DeepSeek has spread.
DeepSeek, a Chinese AI lab, introduced the top of the Apple App Store rankings this week (and Google Play), breaking into mainstream consciousness this week. DeepSeek's AI model is trained using computationally effective technology, which has led Wall Street analysts and technicians to question whether the United States can maintain its leadership in the AI competition and whether the demand for AI chips can be maintained.
But where did DeepSeek come from and how did it so quickly rise to international reputation?
DeepSeek is backed by Gaofei Capital Management, a quantitative hedge fund in China that uses AI to inform its trading decisions.
AI enthusiast Liang Wenfeng co-founded High-Flyer in 2015. Wenfeng reportedly began to get involved in trading, and a student at Zhejiang University launched Advanced Capital Management as a hedge fund in 2019 with a focus on developing and deploying AI algorithms.
In 2023, High-Flyer started with DeepSeek, a lab dedicated to researching the separation of AI tools from financial businesses. The lab uses High-Flyer as one of its investors to spin into its own company, also known as DeepSeek.
From day one, DeepSeek built its own data center cluster for model training. But like other AI companies in China, DeepSeek has been affected by the U.S. export ban. To train one of its latest models, the company was forced to use the NVIDIA H800 chip, a smaller-functioning version of the chip H100 that the U.S. company can use for U.S. companies.
TechCrunch Events
Berkeley, CA | June 5
Book nowIt is said that DeepSeek's technical team is biased towards Young. The company is reportedly actively recruiting doctoral degrees from AI researchers from top Chinese universities. DeepSeek also hires people without any computer science background to help its technology understand the New York Times better.
DeepSeek launched its first set of models in November 2023 - DeepSeek Encoder, DeepSeek LLM and DeepSeek Chat.
DeepSeek-v2 is a universal text and image analysis system that performs well in various AI benchmarks - and runs much cheaper than comparable models at the time. It forces DeepSeek's domestic competition, including Bytedance and Alibaba, lowers the use price of some models and gives others complete freedom.
The DeepSeek-V3 was launched in December 2024 and has only joined DeepSeek's infamous.
According to DeepSeek's internal benchmarks, DeepSeek V3 is superior to downloadable, publicly available models such as Meta's Llama and "closed" models, and is accessible only through APIs, such as OpenAI's GPT-4O.
Also impressive is DeepSeek's R1 "inference" model. DeepSeek claims R1 was released in January, and OpenAI's O1 model is in key benchmarks.
As an inference model, R1 effectively performs fact checks, which helps it avoid some of the traps that usually trip over the model. Compared to typical non-disputed models, the inference model takes longer (usually longer to minutes) to reach the solution. The advantage is that they tend to be more reliable in fields such as physics, science, and mathematics.
However, there is one drawback to the other models of R1, DeepSeek V3, and DeepSeek. As China's developing AI, they are benchmarked by Chinese Internet regulators to ensure that their responses “reflect core socialist values.” For example, in DeepSeek's chatbot app, R1 won't answer questions about Tiananmen Square or Taiwan's autonomy.
In March, DeepSeek had more than 16.5 million visits. “(f) or March, DeepSeek is second, despite daily visits, despite traffic volumes falling 25% from February,” David Carr, editor of SamelyWeb, told TechCrunch. It still pales compared to Chatgpt, which soared 500 million active users a week in March.
If DeepSeek has a business model, it is not clear what the model is. The company puts its products and services at prices well below market value and offers them to others for free. Despite the risks, it also doesn't cost money.
The way DeepSeek tells is that efficiency breakthroughs allow it to remain extremely cost-competitive. However, some experts have objected to the figures provided by the company.
In any case, developers have adopted the DeepSeek model, which is not open source, as the phrase is usually understood but is available under a loose license that allows commercial use. One of the platforms that host DeepSeek models, developers on Hugging Face, have created more than 500 "derived" models of R1 that have added 2.5 million downloads, said Clem Delangue, CEO of Hugging Face.
DeepSeek's success for larger, more established competitors has been described as "high-profile AI" and "super-promotion." The company's success was at least in part due to the 18% drop in NVIDIA's share price in January and triggering a public response from Openai CEO Sam Altman. In March, the U.S. Department of Commerce told staff members that DeepSeek would be banned on their government equipment, Reuters reported.
Microsoft announced that DeepSeek can be available on its Azure AI Foundry Service, Microsoft's platform, which aggregates AI services into enterprise services under a single banner. When asked about the impact of DeepSeek on Meta’s AI spending in the first quarter earnings call, CEO Mark Zuckerberg said spending on AI infrastructure will continue to be Meta’s “strategic advantage.” In March, Openai called DeepSeek a "state subsidy" and "state control" and suggested that the U.S. government consider banning the use of DeepSeek's model.
During NVIDIA's fourth-quarter earnings call, CEO Jensen Huang highlighted DeepSeek's "excellent innovation", which he said and other "inference" models are very useful for NVIDIA because they need more computing.
Meanwhile, some companies are banning DeepSeek, as are the entire country and government, including South Korea. New York State also bans DeepSeek from being used in government equipment.
Microsoft vice chairman and president Brad Smith said at a Senate hearing that Microsoft employees are not allowed to use DeepSeek due to data security and publicity issues.
As for what the future of DeepSeek may be, it is not clear. The improved model is given. However, the U.S. government seems to be increasingly alert to foreign influence it considers harmful. In March, the Wall Street Journal reported that the United States may ban DeepSeek on government equipment.
This story was originally published on January 28, 2025 and will be updated regularly.