The emergence of DeepSeek has revitalized China’s artificial intelligence sector, attracting billions in state investment and enhancing the push for self-reliance. Its achievements have also ignited competition among startups to create products and services utilizing its top-tier open-source technology, as the nation’s leading tech firms race to launch rival AI models.
Increasing U.S. restrictions on chips, together with DeepSeek’s achievements, have accelerated innovation within the Chinese AI ecosystem, improving its standing in the global technology arena, according to experts and investors speaking to Rest of World. This situation has compelled startups to shift their focus toward more practical uses.
“DeepSeek has demonstrated that Chinese AI laboratories can produce cutting-edge models despite export control limitations,” Kevin Xu, founder of the U.S.-based hedge fund Interconnected Capital, which invests in AI, stated to Rest of World. “Its success is also driving more startups to concentrate on developing applications and services rather than squandering time and resources on model creation.”
Last year, China trailed the U.S. in the number of AI models created, but a recent report from the Stanford Institute for Human-Centered AI indicates that Chinese models are quickly closing the performance gap with their American counterparts. Nevertheless, Chinese companies face numerous difficulties, including increased U.S. export regulations that restrict access to advanced chips.
DeepSeek asserts it trained its V3 foundation model — a comprehensive AI system capable of performing various tasks using extensive data sets — on less advanced Nvidia chips at a cost nearing $6 million, in stark contrast to the over $100 million spent on OpenAI’s GPT-4 model.
The efficiency claims made by DeepSeek are likely influencing investor perceptions of AI companies beyond China, as noted by Melanie Tng, an analyst at PitchBook, a global market data provider.
“If high-performing models can be produced at a significantly lower cost, it raises questions about the viability of billion-dollar training budgets in other areas,” she remarked to Rest of World.
Investors in China are now likely to hesitate in supporting smaller AI firms that continue to focus on foundation models, since DeepSeek’s technology makes competition more challenging, according to Xu.
While a few firms will remain serious about developing AI models, “most others will concentrate on creating applications, services, and agents,” he indicated. “That is where investment will be directed.”
In contrast, Chinese tech giants are introducing new AI models and committing billions to research, suggesting a future where only major players will engage in AI model development, experts indicated.
On the first day of the Lunar New Year, Alibabai launched an open-source version of its leading Qwen large language model series, just days after DeepSeek’s announcement. The company also vowed to invest $53 billion over the next three years in its cloud computing and AI infrastructure, with the next generation of the Qwen LLM series set to launch in the coming weeks.
In recent weeks, Tencent Holdings, Baidui, and ByteDance have all released new AI models. Baidu, for its part, made its chatbot Ernie Bot available for free to the public at the start of April, ahead of schedule. Baidu CEO Robin Li alluded to the reasons for eliminating the paywall earlier this year, stating that the costs associated with model training “can be decreased by more than 90% within a year.” He emphasized the necessity of investment to remain at the forefront of this technological innovation or revolution.
Food delivery giant Meituan, as well, has introduced its own AI model, LongCat. The company claims that the model is already enhancing operational efficiency. Meituan founder Wang Xing committed “billions of yuan” to AI development and pledged to “take the initiative” to compete with other firms.
The larger tech companies have a strategic advantage over their smaller competitors when it comes to foundational research, according to Rui Ma, founder of Tech Buzz China, an investment research platform, who shared this insight with Rest of World.
“It’s because monetization from merely selling access to models is no longer viable with DeepSeek offering such affordable options,” Ma stated. “Everyone else must transition into being more product-oriented.”
Even prior to DeepSeek’s rise, smaller startups had already started adjusting to a lackluster investor interest in funding research, she mentioned.
Several of China’s “six little dragons” — the nation’s leading AI startups — have been compelled to change their direction as investor and consumer demand have decreased. Baichuan, for instance, ceased its model pretraining in mid-2024 to prioritize medical AI services, according to the Chinese tech news outlet 36Kr.
Another startup, 01.ai, announced last month that it will adopt DeepSeek as it shifts from creating its own AI models to becoming a solutions provider, particularly focusing on the finance, video gaming, and legal industries.
In contrast, Zhipu has recently introduced a free AI agent based on its proprietary model, asserting that it competes with DeepSeek’s capabilities. AI agents carry out research and personal assistant functions, such as booking flights and ordering food.
Pivoting to address “real-world problems” is a wise strategy, according to Beijing-based venture capitalist Celia Chen, who spoke with Rest of World.
“Instead of engaging in the high-profile infrastructure arms race, [Chinese AI startups] can explore and implement ideas at a much lower cost than what’s required to develop large models,” Chen noted.
Despite more government support and a surge of national pride, venture capital investment in the sector remains subdued.
During the first quarter of this year, venture capitalists invested $1.2 billion across 144 deals in Chinese AI and machine learning, which includes areas like speech recognition and robotic control, as reported by data provider PitchBook. The total value of deals fell by 30% compared to the same quarter last year.
That might not be entirely negative, according to Kayla Blomquist, director of the Oxford China Policy Lab, as it encourages companies to create AI products for monetization.
