All about DeepSeek: its advantages, disadvantages and impact on the China-US tech war

The emergence of China’s DeepSeek AI has sent shockwaves through the tech world, posing a direct challenge to US strategic competitiveness and potentially triggering a new wave of tariffs and export controls.

Interviewed academics pointed out that great power competition in AI has entered a new normal — each breakthrough product not only reshapes market dynamics but could also risk escalating competition and policy confrontations between nations.

DeepSeek’s unveiling of its reasoning model DeepSeek-R1 in late January, which reportedly outperforms OpenAI’s latest models in areas like training costs and testing performance, has sparked widespread interest within the tech community due to its disruptive implications.

Amid the wave of praise, a wave of criticism from American political and business sectors has also followed swiftly.

According to reports from Reuters and Bloomberg, US President Donald Trump’s commerce secretary nominee Howard Lutnick criticised DeepSeek during his confirmation hearing on 29 January for allegedly stealing American technology to create “dirt cheap” AI models. 

He said, “They stole things. They broke in. They’ve taken our IP. It’s got to end… we must stay in the lead.”

In September 2022, the Joe Biden administration ordered chip giant Nvidia to halt exports of its advanced graphics processing units (GPUs) to China in an effort to curb the country’s AI development.

During the hearing, Lutnick accused DeepSeek of circumventing US export controls to obtain large quantities of Nvidia chips, stating that he would take a “rigorous” stance on the issue. 

Lutnick likened export controls unbacked by tariffs to a whack-a-mole model. He asserted, “We’ve got to find a way to back our export controls with tariffs. I do not believe that DeepSeek was done all above board. That’s nonsense.”

Citing people familiar with the matter, Bloomberg reported that the Trump administration is considering additional curbs on the sale of Nvidia chips to China, potentially expanding restrictions to cover Nvidia’s H20 chips. The discussions are said to be in very early stages.

Introduced in July 2024, Nvidia’s H20 chips are a scaled-down product designed to meet existing US curbs on shipments to China. Since 2022, the US government has restricted the export of Nvidia’s advanced H100 and A100 chips to China. Nvidia then introduced the less powerful H800 chips to comply with the regulations, but they also became included in the export ban in 2023.

In the corporate sphere, sources have revealed that OpenAI has obtained evidence which suggested that DeepSeek used OpenAI’s data through “distillation” when training its models, thereby violating its terms of service.

Dataset distillation is a common technique in the AI industry, involving the use of algorithms and strategies to denoise, reduce dimensions and refine raw, complex data into more concise and useful data for training models. The performance of models trained by this distilled dataset can match or even surpass those trained on the original dataset.

Associate Professor Zhu Feida of Singapore Management University (SMU)’s School of Computing and Information Systems stated in an interview with Lianhe Zaobao that OpenAI has never fully disclosed its technology, opting instead to allow developers access through an application programme interface (API), which meant that DeepSeek can only access OpenAI’s public resources.

He added, “Strictly speaking, all advancements in AI technology are built progressively on the foundations laid by predecessors. Every team innovates based on previous work, so technological progress itself is a process of continual accumulation and elevation.”

Faced with a series of allegations, DeepSeek has maintained a low profile over the past few days and did not respond directly. However, major Chinese state media outlets have recently reported actively and promoted its achievements.

Bo Chen, a senior research fellow at National University of Singapore’s (NUS) East Asian Institute, explained in an interview that AI technology has become a core area for global technological competition and a key battleground in China-US rivalry.

He stated that DeepSeek’s rise posed a direct challenge to the core competitiveness of the American AI industry. This competition has transcended the realms of products and business models as the US views DeepSeek as a national security threat, escalating the competition to a strategic contest at the national level.

Chen remarked that “this is likely to become the new norm in future AI competition. When an innovative product emerges suddenly, the market will experience significant disruption, which can potentially spark disputes between countries”.

Chen opined that the current international AI competition lacks clear rules, which has led to competition that easily deviates from legal and fair competition principles. Such unregulated competition could even devolve into chaotic and inefficient confrontations.

He pointed out that fostering a healthy AI industry depends on preserving fair competition. Overemphasising technological advancement, much like the US broadening national security under its “America First” policy, risks undermining market fairness and stifling innovation.

Chen urged the international community to establish clearer rules for technological competition, avoiding its politicisation or conflation with national security. Only with fair competition and clear regulations can AI benefit all of humanity.

The Chinese company DeepSeek rapidly gained global popularity with the launch of its generative AI. Offering performance on par with leading Western models like ChatGPT at significantly lower costs, DeepSeek has stirred a growing sense of urgency among major American AI companies.

DeepSeek’s large language model, V3, is nearly 700 gigabytes — so large that it can only run on specialist hardware. It boasts 685 billion parameters, making it the largest freely available model to date. In comparison, Meta’s Llama 3.1 model has only 405 billion parameters. Tests conducted by AI coding platform Aider show that DeepSeek’s performance ranks just behind OpenAI’s o1 model.  

What is even more impressive is that DeepSeek achieves top-tier performance at an extremely low cost. Independent research firm Melius Research suggests that DeepSeek has likely mastered techniques for reducing model training costs through more efficient memory usage and learning strategies. The cost of generating information with its model is estimated to be just one-tenth that of OpenAI’s ChatGPT model.

DeepSeek’s advantage also lies in its algorithmic improvements and optimisations, which allow it to save computational power. Unlike previous models that required vast amounts of data and computing resources, DeepSeek operates efficiently with smaller input data and corpus size.

Estimates suggest that major Western AI models use around 16,000 specialised chips, while DeepSeek’s documentation reveals that it operates with only about 2,048 Nvidia H800 chips. The H800 was a China-specific version of Nvidia’s hardware, with reduced performance to comply with US export controls introduced in 2022. However, in October 2023, the H800 was also included in the US export ban on China.  

Due to China’s censorship policies, DeepSeek does not always provide unrestricted responses. When The Economist attempted to ask about Taiwan, DeepSeek generated only a few sentences before stopping, deleting its previous response, and prompting the user to “talk about something else”.

For now, US AI models still possess capabilities that Chinese competitors cannot yet match. For example, a Google research project enables users to interact with the Gemini chatbot directly through their web browsers, potentially paving the way for AI “agents” that can seamlessly interact with the internet. Meanwhile, OpenAI’s chatbot not only assists users with coding, but can also help run the code.

OpenAI is also working on multiple groundbreaking technologies, including a rumoured “PhD-level super-agent” with expertise comparable to that of human specialists.

DeepSeek’s founder, Liang Wenfeng, was born in 1985 in Zhanjiang, Guangdong province. In 2010, he earned a master’s degree in Information and Communication Engineering. After graduation, he founded Hangzhou Yakebi Investment Management Co Ltd and later Hangzhou Huanfang Technology Co Ltd, which focused on quantitative investing using mathematics and artificial intelligence. In 2023, Liang founded DeepSeek, officially entering the general AI field.

Chinese media have highlighted that DeepSeek’s success is closely tied to Liang Wenfeng’s unique strategies in team management and technological development. Liang once said his team “has no mysterious geniuses”; instead, it consists mostly of young individuals with little experience. He believes that “innovation requires breaking free from inertia, and experience can sometimes be a burden”.

This article was first published in Lianhe Zaobao as “DeepSeek或引发新一轮关税壁垒出口管制 学者：AI竞争新常态” and “DeepSeek何以震撼全球？”.

Related: OpenAI’s block will spur AI with Chinese characteristics | Chinese AI startups make gains in challenge to US-based OpenAI