The Takeaway
• Alibaba and Chinese state funds have expressed interest in investing in DeepSeek
• Startup has previously rebuffed outside investor interest
• Need to raise more money for chips and servers may change its stance
DeepSeek’s sudden rise to artificial intelligence stardom has created a dilemma for the Chinese startup: whether to raise money.
A two-year-old offshoot of a Chinese quantitative hedge fund, DeepSeek so far has not raised outside funding, partly to avoid pressure from investors to commercialize its products too quickly. But since its AI chatbot app shot to fame a few weeks ago, the need for more AI chips and servers to handle fast-growing usage and support model development has prompted the company to consider outside fundraising, according to people with knowledge of internal discussions.
And in recent weeks, many powerful, deep-pocketed investors, including Alibaba Group and Chinese state–affiliated funds, have expressed interest in financing the company’s next stage of growth, according to two people with direct knowledge of the talks. Among the state investors that have contacted DeepSeek are China Investment Corp., the country’s sovereign wealth fund, and the National Social Security Fund, according to one of the people.
DeepSeek, Alibaba and China Investment Corp. didn’t respond to emailed requests for comment. Calls to the National Social Security Fund were not answered.
In addition to discussing whether they should raise money, executives at DeepSeek and its hedge fund parent, High-Flyer Capital Management, are also talking about whether the startup should pivot away from a primary focus on research toward building a business that generates meaningful revenue and eventually profits, according to people with knowledge of those discussions. It is also considering using data centers in Southeast Asia to gain access to more Nvidia AI chips, according to the people.
DeepSeek founder and CEO Liang Wenfeng—who is also a co-founder of High-Flyer—remains cautious about raising outside capital. It’s unclear what direction he will go in.
Nerdy and Idealistic
People who have worked with Liang describe him as a nerdy and idealistic technologist. He co-founded High-Flyer in 2015 to focus on using AI models for stock trading, then still a novel idea in China. Liang counts Jim Simons, the late founder of New York quant-trading fund Renaissance Technologies, as his role model, according to a preface Liang wrote for the Chinese edition of a biography of Simons written by a Wall Street Journal reporter.
In 2023, after OpenAI sparked the AI revolution with the release of ChatGPT, High-Flyer set up DeepSeek as a separate subsidiary to work mainly on advanced research projects. Liang’s aim was to achieve artificial general intelligence, a milestone where AI achieves human-level cognitive abilities. He set up DeepSeek like an academic laboratory and recruited researchers who also wanted to pursue AGI.
DeepSeek had a leg up over other Chinese AI startups: High-Flyer had stockpiled about 10,000 Nvidia A100 chips before the U.S. restricted sales of those semiconductors to China on national security grounds, according to former employees.
The company showed no interest in raising outside money. Partners of two Chinese venture capital firms who reached out to DeepSeek in late 2023 and early 2024 about possible investment say it made it very clear to them that it had no intention of raising any outside capital.
Meanwhile, Liang was paying close attention to AI developments in the U.S. In September, when OpenAI launched its o1 reasoning model, Liang flew to the U.S. and met with researchers he knows, including some OpenAI employees, to stay up to date, according to two people with knowledge of his U.S. trip.
In an internal meeting with employees late last year, Liang said DeepSeek was lagging behind OpenAI. But Liang also said DeepSeek could contribute to the AI industry’s progress by reducing costs and increasing efficiency, according to a former employee who attended the meeting.
DeepSeek’s older AI models, which were available to users through the company’s website, started to gain recognition late last year among AI researchers in the U.S. At the same time, DeepSeek was holding off on releasing a mobile chatbot app for consumers, even when many other tech companies in both the U.S. and China were rushing to do so.
In the second half of last year, for instance, when other Chinese companies’ AI apps, such as ByteDance’s Doubao, were quickly gaining popularity, Liang told employees in a meeting that the time was still not right for DeepSeek to focus on a consumer app. He said the company’s first priority should always be its research to build better foundation models more efficiently, according to two former employees who heard those comments.
DeepSeek Shock
Meanwhile, DeepSeek was gaining more attention. In late December, DeepSeek released V3, an updated large language model it claimed it had trained with just 2,048 Nvidia chips and at a cost of just $5.5 million, a fraction of what U.S. AI companies were spending on their own LLMs.
On Jan. 20, Liang was invited to a symposium hosted by Chinese Premier Li Qiang. And a few days later, DeepSeek released a new reasoning AI model, DeepSeek-R1. The model shocked Silicon Valley by offering performance comparable to OpenAI’s similar model even though R1 had been developed at a much lower cost by DeepSeek’s account.
DeepSeek’s first mobile app, launched a month ago, suddenly exploded in popularity and racked up 30 million daily active users in China in less than a month, the fastest growth ever in the country, according to research firm QuestMobile.
But it struggled to meet demand. During the Chinese New Year holiday in the first week of this month, when most of the country ground to a halt for the annual festivities, some DeepSeek employees kept working on solving the infrastructure challenge. The company has curbed the usage of its app in an attempt to make the traffic more manageable, but it will likely need to invest more in infrastructure to better accommodate existing users’ activities.
The DeepSeek shock has also sparked concerns in Washington, where lawmakers have long viewed China’s AI advancements as a threat to U.S. national security. Earlier this month, they announced a plan to introduce a bipartisan bill banning DeepSeek’s app from government-owned devices.
Meanwhile, the company’s fundraising plans are a fraught issue. Raising money from any major Chinese investor, especially a state-owned one, could add more fuel to Washington’s concerns and eliminate or limit DeepSeek’s future business opportunities in the U.S. Last month, the U.S. Commerce Department added Chinese startup Zhipu AI, whose investors include multiple state-backed funds, to an export blacklist, alleging that its AI research contributes to China’s military advances.
U.S. investors and funds that raise money from American limited partners would likely face difficulties backing DeepSeek because of U.S. rules that came into effect last month restricting investments in Chinese AI companies.
It is unclear whether High Flyer’s hedge fund business, which has faced more competition from other Chinese quant funds in recent years, can continue to provide all of the capital DeepSeek needs. DeepSeek’s own revenue so far is insignificant. Its chatbot app is free, and the application programming interface for its R1 model charges only 14 cents for input and a little over $2 for output per 1 million tokens—which refers to the units of text that language models process. Those prices are a fraction of those for OpenAI’s o1 reasoning model, which charges $15 for input and $60 for output per 1 million tokens.
Earlier this month, Tencent’s WeChat, the ubiquitous messaging and social networking app in China, started testing a feature that allows users to access DeepSeek’s AI model. It is unclear how much this could add to DeepSeek’s revenue.
Another big challenge for DeepSeek is how to secure its access to more Nvidia AI chips powerful enough for its AI development. While DeepSeek has been using Nvidia’s A100 chips, which High-Flyer acquired in 2021 before the U.S. imposed export restrictions, staffers at the company say they will need to get more Nvidia chips.
Late last year, in a team meeting, Liang talked to employees about the idea of using a data center in Malaysia to gain access to more Nvidia chips superior to H20, according to a former employee who attended the meeting. Other major Chinese AI competitors, such as ByteDance, have already made arrangements to use Nvidia chips without violating U.S. export controls by renting Nvidia-powered servers at data centers outside China.