“I have observed that many startups are concentrating on building upon some of these foundation models, such as DeepSeek’s,” Blomquist stated. “This implies that perhaps [startups] will not require such massive levels of investment and VC support as we have witnessed … it will democratize AI in various ways.”
With anticipation surrounding the launch of DeepSeek’s next-generation model, another previously lesser-known Chinese startup, Butterfly Effect, introduced Manus, an invite-only AI agent, last month. The company asserts that Manus is the world’s first general AI agent, capable of performing tasks with a level of autonomy that existing AI models do not possess.
Recently, the company revealed a partnership with Alibaba. Manus’ co-founder, Yichao “Peak” Ji, explained that the agent was developed using several foundation models, including Anthropic’s Claude and Alibaba’s Qwen. Despite experiencing crashes and glitches, Manus has sparked enthusiasm for a brand-new generation of AI tools with diverse applications.
For Chinese AI firms, that might be the optimal combination.
“The true victors will be those who merge AI with industry knowledge to provide solutions that are difficult for large tech companies to easily imitate — such as firms that enhance the accuracy of medical diagnostics or expedite business workflows,” Chen stated.
“This is where significant opportunities are surfacing — not only for major players but also for founders targeting the mass market.”
The U.S. pioneers, while China refines. This saying—or its less flattering variant, “China copies”—permeates many discussions regarding the respective advantages of the world’s top two economies. The common narrative suggests that the U.S. stands at the forefront of technology, producing revolutionary products and services that influence the global economy. In contrast, China is seen as taking these concepts and enhancing them, delivering less expensive—and arguably inferior—versions.
In the field of AI, this notion seemed to hold true for years, as Chinese firms found it challenging to compete with the lavishly funded, talent-rich American tech giants. However, a Chinese startup challenged this view earlier this January.
DeepSeek, a startup from Hangzhou and not strictly a tech firm but an extension of a hedge fund called High-Flyer, introduced R1, a “reasoning” large language model (LLM) that matched the performance of OpenAI’s o1, which had debuted just a few months prior. Not only did R1 seemingly emerge out of nowhere, but it was also impressively innovative and incredibly affordable. DeepSeek claimed that the final “training run” for its predecessor, V3, cost a mere $6 million—“a laughable budget,” according to former Tesla AI scientist Andrej Karpathy, in comparison to the tens or hundreds of millions spent by some of its U.S. competitors.
The ramifications of this news were vast: As R1 soared to the forefront of popular download lists, investors in Big Tech were unsettled, resulting in a loss exceeding $1 trillion in value from tech stocks like Nvidia and Microsoft. Leaders such as OpenAI CEO Sam Altman publicly expressed their distress and pondered shifting to open-source models, similar to DeepSeek’s approach of making its model available for public modification, thus reducing costs.
“A lot of us, myself included, underestimated China’s capacity for delivering these cutting-edge breakthroughs,” states Jeffrey Ding, an assistant professor of political science at George Washington University and author of the ChinAI newsletter.
While the U.S. feels uneasy, there is excitement in China. DeepSeek’s founder, Liang Wenfeng, received a coveted invitation to a February meeting with Chinese President Xi Jinping and business leaders, sitting alongside notable figures like Alibaba founder Jack Ma and Huawei founder Ren Zhengfei. Major Chinese firms, including EV manufacturer BYD and home appliance producer Midea, are eager to integrate DeepSeek’s model into their products.
This development has injected a surge of optimism into a China that has been steeped in pessimism. “DeepSeek could single-handedly rejuvenate the economy in ways that the government has struggled to achieve,” asserts Paul Triolo, technology policy lead at advisory firm DGA–Albright Stonebridge Group.
However, DeepSeek represents just one aspect of a dynamic Chinese AI landscape, which many U.S. CEOs are largely unaware of. Significant players like Alibaba and ByteDance (the parent company of TikTok) are unveiling AI models that surpass Western products on reasoning benchmarks. Additionally, a new wave of smaller “AI dragons” is applying China’s cost-effective and efficient AI in practical ways through mobile applications, AI agents, and robots.
Meanwhile, investors have quickly returned to Chinese tech stocks. The Hang Seng Tech Index, which tracks technology companies listed in Hong Kong, has risen by 35% year-to-date, driven by shares from Alibaba, Kuaishou, which developed the text-to-video AI model Kling, and SMIC, China’s “national champion” chipmaker, aiding Huawei in producing AI chips.
Few long-term observers of China’s economy were caught off guard by DeepSeek’s emergence. AI may become the latest sector where being a “fast follower” allows the country to reach or even surpass its competition. Chinese solar panels and wind turbines are already facilitating the world’s transition to renewable energy. Chinese electric vehicles have contributed to making China the leading exporter of cars—and even non-Chinese electric vehicles utilize Chinese batteries. In fields such as drones, robotics, and biotechnology, Chinese firms are already global frontrunners.
