The seemingly overnight success of DeepSeek wiped billions of dollars from the fortunes of the world’s richest people—and catapulted the Chinese AI firm’s founder, Liang Wenfeng, into the three-comma-club.
By Giacomo Tognini and Phoebe Liu
On January 20, on the same day as Donald Trump’s inauguration in Washington, Chinese premier Li Qiang held a meeting with experts to consult on the Chinese government’s policies for the year ahead. It was a low-key event that got little attention outside of China. One of the few people to speak at the gathering was Liang Wenfeng, a bespectacled hedge fund founder and AI entrepreneur who was then little-known outside the country. That was also the day his firm DeepSeek launched its latest model, R1, and claimed it rivals OpenAI’s latest reasoning model.
Within a week, DeepSeek’s app had rocketed to the top of app stores in the United States, dethroning those of its competitors including OpenAI’s ChatGPT and Anthropic’s Claude and turning Liang and his AI firm into household names.
Liang Wenfeng at the January 20 meeting with Chinese premier Li Qiang.
CCTV
The best-known AI startups in the U.S. are worth tens of billions of dollars—ranging from $50 billion for Elon Musk’s xAI to $157 billion for OpenAI—and have raised huge sums from the world’s most prominent investors, including Microsoft, Amazon and Silicon Valley’s top venture capital shops. Unlike its U.S. competitors, DeepSeek appears to have no external investors outside of Liang and his three cofounders. According to Chinese corporate records, Liang owns about 84% of the Hangzhou-based firm, which he founded in 2023 and financed with funds from High-Flyer Capital Management, the quantitative trading hedge fund he cofounded in 2015.
How much is Liang’s open-source AI model worth? spoke to five analysts and investors who offered a range of potential valuations, but the three who provided a specific number agreed DeepSeek is worth at least $1 billion—and potentially far more, despite the fact that it doesn’t yet generate much revenue. Right now, DeepSeek has one paid product: developer access to its models. Its reasoning model costs $2.19 per million output tokens (on average 750,000 words), far less than OpenAI’s $60. That low price could simply be to undercut larger competitors OpenAI and Anthropic to capture market share, but estimates that DeepSeek generates around $6 million in annualized revenue per million paying users, using estimates of a typical user’s token usage from investment firm D.A. Davidson.
Apply that to a revenue multiple of 65 (somewhere between that of Anthropic and Chinese open-source AI startup 01.AI, which raised outside capital in late 2023), and DeepSeek would need around 3 million paying users to reach a billion-dollar valuation. More than 3.6 million people downloaded DeepSeek’s app in its first two weeks, per Appfigures data. William Blair partner and software analyst Arjun Bhatia thinks that less than 10% of that number are paying users—but also says the multiple should be applied to DeepSeek’s user count a year or two from now.
Regardless of DeepSeek’s exact revenue, estimates it’s worth at least $1 billion—in part because state-of-the-art models like DeepSeek’s are often valued at their “blue-sky potential” rather than as a multiple of revenue, according to tech investor and Corpora.ai CEO Mel Morris. DeepSeek is arguably in the “top five AI labs in the world right now,” says D.A. Davidson analyst Alexander Platt, adding that it should be worth more to “account for the research horsepower there which isn’t necessarily monetizable.”
Bhatia says $1 billion “seems like a very low number” even after the “China discount” (due to geopolitical uncertainty since DeepSeek is a Chinese company), while Morris pegs DeepSeek’s value at $10 billion. “There might even be an annoyance factor where someone might actually be prepared to pay that just to take them out of the picture,” he says.
There’s also the case of DeepSeek’s Chinese competitors—none of which seem to have achieved performance as good as DeepSeek’s, but all of which external investors have valued at $1 billion or more in various funding rounds. DeepSeek hasn’t raised VC funding in part because Liang believes VCs want to “exit and hope to commercialize products as soon as possible,” which didn’t align with DeepSeek’s research priorities, he told Chinese tech outlet 36Kr in 2023.
On top of that, Liang also owns at least 76% of High-Flyer, which has $8 billion in assets according to financial data provider Preqin. values High-Flyer at $240 million, with Liang’s stake worth about $180 million. (Chinese corporate records also show that he owns 85% of another High-Flyer entity that was first registered in 2015 and manages 65 of the firm’s 503 active funds, meaning his equity in the firm could be even higher.) Put it all together and Liang is likely worth at least $1 billion, making him the latest founder to mint a fortune from AI.
Got a tip or have additional thoughts? Contact Phoebe Liu at pliu@.com or 678.834.4200 on Signal, and Giacomo Tognini at gtognini@.com or giacomo.na.tognini@proton.me.
Not that he likely cares much. “An exciting thing cannot be measured purely by how much it is worth,” Liang told 36Kr, speaking of DeepSeek and adding how he’d been interested in testing the limits of computing power since 2012. “It’s like buying a piano for the home. On one hand, you can afford to buy it; on the other, it’s because there’s a group of people eager to make music with it.” Liang and his firm, which are likely closed due to Chinese New Year, did not yet return ’ requests for comment.
Liang was born in 1985 and grew up in the port city of Zhanjiang in southern China. The son of a primary school teacher, he studied artificial intelligence at Zhejiang University in Hangzhou and got a bachelor’s in electronic engineering in 2006 and a master’s in information and communication engineering four years later.
