutting-edge chips, Beijing’s AI ambitions would falter, ensuring continued American dominance in the field. Chinese officials also saw the measures as an act of technological “strangulation.”
Before the ban was issued, as mentioned earlier, Chinese developers, including DeepSeek, had already begun stockpiling a large number of chips that were about to be restricted. At the same time, China adopted various methods to bypass U.S. export controls—essentially, smuggling. This included using shell companies based outside China to purchase restricted chips or renting chip access through offshore cloud providers. Reuters reported that a product manufactured by Huawei was found to contain restricted TSMC chips.
Moreover, some American companies are also working to circumvent government restrictions by selling technology to China for lucrative profits. Some lower-spec chips remain available for export. In late 2022, Nvidia designed a “China-specific” downgraded version of its chips, the H800 and A800 (which DeepSeek claims to use), in a bid to maintain its market share in China. However, the U.S. Department of Commerce ruled that these chips still had excessive capabilities, prompting a tightening of regulations at the end of 2023, effectively halting exports of these chips to China. Nvidia’s subsequent release, a second-generation “China-special” model, featured just a fraction of the original chip’s processing power — only one-seventh of the top-tier version. Many Chinese clients canceled their orders, mockingly dubing the chip the “castrated version.”
These measures did slow the pace of AI development in China in the early stages of export controls but not two years after their implementation. In 2024, Tencent, Kuaishou, and Alibaba launched several impressive AI applications, prompting observers to alert that China is closing the gap with the U.S. Former Google CEO Eric Schmidt asserted in May 2024 that America held a two-to-three-year lead over China in the AI sector. Yet by November, Schmidt had changed his tune. Citing the advancements of Alibaba and Tencent, he acknowledged that China was rapidly catching up, saying, “I thought the restrictions we placed on chips would keep them back.”
This resilience is partly driven by China’s “national system,” a centralized, government-led approach that directs national funding to AI projects and invests in critical infrastructure, including chip manufacturing plants and supercomputing centers. The private enterprise-led model in the U.S. is no match for the “concentrating resources to achieve a big goal” model. China’s vast data resources and a big talent pool further bolster its AI development.
In the end, the effectiveness of the chip ban may be overstated by American hubris. The hardware restrictions not only failed to weaken China’s AI ambitions, but instead boosted creativity. This is not the first time China has thrived under restrictions. A popular Chinese saying from the 1960s goes “Chinese people tightened their belts to develop atomic bombs.” It speaks to the country’s ability to overcome political isolation and economic hardship to achieve breakthroughs — in this case, becoming a nuclear power.
AI safety and ethics at risk
Within two weeks of DeepSeek’s debut, Alibaba, a leading Chinese tech company, announced its plan to roll out a new AI product that would surpass the capabilities of DeepSeek’s R1 model. OpenAI also revealed plans to release its more advanced O3 model, hoping to widen the gap with its rivals. The rapid pace of technological advancement means the shockwaves created by DeepSeek may soon dissipate as the next generation of models hits the market. However, this is unlikely to soften Washington’s stance. It may further tighten the export controls or impose more tariffs to curb China’s ambitions.
In the midst of this technological arms race, the mission of AI — to be “safe and broadly beneficial” — is at risk of being sidelined.
In 2023, OpenAI underwent a significant infight, marked by restructuring and departures of top executives. The company recalibrated its goals, officially placing product promotion and profitability before AI safety. Ilya Sutskever, co-founder and former chief scientist at OpenAI, left the company after years of championing the need to embed human values and safety measures into AI model design, ensuring that these technologies benefit humanity rather than pose a threat.
Initially founded as a non-profit research organization, OpenAI’s mission was to democratize AI research, making patents and findings publicly available through open collaboration. This was intended to ensure that artificial intelligence serves the collective good of society. However, the prohibitive cost of AI research soon prompted OpenAI to pivot towards commercialization, leaving its idealistic vision behind and leading to lawsuits from Elon Musk. Meanwhile, Meta also disbanded its Responsible AI team, reallocating most members to teams focused on developing generative AI products.
After taking office, Trump allowed the “national team” to join the game, announcing a $500 billion investment plan called the “Stargate Project,” which would unite several U.S. tech companies to build new AI infrastructure, accelerate R&D in advanced chips and algorithms, and cultivate talent. At the same time, Trump rescinded the executive order signed by former President Biden in 2023 on “Safe, Secure, and Trustworthy Artificial Intelligence.” The order sought to enhance federal oversight of AI to mitigate risks to consumers, workers, and national security. With its repeal, AI development will no longer be subject to a “speed limit.”
In the foreseeable future, the Sino-U.S. competition in artificial intelligence is set to intensify. Yet, both Chinese and American authorities risk overlooking the key consequence of this zero-sum game. The unchecked acceleration of AI — once rooted in the ideal of creating a safe, controllable, and beneficial technology for humanity — may soon stray from its original vision, even diverge entirely.
While many American hawks remain vigilant and hostile to the idea of China overtaking the U.S. in AI, accusing Beijing of using the technology for surveillance and censorship, there is a wider concern that global AI developers and leaders might exploit and abuse it in troubling ways. Paul Buchheit, the creator of Gmail, predicted after DeepSeek’s launch that if China emerges victorious in the AI arms race, humanity will find itself trapped in a digital zoo, permanently locked down and monitored. In fact, this warning is relevant to all leaders around the world who covet the power of artificial intelligence.
(The views and opinions expressed in this article are those of the authors and do not necessarily reflect the views or positions of WHYNOT or GNS.)
To read the original story in Chinese, click here.