Some Western CEOs dismiss these accomplishments, attributing them to government subsidies, intellectual property theft, smuggling, and violations of export controls. However, the elements that have allowed China to keep pace in various technological domains have proven to be both strong and enduring: a vast manufacturing base; a willingness to emulate foreign competitors; a deep talent pool; and a government capable of acting as a supporter and promoter.
Chinese innovators may prioritize “tailor-made problem-solving” rather than the “breakthrough, systemwide thinking” predominant in the U.S., as explained by economist Keyu Jin, author of The New China Playbook. Yet this method of directed, “good enough” innovation enables China to mass-produce technology that remains affordable, even when it approaches cutting-edge, as seen with DeepSeek. As Western business leaders grapple with the costs of AI advancements, China could soon provide what the rest of the world requires.
A quick recovery
Just two years ago, it seemed that China would lag significantly in AI. In 2020, Beijing initiated a prolonged effort to rein in China’s tech industry, which some political leaders perceived as excessively powerful and irresponsible. The once-consistent stream of Chinese tech IPOs came to a halt as Beijing tightened regulations surrounding data privacy.
The launch of OpenAI’s ChatGPT in 2022 highlighted the extent of the AI gap. Chinese-developed LLMs released afterward generally delivered inferior performance compared to ChatGPT, even in the Chinese language. At the same time, U.S. export restrictions prevented Chinese firms from obtaining Nvidia’s AI chips, which are crucial for training and operating LLMs.
Ding observes that the narrative began to change in the fall of 2024. “You started to see the gap closing,” he notes, especially in the open-source sector, where “Chinese companies began optimizing for smaller-sized models that could be trained more effectively.”
China’s AI industry has been nurturing a surge of new startups. Initially, there were the “little dragons,” firms focused on machine learning and computer vision like SenseTime and Megvii that gained global recognition. As focus shifted to generative AI, attention turned to the “AI tigers”—Baichuan, Moonshot, MiniMax, and Zhipu. Currently, they’ve been outshined by the latest group of “dragons”: six startups based in Hangzhou, including DeepSeek.
Hangzhou, home to Alibaba, is the center of China’s AI innovation. “It benefits from being distanced from Beijing to evade bureaucratic hurdles, yet close enough to Shanghai to tap into international capital and talent, along with a robust talent pool thanks to Alibaba, NetEase, and others,” explains Grace Shao, founder of Proem, an AI consulting firm. Alibaba itself has been a supporter of open-source creators: The top 10 LLMs ranked by performance on Hugging Face, an open-source AI community, are based on Alibaba’s Tongyi Qianwen models.
Among the city’s exciting new dragons are Unitree, whose dancing robots showcased at this year’s televised Spring Festival Gala, attracting hundreds of millions of viewers; Game Science, the studio behind Black Myth: Wukong, one of 2024’s best-selling video games; and Manycore, a company specializing in “spatial intelligence” and 3D rendering.
What has allowed China’s AI sector to close the gap so rapidly? One factor is China’s sheer size. Shao points out that DeepSeek’s user base significantly expanded when Tencent, the operator of the widespread WeChat platform, provided DeepSeek’s LLM to its more than 1 billion users—immediately propelling the startup into prominence in China’s AI landscape.
The government plays a crucial role in how to utilize that scale. Through various policies, regulations, and subsidies, officials nationwide create a “state-coordinated” innovation framework, and the private sector typically aligns with the system’s objectives. The government acts as a motivator, Triolo states: “When Liang Wenfeng meets with Premier Li Qiang and President Xi Jinping, that sends a clear message.” Indeed, that high-profile meeting in February sparked the excitement surrounding China’s DeepSeek: Initially, Chinese telecom companies adopted and promoted its LLMs, which spurred tech and consumer companies to follow suit, ultimately influencing local officials.
Ironically, U.S. restrictions on semiconductors may have sped up China’s innovation. “Money has never been an issue for us; the ban on advanced chip shipments is the concern,” Liang shared with Chinese media last year. For years, China’s chip sector struggled to gain traction, as manufacturers were able to procure better chips from overseas. However, the U.S. trade restrictions “mobilized the entire nation to pursue advancements,” says economist Jin.
Huawei, the telecom giant, is now at the forefront of China’s advanced-chip supply chain. Its Ascend AI chips—although not yet as powerful as Nvidia’s—are being utilized by startups like DeepSeek for “inference,” which refers to running AI models in real-world applications once they have been trained.
The final aspect is talent: China’s universities are producing enthusiastic engineers that AI startups can recruit. While some engineers at DeepSeek have received Western education, Triolo mentions that “Liang Wenfeng has actively recruited these top talents—young individuals who lack experience in the West and weren’t trained at MIT and Stanford.” CEOs are “astounded by the caliber of graduates emerging from second-, third-, and fourth-tier universities in China. Such talent in those numbers is hard to find at U.S. institutions.”
Some analysts also perceive a significant shift in mindset among founders from China’s “’90s generation.” Older tech founders believed it was acceptable to “copy, but improve,” as Shao explains. “Now [entrepreneurs] discuss open-source as a fundamental choice. China can innovate instead of just imitating.”