While he was still in graduate school, he started exploring how to fully automate trading on China’s domestic stock market. In 2013 he founded his first investment firm, Hangzhou Jacobi—named for German mathematician Carl Jacobi—with his college classmate Xu Jin. Two years later the pair cofounded High-Flyer with another classmate, and the trio used math and AI techniques to build a hedge fund.
At that point, Liang already had 100 graphics processing units—the high-tech chips that help train AI models—powering High-Flyer’s investment decisions. By 2019, High-Flyer had become one of the largest and best-performing quantitative trading firms in China, and Liang spent nearly $30 million expanding that footprint to 1,100 chips and building his own facilities to house them. He also doubled down on AI, setting up a separate company—Hangzhou High-Flyer AI—to research AI algorithms and their applications and expanded High-Flyer overseas, setting up a fund registered in Hong Kong. The firm has a similar structure to most hedge funds, charging 2% of assets in annual management fees and 20% of profits for its “enhanced funds” while charging a higher profit fee for investors in its quantitative hedging and Hong Kong-based funds.
In 2021, as High-Flyer reached a peak of around $14 billion in assets under management—generating an estimated windfall of more than $200 million in management fees for the firm—Liang spent another $155 million to buy 10,000 of Nvidia’s A100 chips. In a 2021 pitch deck for High-Flyer viewed by , the firm revealed it spent 60% of its research funding on its AI lab. That investment came after one of High-Flyer’s best years in 2020, when one of the firm’s earliest and flagship funds—targeting the Chinese CSI 500 stock index—outperformed the index by 50%, posting an annual return of 71% thanks to its use of an AI-powered prediction model that forecast which stocks would perform better. A graph of the firm’s history in the deck showed an upwards-sloping line with a picture of a rocket next to the year 2020, noting that High-Flyer was “looking towards the future.”
But the funds’ performance started to decline in late 2021, partly due to its AI mistiming trades on the market. High-Flyer closed new subscriptions to its funds in November that year and an executive apologized on social media for the poor returns a month later. The firm then put about $55 million of its own cash into its funds in January 2022, and by the end of the year it had recovered, with High-Flyer’s three main funds posting returns higher than 15% compared to a more than 20% loss for the broader market in 2022.
Then came DeepSeek. In April 2023 the firm announced in a post on its WeChat account—titled “High-Flyer’s New Journey”—that it had established a new firm named DeepSeek to build an artificial general intelligence (AGI) model. Liang funded DeepSeek himself, in part with High-Flyer proceeds, and enlisted his team of mostly new grads from top Chinese universities. Liang was driven by his curiosity to push the boundaries of what AI can do, and not necessarily profit, he told 36Kr in 2023.
Around that time, perhaps realizing DeepSeek’s potential and amid the Chinese government’s regulatory crackdown on quantitative trading, Liang started to scale down High-Flyer to focus on the new firm. In October 2023, a High-Flyer representative told Chinese financial outlet Cailianshe that High-Flyer’s business would be divided into two units—artificial intelligence research and the hedge fund—which would be “managed by two parallel companies under the same controller,” a reference to Liang. High-Flyer also reduced its scale to about $6 billion in assets under management at the time.
Last year, as DeepSeek built out its models, High-Flyer adjusted its strategies and abandoned its market-neutral products, which pick both long and short positions on stocks, focusing only on long positions instead. The firm is still active—it invested $35 million of its own cash into its funds in February 2024 and its assets appear to have ticked up again—but its performance last year was middling. Out of 65 High-Flyer funds that report financials, 36 of them showed losses of up to 6% in value in 2024 while another 29 posted gains up to nearly 18%, according to Chinese financial data firm Tonghuashun.
Now, all the attention is on DeepSeek. That’s largely thanks to its claims that it trained its V3 model, released in December, with far fewer resources than the much larger OpenAI—$6 million vs. $100 million for GPT-4—and with less sophisticated chips, though the GPT-4 number could include chip and personnel costs. (A few differences in the structure of DeepSeek’s model makes it smaller, and therefore cheaper, while losing minimal accuracy: akin to writing numbers with fewer decimal places, reading in whole phrases at a time instead of word-by-word, and creating a system that divides the model’s knowledge among “experts,” only a few of which need to be active at a time.) On top of that, it claims that its reasoning model R1, released in January, can rival OpenAI’s “o1” model on tasks like coding and solving complex math problems. Notably, unlike those of OpenAI, DeepSeek’s models are open-source, meaning anybody can access the code for free.
DeepSeek made “what was once thought to be such a capital intensive process much less capital intensive,” says William Blair’s Bhatia.
For now, DeepSeek’s valuation is a moving and yet-unproven target. It will depend on whether its models continue to be as efficient and advanced going forward and at scale. Plus, with users inputting information into DeepSeek’s models that is “more sensitive, from a national security standpoint, than anything that ever will appear on TikTok,” it’s unclear how much of the global market DeepSeek will be able to tap into, per Chris Franzek, who leads financial firm Stout’s valuation services. (One of DeepSeek’s Chinese competitors, Zhipu AI, was added to a U.S. Department of Commerce list of foreign companies considered to be a national security risk on January 16, subjecting it to export controls.) Also crucial is how aggressively Liang—who reportedly started DeepSeek without big plans for commercialization—decides to pursue plans for profits.
“What we are thinking now is that we can share most of our training results publicly, so that we can integrate it with commercialization,” Liang said in 2023. “We hope that more people, even a small app, can use the big model at a low cost, rather than having the technology be monopolized by a few people and companies.”
