DeepSeek is taking the spotlight.
This week, the Chinese AI research entity DeepSeek made waves in the tech world as its chatbot application skyrocketed to the top of the Apple App Store charts, with a similar ascent on Google Play. The company’s AI models, developed through compute-efficient strategies, have prompted Wall Street analysts and technologists alike to ponder the sustainability of the U.S. lead in the AI sector and the continuing demand for AI hardware.
But what is the story behind DeepSeek, and how did it attain such rapid global recognition?
The Trading Roots of DeepSeek
DeepSeek is financed by High-Flyer Capital Management, a Chinese quantitative hedge fund that leverages AI for its trading strategies.
Liang Wenfeng, an AI enthusiast, co-founded High-Flyer in 2015. His journey in trading began as a student at Zhejiang University. In 2019, he established High-Flyer Capital Management, a hedge fund aimed at creating and implementing AI algorithms.
In 2023, High-Flyer introduced DeepSeek as a lab focused on researching AI technologies independently from its financial operations. With High-Flyer included among its investors, DeepSeek became an independent entity.
From its inception, DeepSeek has constructed its own data centers for model training. However, like many Chinese AI firms, it has faced challenges due to U.S. export restrictions on hardware. For one of its latest models, the company had to resort to using Nvidia H800 chips, which are less powerful compared to the H100 model available to U.S. enterprises.
DeepSeek’s technical team is reportedly quite young, and the company actively recruits PhD holders in AI from leading Chinese universities. Additionally, DeepSeek also employs individuals without formal computer science training to enhance the model’s comprehension across diverse fields, as noted by The New York Times.
DeepSeek’s Powerful Models
DeepSeek launched its initial model lineup—including DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat—back in November 2023. However, it was the release of the next-generation DeepSeek-V2 models in spring that caught the industry’s attention.
DeepSeek-V2 is a versatile text and image analysis system that performed admirably in multiple AI benchmarks, proving to be significantly more economical to operate than similar models available at the time. This forced domestic competitors like ByteDance and Alibaba to lower their model usage fees and even offer some for free.
The launch of DeepSeek-V3 in December 2024 further bolstered the company’s reputation.
Internal benchmarks from DeepSeek indicate that DeepSeek V3 surpasses both downloadable models like Meta’s Llama and “closed” models accessible only via APIs, such as OpenAI’s GPT-4o.
Moreover, DeepSeek’s R1 “reasoning” model, introduced in January, reportedly matches OpenAI’s o1 model on crucial performance metrics.
R1, as a reasoning model, self-verifies its answers, mitigating common errors that many models encounter. Although reasoning models might take slightly longer to deliver results—usually from seconds to minutes—they tend to be more reliable in areas like physics, science, and mathematics.
However, there’s a caveat concerning R1, DeepSeek V3, and other models developed by DeepSeek. Being AI crafted in China, these models are subject to regulatory benchmarking by China’s internet authorities to ensure compliance with “core socialist values.” In its chatbot application, for example, R1 is restricted from responding to inquiries about sensitive topics like Tiananmen Square or Taiwan’s sovereignty.
An Innovative Strategy
The specifics of DeepSeek’s business model remain ambiguous. The company consistently prices its offerings below market standards, with some provided at no cost.
DeepSeek attributes its remarkable cost-effectiveness to breakthroughs in efficiency, although certain experts challenge the accuracy of the company’s financial assertions.
Regardless of the underlying details, developers are embracing DeepSeek’s models, which, while not open-source as traditionally defined, come with permissive licenses that facilitate commercial application. Clem Delangue, CEO of Hugging Face, noted that over 500 “derivative” models based on R1 have been created on the Hugging Face platform, accumulating a total of 2.5 million downloads.
DeepSeek’s achievements against larger, established competitors have been termed as “disrupting AI” and “over-hyped.” The company’s rise somewhat contributed to Nvidia’s stock dropping by 18% in January, and even drew a public reaction from OpenAI CEO Sam Altman.
Microsoft has announced the integration of DeepSeek into its Azure AI Foundry service, a platform consolidating AI services for businesses. When asked about DeepSeek’s potential impact on Meta’s AI expenditure in a recent earnings call, CEO Mark Zuckerberg articulated that continued investment in AI infrastructure will remain a “strategic advantage” for Meta.
During Nvidia’s fourth-quarter earnings briefing, CEO Jensen Huang applauded DeepSeek’s “exceptional innovation,” emphasizing that models like DeepSeek’s require significantly more computational resources.
However, some organizations and even entire nations and administrations, such as South Korea, have imposed bans on DeepSeek. Moreover, New York state has also barred the use of DeepSeek in state government devices.
As for the trajectory of DeepSeek, its future is somewhat uncertain. While advancements in its models are expected, the U.S. government appears to be growing increasingly cautious regarding what it perceives as detrimental foreign influences.
TechCrunch offers a dedicated AI newsletter! Subscribe here to receive it in your inbox every Wednesday.
This article was initially published on January 28, 2025, and is subject to continuous updates.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
