DeepSeek’s emergence has introduced fresh challenges to US and Chinese AI firms and policymakers alike
The launch of DeepSeek’s latest open-source AI model on 20 January — which erased USD1tn from the US tech market amid fears of overheated AI investment — has caused a wide range of reactions from policymakers, academics and industry leaders. As Washington evaluates the effectiveness of its tech restrictions against China, researchers are raising alarms over DeepSeek’s security risks — from cross-border data transfers to the spread of Chinese propaganda.
What next
Subsidiary Impacts
- The share price of semiconductor firms will recover, as lower costs would lead to a wider availability of AI.
- DeepSeek’s achievement will ramp up pressure on domestic rivals to roll out new models under tighter deadlines.
- In the long run, Beijing may put guardrails around open-source AI over safety concerns.
Analysis
According to DeepSeek, training its latest DeepSeek-v3 model — which matches the coding, mathematical and reasoning capabilities of the most prominent Western rivals, including OpenAI’s latest ChatGPT-4o — only cost about USD6mn in compute cost, using Nvidia H800 chips.
Although the USD6mn is widely disputed and only covers compute outlays — and not, for instance, staff salaries and other operational costs — DeepSeek still showcases unmatched efficiency. This stands in stark contrast to Western firms, which have largely driven progress through massive upscaling of training data and compute power.
Increased self-sufficiency
Chinese companies, of course, face increasingly strict export sanctions and, despite widespread semiconductor smuggling, have to develop models that maximise relatively scarce computational resources.
DeepSeek’s breakthrough under these constraints has had a profound impact on Nvidia’s stock price, challenging the assumption that continued AI advancement depends solely on massive investment in cutting-edge chips. This, on January 27, led to the loss of more than USD500bn from Nvidia’s market capitalisation.
Nvidia’s stock, however, soon started recovering, as researchers pointed out that such concerns are likely overstated. Jevons’ paradox suggests that, as the cost of a resource decreases, its consumption increases. In other words, making AI more compute-efficient will not slow development; rather, it will likely accelerate AI adoption across a broader range of applications.
Although this trend is likely to benefit Nvidia, DeepSeek’s efficiency-focused approach could pose a serious challenge to OpenAI and other earlier-generation LLM developers. As AI models become increasingly commodified, these firms may struggle to justify the premium pricing that their more compute-intensive models demand.
However, they may still retain a reputational edge that helps sustain investment. DeepSeek has already come under heavy scrutiny, with researchers raising serious safety and security concerns.
Recent findings suggest that DeepSeek-R1 is eleven times more likely to generate harmful content than OpenAI’s o1 model. Meanwhile, the company has also faced accusations of spreading Chinese propaganda and has been heavily criticised for its data privacy policies, which require all user data to be sent to China. These worries are likely to further deter foreign investment in Chinese AI, ultimately strengthening US competitors.
Domestic reaction
If the emergence of DeepSeek was a surprise to the outside world, it was seemingly a relative novelty for policymakers in Beijing as well.
DeepSeek is not affiliated with the large tech giants that have hitherto dominated the Chinese AI landscape, such as Baidu and Alibaba. In fact, it was only established as recently as 2023. Liang Wenfeng, the company’s 39-year-old founder, had not gained much of a public profile until mere weeks ago, when he was one of nine tech luminaries invited to deliver a speech at a closed-door roundtable organised by Premier Li Qiang in late January.
Wenfeng became a billionaire through running a hedge fund, HighFlyer, which used AI tools for investment decisions, and only started working on advanced AI development in 2023. It could be said that DeepSeek is one of the “little giant” start-up companies envisaged in Chinese technology development policy, but it is unclear whether it actually benefited from specific state support beyond general investments in infrastructure and education.
DeepSeek’s model also differed from the approach taken most often in the Chinese tech ecosystem. It eschewed app building and immediate profit seeking in favour of developing its technology efficiently. In those efforts, it was assisted by the fact that HighFlyer had made significant investments in computing power, purchasing thousands of high-end Nvidia A100 chips in advance.
US sanctions
The alleged use of advanced Nvidia chips will be scrutinised in the growing debate over what DeepSeek’s success reveals about the effectiveness of US AI sanctions, which were further tightened in the final days of the Biden administration — although their enforcement now rests with the Trump administration.
DeepSeek has stated that they have only used less performant — and sanctions-compliant — H800 and H20 chips to train their latest models. However, some tech executives have publicly voiced disbelief about this claim, suspecting that DeepSeek did use chips that were either acquired before sanctions or in violation of sanctions (see CHINA: Semiconductor smuggling presents new challenges – September 10, 2024).
Certainly, a lot of US corporate and state intelligence efforts will be directed towards auditing DeepSeek’s actual technological capabilities. However, if the company’s claims are proven to be false, it will be vastly self-defeating, as DeepSeek has drawn considerable attention to itself.
In Washington, policymakers will weigh whether the current AI-related sanctions are fit for purpose, and whether they have pushed China towards more innovation and self-sufficiency (see US/CHINA: US AI export controls on China have limits – May 29, 2024).
The Trump administration has, in response to the DeepSeek story, stated that Washington must remain the global leader in AI. Further sanctions, including a strict implementation of the last Biden sanctions package, are therefore highly likely.
Such sanctions would make Chinese AI ambitions more complicated in the short-term, limiting domestic firms’ access to critical AI components. In the long-term, however, they may continue having the opposite effect to Washington’s intentions: stimulating Chinese entities to innovate with greater urgency amid intensifying domestic rivalry.
Open-source
A major component of that innovative approach would likely lie in the embrace of open-source. Where Chinese companies are often criticised for being secretive, DeepSeek is seemingly transparent both in the actual code of its models, as well as in the developmental process leading to them.
This has multiple benefits for further technological improvement, including facilitating adoption by broader development communities. Perhaps counterintuitively, it is also commercially beneficial — enhancing trust with customers while enabling DeepSeek to provide tailored, client-specific solutions.
