TheChek News Flow

Up until two months ago, DeepSeek, the three-year-old Chinese AI lab, was an anomaly in the increasingly costly global AI battle. It had relied entirely on CEO Liang Wenfeng’s personal wealth and never raised outside money.

That changed in the middle of this month, when DeepSeek completed a $7.4 billion fundraising that valued the startup at more than $50 billion, the biggest ever first-time fundraising by a Chinese startup. The change of heart was prompted, say three people familiar with the CEO’s thinking, by Anthropic’s release in April of a preview of Mythos, a new model that the U.S. company said was so powerful that it can find and exploit software vulnerabilities, opening the door to potential chaos and misuse.

The Takeaway

Anthropic’s Mythos model prompted DeepSeek’s $7.4 billion fundraising.

DeepSeek plans to double its workforce and embrace Huawei chips.

DeepSeek V4 became third-largest model on Vercel due to low cost.

Powered by Deep Research

After seeing how Mythos achieved strong capabilities from training on enormous amounts of compute and data, Liang realized DeepSeek couldn’t compete without a massive war chest. Now the firm, which currently has a staff of around 300, is expanding both its workforce and its computing capacity.

In a rare public announcement, DeepSeek on Thursday said it is planning to “at least double” the headcount in all of its departments, including AI system development, infrastructure, product development and deep learning research. The DeepSeek Harness team, which works on transforming DeepSeek AI models into autonomous AI agents, is interviewing candidates every day, team leader Tianyi Cui, who joined the company from Jane Street in March, said on X earlier this month.

The company is also intensifying its efforts to adapt to Huawei chips, in light of U.S. export controls, while continuing to train its models on Nvidia chips stockpiled via the black market.

DeepSeek’s expansion could be a watershed moment for the global AI race. The Trump administration recently banned any foreigners, including those living in the U.S. and working for Anthropic, from accessing Mythos and a neutered version of the model, Fable 5. DeepSeek, along with another startup, Zhipu, represent China’s best hope for catching up to the U.S. in AI.

DeepSeek’s recent fundraising “marks its transition from a high-efficiency research outlier into a scaled national AI platform,” said Paul Triolo, a partner at Washington-based advisory firm DGA-Albright Stonebridge Group. “DeepSeek and Zhipu are now the top two Chinese AI labs and, along with Anthropic and OpenAI, are now part of the top four frontier model AI labs in the world. The emergence of a Mythos caliber model in China is most likely to come from either DeepSeek or Zhipu,” he said. (Beijing-based Zhipu, also known as Z.ai, recently released the open-weight GLM 5.2, which boasts capabilities similar to those of OpenAI’s GPT 5.5 and Anthropic’s Claud Opus 4.8.)

Liang believes the only way for China to lead global AI development, especially given U.S. export restrictions on advanced chips, is by having a research-focused lab free from commercial pressure.

With that in mind, he has told people close to him his strategy for DeepSeek isn’t changing. It will continue to offer open-source technology, maintain low prices for its models and maintain its focus on achieving artificial general intelligence. He defines AGI as machines having human-level capabilities in understanding, reasoning, learning, planning and adapting across a broad spectrum of tasks.

The 41-year-old Liang has told associates AI shouldn’t be controlled by a handful of people. While other tech CEOs would likely make the same point, DeepSeek is the only major AI lab to make the code underlying all of its models available to the public.

This account of DeepSeek’s fundraising is based on the accounts of multiple people close to Liang, DeepSeek’s investors and former and current employees. The company has not made any announcements on its fundraising, and didn’t respond to a request for comment.

Born in 1985 in a village in Zhanjiang, a small coastal town in the southern tip of China’s affluent Guangdong province, Liang attended Zhejiang University, where he studied electrical engineering—a discipline equivalent to computer science studies in China at the time. Zhejiang University, located in the eastern Chinese city of Hangzhou, had earlier spawned the first wave of Chinese tech giants, including e-commerce powerhouse Alibaba and gaming publisher NetEase Games.

In 2015, he co-founded a hedge fund, High-Flyer Capital Management, with some university classmates. The firm became a pioneer in quantitative trading in China’s stock market. It couldn’t be learned how much wealth Liang accumulated from his trading years, but he wrote the biggest check in DeepSeek’s funding round, 20 billion yuan (about $3 billion), or two-fifths of the total amount raised.

In 2023, after ChatGPT took off, Liang set up DeepSeek as an arm of High-Flyer. The hedge fund already had a large cluster of Nvidia chips, which it had stockpiled before the U.S. cracked down on export of the chips to China. That gave DeepSeek a leg up compared to other Chinese AI firms and helped with its early success.

Liang was not initially resistant to seeking venture capital, as he understood that AI development was an expensive game. But he turned off prospective investors in 2023 by telling them DeepSeek would be dedicated to deep research and scientific exploration with no commercial or product road map. Liang ended up funding DeepSeek himself.

In January 2025, DeepSeek’s R1 achieved capabilities on par with those of OpenAI’s latest models at significantly less computing power, sending shockwaves across Silicon Valley and Wall Street. The explosion put China-made AI on the world map and activated an open-source movement in China.

