Recently, the AI startup DeepSeek, which has only been around for two years, finds itself at a crucial crossroads, confronting a myriad of significant decisions impacting its future directionThe surge in popularity of its chatbot, DeepSeek ChatBot, a few weeks ago has led to an urgent demand for computational power expansion, prompting internal discussions on whether to accept outside financing in order to acquire more AI chips and server resources.

According to two individuals with direct knowledge of the situation, notable investors such as Alibaba Group, the China Investment Corporation (CIC), and the National Social Security Fund have approached DeepSeek, expressing a strong interest in investingThis trend underscores the recognition of DeepSeek's potential in the AI field by the capital markets.

It’s noteworthy that prior to this, DeepSeek and its parent company, Huanshe Quantitative, maintained a cautious stance, turning down multiple requests for cooperation from venture capitalistsThe company long adhered to a "pure research-oriented" operational model, focusing on deep technological development while being reserved about the involvement of commercial capitalHowever, with the rapid growth of its operations, the bottlenecks associated with computing power have become increasingly prominent, compelling the company to reevaluate its financing strategy and consider a strategic shift from focusing solely on technical research to commercial operations.

Leading DeepSeek is its founder and CEO, Liang Wenfeng, affectionately dubbed a "tech geek" by his colleaguesHe idolizes the tech founder of Renaissance Technologies, Jim Simons, driven by pure passion and an unwavering pursuit of technologyIn 2015, Liang founded Huanshe Quantitative, achieving notable success in the field of AI quantitative tradingWith the global AI revolution sparked by ChatGPT in 2023, Liang astutely recognized the immense potential in General Artificial Intelligence (AGI). He decisively established DeepSeek as a standalone AGI research lab, operating in an academic-like framework.

In the startup’s early days, Liang leveraged around 10,000 Nvidia A100 chips previously reserved by Huanshe Quantitative, granting DeepSeek a crucial first-mover advantage

Advertisements

Under his leadership, the team was singularly focused on tackling core challenges in the AI space through rigorous technical researchLiang's idealistic approach and steadfast commitment to technological fundamentals cultivated a vibrant atmosphere for innovation, enabling DeepSeek to achieve a remarkable series of outcomes in technological development.

In December 2023, the release of DeepSeek’s V3 model sent shockwaves through the industryBuilt using only 2048 chips and incurring a training cost of $5.5 million, the model outperformed several benchmark tests, surpassing Alibaba's Qwen 2.5-72B and Meta’s Llama 3.1-405B among other open-source modelsIt even compared favorably to closed-source giants like OpenAI’s GPT-4o and Anthropic’s Claude 3.5 - SonnetThe innovative technological path and minimal training costs this model showcased present a new direction for the development of large models.

The DeepSeek-R1 inference model launched in January this year, achieved performance on par with OpenAI's models at a significantly lower costThis model excelled in mathematical, coding, and natural language reasoning tasks, showcasing vastly improved reasoning abilities and delivering comprehensive, well-argued answersMoreover, the competitive pricing of DeepSeek-R1 API services has further increased its market attractivenessWithin just a month of its chatbot app launch, DeepSeek amassed 30 million daily active users, setting a new record for Internet growth in China while rapidly building a vast global user base.

Despite achieving monumental technological breakthroughs, DeepSeek faces considerable challenges amid its rapid advancementDuring the Spring Festival holiday, some employees were still working overtime to resolve server overload issues, indicating that the computational power crisis is a critical constraint on the company’s growthConcurrently, geopolitical pressures loom over DeepSeek like dark clouds.

The U.S. government has begun to view DeepSeek as a national security threat, with plans to legislate against its applications on government devices

Advertisements

Earlier, NASA had prohibited its employees from utilizing DeepSeek’s AI technology and blocked its systems from accessing the DeepSeek platform, citing concerns over servers operating outside U.S. territory which pose national security risks and privacy issuesTexas has become the first state to ban DeepSeek applications from government-issued devicesAdditionally, data protection authorities in countries like Italy and Ireland have demanded information from DeepSeek regarding its handling of personal data, with Australia and South Korea also announcing similar bans on DeepSeek's use in government systemsSuch geopolitical pressures raise alarm for DeepSeek, especially in light of last month’s inclusion of another AI company, Zhipu Zhihua, with state-owned backgrounds, on the entity list, signaling potential repercussions for DeepSeek should it pursue state-owned investment.

Currently, DeepSeek’s primary revenue stems from charging for API interfaces, priced at just 1% of OpenAI’s ratesWhile this pricing strategy aids in garnering more users, it concurrently restricts the company’s profit marginsWeChat has begun testing integration of its model, hinting at glimmers of hope for commercial prospects, but overall, the journey towards commercialization remains fraught with uncertainties.

On the computational front, the company plans to enhance performance by obtaining higher-density chips through a data center in Malaysia as a means to alleviate computational stressNevertheless, the latest U.S. export control orders present substantial challenges to its supply chain, drastically increasing the difficulty and cost of securing the necessary chipsMoreover, Huanshe Quantitative, its parent company, has faced intense competition in the quantitative sector, with recent performance results revealing that out of 65 funds managed by Huanshe, 36 have seen declines this year, while only 29 managed to register over a 10% increase

Advertisements

Advertisements

Advertisements