The training only costed less than 6 million dollars, while OpenAI spent 63 million. The user cost is also significantly low, with the access to its most powerful model costs 95% less than its competitors. The design of Deepseek model emphasizes modularity and efficiency, providing free API access to developers. (APIs, Application Programming Interfaces, tools that allow developers to integrate Deepseek’s model into their own applications). Besides using API tools, as the model is completely open-sourced, developers can freely access, develop and deploy its model based on their specific needs – unlike models such as ChatGPT, which only provide limited access through paid APIs. Users can even download Deepseek models to local or private servers, which prevents the privacy risks associated with cloud-based models. The local server can also avoid latency caused by network transmission and allows the model to be used even without an internet connection. Another factor behind the strong market reception of Deepseek is that the model breaks down the thought process of AI and present it transparently to the users. Previously, users were bothered by models making up false responses, while this function allows fact checking easier. This cost-effective product has been greatly recognized by the market. On January 28, Deepseek become the No.1 downloaded free app on Apple’s iPhone store.

How Did Deepseek Achieve Success?

There are many factors that contributed to the success of this model. The CEO of High-flyer Capital (owner of Deepseek), Wenfeng Liang, earned his degree at Zhejiang University studying electronic information engineering. Driven by the obsession with AI, Liang and his High-flyer Capital developed AI to spot patterns in stock price, which made huge amount of money, sufficient enough to support his ambition in exploring AI. Instead of business logic, the decision to step into AI market was purely out of curiosity as Liang told the media. He claimed that the development of their AI is not for profit, but to “be at the forefront of technology and promote the development of the entire ecosystem”. 

It is a common sense that the development of AI requires numerous high-performance chips. US and China are having competitions in almost all technological areas, and in order to slow China down in the race of developing AI, since 2022, the US issued export control on high-performance semi-conductors to China, stopping the country from accessing advanced equipment necessary for AI development. Since then, the limited access to high-performance chips had deeply restricted the development of China’s AI industry.

However, these restrictions have instead forced China’s AI industry to explore how to build efficient AI models with limited computation power and money. Traditionally, the training of AI model requires huge amount of external data, but Deepseek R1 used its previous model to generate data, and use them to train the latest R1 model, which requires less data input. Most importantly, Deepseek adopted Mixture of Experts (MOE) and Knowledge Distillation method. MOE structure significantly reduces computational resources requirements while maintaining good performances. With a total of 671 billion parameters, the model only activates around 37 parameters per task, since the model select the most suitable expert modules based on the specific task. Additionally, as different experts in the model specialize in different tasks, it is very professional in handling field-specific problems. Knowledge Distillation is also essential in reducing the cost. It allows Deepseek to use the existing LLM (such as ChatGPT) to guide the training of Deepseek’s own model, which directly learns the distilled from the teacher model, which eliminate the need for redundant trial and error. Imagine teaching someone to drive, the traditional way is to learn by crashing 10 cars through trial and error, while this “knowledge distillation” way is to master in 3 tries with a trainer’s simulated demo. The computation cost for Deepseek V3 is only 2.8 million GPU hours, while the cost for majority of other models is more than 30 million GPU hours.

No matter how efficient their strategy was, the requirement for a moderate amount of chips is still inevitable. Liang told media that the biggest challenge they encountered was never about funding, but rather securing access to high-performance GPUs. Fortunately, before Biden’s prohibition took effect, Liang and his company had already collected 10000 GPUs in anticipation for future need. Though the chips they used were merely Nvidia’s H800 (designed specifically for Chinese market, which has limited computation power), while OpenAI used 25000 pieces of Nvidia’s A100 (two times better than H800), their efficiency in strategy bridged this gap.

Another interesting fact about Deepseek is that Liang himself also personally involved in the research. Typically, employer would favor candidates with computer science degree from the US, but Liang is reported to prefer local talents. Deepseek is also known for offering significantly higher salary compared to the industry average. It provides salaries of up to 60000 RMB (around 8000 Euros) per month, while the average salary at Tencent is about 35000 RMB. With a total number of around 200 employees, the generous compensation for each employees contributes to its high efficiency.

Impact of Deepseek’s Success

The success of Deepseeks has great significance to the global AI industry. The CEO of OpenAI, Sam Altman once claimed that it is hopeless for small companies to compete with them in AI, while the success of Deepseek’s R1 model has refuted this claim. Previously, the entire industry fostered an atmosphere that the development of advanced AI model requires massive funding and numerous high-performance semi-conductors, while Deepseek has just challenged it to be a lie that deters the startups from joining the competition. Meanwhile, from a technological perspective, it makes the industry to reconsider the path for developing AI.

The free access and open-source characteristics further promote the democratization process in AI industry, as it forces its competitors to reprice their services, more and more people can benefit from the convenience brought by technology. Meanwhile, this breakthrough has also led people to reconsider whether the enormous market demand for high-performance semi-conductors has been over estimated, which is one of the reasons behind the sharp decline in Nvidia and the US tech stocks.

From the geopolitical perspective, as the Guardian has reported, the success of Deepseek is a “Sputnik” moment. The arms race between US and China in technology had undergone for a while, both sides tried to slow their opponents down in developing core technologies. The breakthrough made by Deepseek symbolizes the failure of US’s attempts to stifle China’s AI development through semi-conductor prohibition. It collapses the AI domination established by the US and several giant companies, providing a cheap replacement, which reduces the global reliance to few companies or a country. As developers’ worldwide download Deepseek’s codes and startups emulate its strategies, one question comes up: if a 200-person team in Hangzhou can shake the Wallstreet, what’s next? There will be more Sputnik moment coming. 

By Xingchen Liu