On January 27, Chinese AI startup DeepSeek burst onto the global scene, dominating media coverage and sparking intense discussions across the AI innovation community. What exactly is DeepSeek? What makes it unique? How could it shape the future of global AI, and what implications does it have for China's overall AI development?
In this article Professor Fang Yue, EVE Energy Chair in Economics and Decision Sciences and Director of the Research Centre for AI and Management Innovation at CEIBS, answers these questions and more.
As most readers are likely aware, DeepSeek recently launched its R1 reasoning model, building on the DeepSeek V3 base model, which shocked the world with its ability to match its far more famous rival, OpenAI's latest model o1, in quality of output. By virtue of both their quality and their cost efficiency, Deepseek's models have been praised by Silicon Valley executives and US tech company engineers alike. On the day of DeepSeek R1's release, DeepSeek became the most downloaded free app on Apple's iPhone store in both China and the US.
On the same day, US tech stocks took a beating. Nvidia (NVDA), which has benefited greatly from the use of its chips in the vast infrastructure that many assumed was needed for the effective utilisation of AI technology before DeepSeek's more cost-efficient product was released, lost about 17% or $588.8 billion in market value. Tech giants such as Meta (META), Alphabet (GOOGL), Marvell, Broadcom, Palantir, and Oracle also saw their shares drop sharply, sending the tech-heavy Nasdaq down 3.1%.
The implication that DeepSeek's new technology could make AI models more energy-efficient also worried investors and sent some energy stocks tumbling. GE Vernova, which produces wind power and gas turbines, for example, plunged 21%. Vistra, another power generation company, plummeted 28%. Over the subsequent few trading days, shares of Nvidia and other major companies continued to fluctuate wildly as the market weighed up DeepSeek's potential impact on US stocks and the wider AI sector.
Fundamentally, DeepSeek's high profile arrival on the global stage raised investor doubts about US companies' dominance in AI, calling into question big tech's massive spending on building AI models and data centres.
01
What is DeepSeek?
Headquartered in Hangzhou, China, DeepSeek was established in July 2023 by Liang Wenfeng, a graduate from the Department of Information and Communication Engineering, Zhejiang University. It grew out of High-Flyer, a hedge fund Liang founded in 2015.
In March 2023, High-Flyer announced on its official WeChat account that it was "restarting" to create a new and independent research group dedicated to exploring the application of Artificial General Intelligence (AGI) beyond trading.
This initiative led to the establishment of DeepSeek later that year. Although it remains unclear how much High-Flyer has invested in DeepSeek, according to existing publicly available company information, High-Flyer has an office located in the same building as DeepSeek and owns patents related to chip clusters used to train AI models.
For ambitious tech firms to success, it is essential that they have a clear vision. Like OpenAI, DeepSeek ultimately aims to achieve AGI. And, like other prominent AI startups, including Anthropic and Perplexity, it has released multiple competitive AI models over the past year. It was, however, relatively unknown and captured little industry attention.
Until now. This year, DeepSeek R1's impressive performance, coupled with the company's Chinese origins and relative obscurity, has rocked the global AI community and caused panic in the US stock market.
02
What makes DeepSeek different?
Using an innovative framework, DeepSeek has made remarkable breakthroughs in AI algorithms and developed competitive models with capabilities on par with ChatGPT-4. Notably, the model performance described in the company's research papers has largely been validated by independent benchmark tests.
Instead of following a distillation approach (where smaller AI teams with limited funds build specialised models based on existing large models), DeepSeek adopts a groundbreaking, generalisable method. Specifically, it uses a "mixture-of-experts" system that splits its models into sub-models, each specialising in a particular task, and continuously refines them.
In addition, DeepSeek's reasoning architecture eliminates the need for supervised fine-tuning (SFT) - a crucial optimisation technique in which a pre-trained model is further trained on labelled data to improve task-specific performance.
As demonstrated by DeepSeek R1, this approach, based on combining algorithmic optimisation and new architecture, can overcome computational constraints and even increase data utilisation efficiency and iteration speed, ultimately leading to a substantial reduction in development costs.
While DeepSeek R1 may not beat the most advanced AI model in Silicon Valley, it represents a technical path that the industry has yet to explore, characterised by low cost and high response speed, earning a reputation as "an excellent AI advancement" (as it was described by an Nvidia spokesperson on January 27, 2025). With a talented team, high-quality training data, and exceptional innovation, DeepSeek offers a prime example of how to do more with less in terms of funding and resources.
As Perplexity CEO Arvind Srinivas recently said in an interview with CNBC: "The reason it (DeepSeek R1) is so eye-opening and the reason that a lot of researchers have been compelled by what DeepSeek has achieved is that it is a really interesting step in terms of how fast you can get to the frontier or close to the frontier with as little capital as has been claimed here."
Last but not least, training AI large language models (LLMs) at a lower cost is an inevitable industry trend. Through a series of technical efficiency innovations, DeepSeek has managed to deliver capabilities comparable to industry leaders at a fraction of the cost. It is worth noting at this point that DeepSeek's total training costs for its V3 model and the equally acclaimed R1 model have not been disclosed, and since AI training costs are dropping significantly every year it may not be entirely fair to compare DeepSeek's training expenditure to some tech giants' tremendous early-stage development expenses.