But as DeepSeek’s fame has grown, its orbit has been pulled closer to Beijing. Liang has to notify authorities if he intends to travel overseas, and governments across the country provide security details as he travels domestically, according to a person close to him. The Information reported last year that some of DeepSeek’s key researchers were barred from freely traveling overseas and were asked to hand in their passports to the company.

Liang’s AGI aspiration will be complicated when DeepSeek’s models reach Mythos-like levels and the Chinese government considers how to control the release of frontier AI models with national security–related capabilities, said Triolo. “This ultimately could force DeepSeek into a much closer relationship with Chinese government authorities than Liang Wenfeng may be comfortable with, but this appears inevitable if DeepSeek’s driving mission remains AGI,” he added.

OpenAI and Anthropic have repeatedly accused Chinese AI labs including DeepSeek of “distilling” their models, or using answers from advanced models to train new models. It couldn’t be learned what Liang thinks of those allegations. People close to him said he seldom comments on competitors.

Liang has run DeepSeek as a research lab. The vast majority of its researchers, split between Beijing and Hangzhou, are graduates from top Chinese universities. Applicants without top-school credentials often struggle to get an interview, while many who do make it through describe the written tests as unusually difficult. Even interns are expected to have published influential research papers.

Bureaucracy is thin. DeepSeek has no dedicated human resources or public affairs departments, and every researcher reports to Liang himself. He remains hands-on and attends most meetings on research discussion. As part of the company’s current workforce expansion, it is hiring for human resources, finance, legal, procurement and administrative positions.

The company’s office signages in Beijing and Hangzhou are so subtle that if visitors are not familiar with DeepSeek’s blue whale logo, they can easily miss the entrance.

Researchers work together. They rarely go out to lunch. Instead, they order either takeout from KFC or Chinese meal boxes and eat together in the office.

That said, Liang, an outdoorsy man who enjoys cycling across nearby cities and country roads, encourages his staff to have a work-life balance—almost unheard of in China’s tech industry, where employees at the biggest companies have to endure long working hours. He believes any individual’s optimal productivity can only last six to eight hours a day, and there’s no point in toiling away beyond that.

Nonetheless, about two dozen DeepSeek researchers have left the company since the lab shot to global stardom in early 2025. Higher salaries and higher rewards from employee stock options lured most of them away to companies including Alibaba, ByteDance and Tencent. The most notable departure was Guo Daya, a key contributor to DeepSeek’s previous models who’s now leading ByteDance’s AI coding effort.

Some of DeepSeek’s former researchers see a tinge of hypocrisy in Liang’s AGI vision and not-for-profit drive, as the CEO has accumulated substantial wealth himself from his quant-trading years. Those who have stayed take comfort in the recent fundraising, as the company has set up an employee stock ownership plan, which distributes shares with an actual valuation.

Embrace Huawei

Liang believes that it’s only a matter of a few years before Huawei chips will become as good as Nvidia’s and that DeepSeek should move first by adapting to semiconductor hardware outside Nvidia’s dominance. Huawei only began working directly with DeepSeek last year after learning of the company’s experiments with Huawei’s chips. Huawei didn’t respond to a request for comment.

The attempt to adapt to Huawei chips prolonged DeepSeek’s model release last year. DeepSeek had built its training and deployment systems around Nvidia’s Cuda software, and engineers had to rework the software so the model could use the Huawei chips efficiently and run at a competitive speed and cost. As a result, the company didn’t release any new-generation model for 15 months, a conspicuously long gap in an era when other top-tier developers release new models every couple of months.

The gap also made DeepSeek late to the coding frenzy Anthropic’s Claude Code tools unleashed in the second half of last year. That didn’t faze Liang, who told investors during a road show that the coding tool was, like AI chatbots, just temporary products in the AI evolution. If DeepSeek were to bet heavily on these short-term products, that would divert it from the ultimate goal of attaining AGI, he argued.

In the U.S., DeepSeek is gaining popularity among developers. The company launched its latest flagship model, V4, in April. V4’s shares of token usage on AI Gateway, the model aggregator platform of U.S. startup Vercel, jumped in May from under 1% to 17% in a single month, making it the third-largest model provider after Anthropic and Google on that platform. The growth has continued into June, according to Vercel. DeepSeek V4 Flash, the affordable lightweight version of V4, is 20 to 50 times cheaper than Anthropic models.

“We’re seeing increasing pricing sensitivity among customers. Teams are routing inexpensive, capable models to lower-risk work while continuing to use frontier models for high-stakes tasks,” said Harpreet Arora, head of AI infrastructure at Vercel.

DeepSeek’s continued rise will inevitably draw more attention and scrutiny from Washington. “The U.S. government is likely considering restrictions on the use of Chinese open-weight models for specific applications in the U.S., such as critical infrastructure,” Triolo said. “But the company’s models will remain popular in markets sensitive to price where there are lower or no geopolitical or national security considerations.”