DeepSeek: the Chinese AI lab that shocked Silicon Valley
DeepSeek, a Chinese AI company owned by a hedge fund, released a powerful, low-cost open-weight reasoning model in January 2025 that wiped hundreds of billions off US tech stocks. Here's its story.
First published · AI Digested
In late January 2025, a Chinese AI app few people in the West had heard of shot to the top of Apple's App Store in the US and UK. Within days, it had wiped hundreds of billions of dollars off American technology shares. The app was DeepSeek, and it changed assumptions about who could build cutting-edge AI.
A hedge fund's AI lab
DeepSeek was founded in July 2023 in Hangzhou, China, by Liang Wenfeng. Liang had co-founded High-Flyer, a successful hedge fund that used AI to trade shares, and had stockpiled thousands of Nvidia GPUs before US export controls restricted sales to China.
Unusually for a Chinese tech company, DeepSeek focused on research rather than quick products, hired mainly young graduates from Chinese universities, and published its methods openly. It doesn't rely on outside investors; High-Flyer funds it.
The models
DeepSeek released a series of increasingly capable open-weight models in 2023 and 2024. Two made headlines:
- DeepSeek-V3 (December 2024), a large language model that performed close to leading US models. DeepSeek said the final training run cost about $5.6 million in computing, a tiny fraction of what US labs were thought to spend.
- DeepSeek-R1 (January 2025), a reasoning model that "thinks" step by step, matching OpenAI's o1 on many maths and coding tests. Its weights were released under the permissive MIT licence, and it was dramatically cheaper to use.
The DeepSeek shock
The free DeepSeek app became the most downloaded app on the US App Store in late January 2025. Investors asked: if a Chinese lab could build a top model so cheaply, with restricted chips, were US companies overspending, and would demand for chips fall?
On 27 January 2025, Nvidia's shares fell about 17%, wiping roughly $590 billion off its market value, then the biggest one-day loss for any company in history. Other tech and energy stocks fell too. Most recovered in the following months.
Questions and pushback
- True costs: the $5.6 million figure excluded research, failed experiments, salaries and hardware. Analysts estimated DeepSeek's total spending in the hundreds of millions or more.
- Distillation: OpenAI said it had evidence DeepSeek may have used outputs from OpenAI's models to train its own, which would break OpenAI's terms. DeepSeek did not publicly respond in detail.
- Censorship: the DeepSeek app refuses or follows the official line on topics sensitive to the Chinese government.
- Privacy and security: DeepSeek's privacy policy says user data is stored in China. Italy's data protection authority blocked the app, and governments including Australia, South Korea and Taiwan banned it on official devices.
Why it matters
DeepSeek showed that the gap between US and Chinese AI was smaller than many assumed, and that clever engineering could partly make up for fewer chips. It also gave a huge boost to open-weight AI: R1 and its successors became some of the most widely used open models in the world, run by companies everywhere.
The episode intensified debate in Washington over chip export controls and in boardrooms over the value of hundreds of billions in AI spending.
Who runs it
Liang Wenfeng is founder and chief executive. He rarely gives interviews.
This profile is reviewed regularly. Last reviewed on the date shown above.
Read on
Common questions
Is DeepSeek safe to use?
The models themselves can be downloaded and run privately on your own computer, which keeps your data with you. DeepSeek's own app and website store data on servers in China, which has led governments including Australia, South Korea, Taiwan and Italy to restrict or ban it on official devices or investigate it.
Does DeepSeek censor answers?
DeepSeek's app avoids or follows the Chinese government line on politically sensitive topics, such as Tiananmen Square or Taiwan, as Chinese law requires. Versions of the open model run elsewhere can behave differently.
Did DeepSeek really train its AI for $6 million?
That figure was for the computing cost of the final training run of its V3 model. It excluded research, earlier experiments, staff and hardware, so the total cost was much higher. Even so, its methods were notably efficient.
Sources and further reading
Spotted a mistake or something out of date? Tell us and we will fix it. Read how we research and write our articles.
Read next
Meta AI: Llama, Meta's AI assistant and the Superintelligence Labs
How Meta became a major AI player: its FAIR lab, the open-weight Llama models, the Meta AI assistant in WhatsApp, and its costly superintelligence push.
Updated 22 November 2026
Mistral AI: Europe's answer to OpenAI
Mistral AI was founded in Paris in 2023 by former Google DeepMind and Meta researchers. How it became Europe's leading AI company, what Le Chat and its open models offer, and who backs it.
Updated 22 November 2026
OpenAI: who owns it, how it started, and what it makes
The story of OpenAI, the company behind ChatGPT: its founding as a non-profit in 2015, Elon Musk's departure, Microsoft's billions, the 2023 boardroom crisis, its restructure and who owns it today.
Updated 22 November 2026