03
How could DeepSeek shape the future of global AI?
The scaling law
The "scaling law" is a key concept in large model development introduced by OpenAI researchers in 2020. It posits that enhancing AI systems inevitably involves increasing computational power and data input, thereby necessitating additional chips and high-quality data.
Sam Altman, CEO of OpenAI, last year said the AI industry would need trillions of dollars in investment to support the development of high-in-demand chips. DeepSeek's arrival challenges this assumption, having achieved capabilities similar to the most powerful AI models for a small fraction of the cost and on less powerful chips. This raises two key questions. Does the scaling law still apply? And should we continue to invest so heavily in expensive AI infrastructure?
I believe that the future of AI will be driven by energy efficiency and cost-effectiveness, but that demand for computing power and data will remain high. DeepSeek R1's success could prompt Wall Street to broadly reassess the AI industry, raising questions as to whether Nvidia's stock is overvalued and whether we really need quite so many AI data centres.
DeepSeek's algorithmic innovations will undoubtedly disrupt the AI supply chain. While DeepSeek may challenge the dominance of established US companies like OpenAI, access to powerful chips and computing power will continue to be indispensable advantages in the AI space. Going forward, AI companies will likely explore a number of different development paths.
Inspiring the open-source AI community
In contrast with closed-source models like OpenAI's ChatGPT, DeepSeek is open-source. It's no exaggeration to say that DeepSeek's success represents a significant breakthrough and a victory for the open-source AI community.
Over the past two years, a growing number of Chinese companies have embraced open source technology. Alibaba Cloud, for example, has released over 100 new open-source AI models, supporting 29 languages and catering to various applications including coding and mathematics.
Similarly, startups like Minimax and 01.AI have open-sourced their models. In this sense, DeepSeek's success could potentially reshape the interplay between open- and closed-source AI development.
While exciting the AI community, DeepSeek has also attracted the attention of open-source competitors. According to The Information, Meta has set up several "war rooms" with engineers to figure out how DeepSeek was made so efficient. This could fuel a collaborative effort to drive the development of open-source models, ensuring that they develop hand-in-hand with closed-source models.
The impact on AI implementation
It is still unclear what DeepSeek means for the AI industry, but I believe it will play a pivotal role in advancing AI applications. I also believe that the potential long-term revenues and costs associated with large AI models will decrease.
We must note, however, that training an AI LLM on a smaller budget is one thing; deploying it at scale to create societal value is quite another. Meeting the huge demand for AI technology will still require a lot of infrastructure and time, too.
04
What implications does DeepSeek have for China's AI developments?
According to a 2024 white paper by the China Academy of Information and Communications Technology, the number of AI LLMs worldwide has reached 1,328, with 36% originating in China. This makes China the second-largest contributor to AI development, after the US.
Over the past two years, Chinese companies have rolled out many ChatGPT-like products. Nevertheless, most of these products have met a lukewarm reception. DeepSeek's success is a shot in the arm, boosting the confidence of innovation-driven Chinese AI companies and the industry as a whole.
Building a basic AI LLM can be resource-intensive and technically complicated. In a July 2024 interview with Chinese media outlet 36Kr, Liang stated that in addition to chip sanctions, another challenge for Chinese AI companies was engineering efficiency issues. "We (most Chinese companies) need to consume twice as much computing power to achieve the same effect. Furthermore, there is a data efficiency gap, meaning we may need to consume four times more computing power. What we have to do is keep closing those gaps." DeepSeek could offer invaluable lessons for other AI companies in their choice of innovative engineering strategies.
From my perspective, DeepSeek's R1, much like OpenAI's ChatGPT, is a key prelude to the AI era. "The models they (DeepSeek) built are fantastic, but they aren't miracles either," said Bernstein analyst Stacy Rasgon. "They're not using any innovations that are unknown or secret or anything like that. These are things that everybody is experimenting with. The reason why DeepSeek can take the market by storm is because it does more with less."
Believe it or not, the very human drive to accomplish more with fewer resources will translate into amazing AI advancements in the coming years.
Prof. Fang Yue is Professor of Economics and Decision Sciences, EVE Energy Chair in Economics and Decision Sciences, Department Chair of Economics and Decision Sciences, and Research Area Director of AI-powered Enterprise and Management at CEIBS.
Since the beginning of 2024, he has headed the AI and Management Innovation Research Centre at CEIBS, which he established. The centre focuses on the impact of AI on business management and industrial development, and how to build AI-driven organiations. It is committed to creating an AI industry-academic-research platform with business school characteristics, as well as a high-end think tank for AI and management innovation.
Recommended Reading
How can leaders spread CSR values effectively within a company? |
An ultimate guide for businesses expanding to Southeast Asia |
Want to learn more about CEIBS programmes? Click 'Read more' below
No comments:
Post a Comment