Parsing the geopolitics surrounding DeepSeek

The AI race between the US and China looks set to be at the forefront of the competition between them for decades to come, and this could have wider consequences

    • The importance of DeepSeek is not so much about its technical ability to substitute US AI platforms as it is about making clear that the AI race is open.
    • The importance of DeepSeek is not so much about its technical ability to substitute US AI platforms as it is about making clear that the AI race is open. PHOTO: BLOOMBERG
    Published Tue, Feb 25, 2025 · 06:41 PM

    THE sudden eruption of DeepSeek, until now a relatively unknown Chinese artificial intelligence (AI) startup, has shaken the world. DeepSeek’s success – exemplified by it reaching the No 1 ranking on Apple’s app store – has led to a massive correction in the stratospherically high valuations of the US’ biggest tech giants involved in AI.

    DeepSeek’s R1 model has been claimed to surpass Open AI’s cutting-edge o1 model family with a much smaller investment and without access to the most advanced chips due to US export controls. The silver lining of the financial and technological scarcity embedded in DeepSeek’s model is that R1 seems to run at a much lower cost and consume much less energy that its American peers.

    The takeaway, reflected in the sudden massive correction of US AI giants, is that the United States’ hegemony in AI is no longer guaranteed, as a much smaller financial investment, with the right talent, can obtain similar results. Furthermore, the US reliance on export controls to contain China technologically does not seem to be effective.

    In principle, such conclusions should be positive for the world, given the potentially massive efficiency gains and the related commoditisation of AI brought about by DeepSeek. The European Union, so far a follower in generative AI – as China seemed to be until DeepSeek’s surprise – could potentially find itself with a home-grown AI platform. This seemed impossible until now, based on the sheer amount of investment needed to develop large language models.

    The positive aspects of DeepSeek’s success are undeniable, but there are also less-positive angles that need to be considered. Looking at the technical aspects, for a start, DeepSeek can hardly be compared with US AI platforms since its key role is to optimise existing models, rather than develop new ones that can compete with those from the US. The positive aspect of optimising models instead of creating them is that the massive cost of traditional pre-training can be minimised.

    DeepSeek’s optimisation, achieved through a combination of techniques, including Mixture of Experts and reinforcement learning, can massively reduce computational costs. But it does not fundamentally break the “scaling law”, which says larger models deliver better results.

    BT in your inbox

    Start and end each day with the latest news stories and analyses delivered straight to your inbox.

    As a result, the development of the most powerful AI systems still needs powerful infrastructure, which brings the race back to huge financial resources.

    Another important issue to consider is that DeepSeek is not fully open access, as certain components, such as training data, fine-tuning methodologies, and parts of its architecture, remain undisclosed. The latter is all the more important, considering that DeepSeek, as with any Chinese AI company, needs to be in full compliance with China’s national security law and the most recent AI regulations introduced in 2022. The regulations ensure that any AI output from a Chinese AI platform is censored to avoid any criticism of the Chinese political regime, and also that responses follow Chinese propaganda.

    Whether Western governments will accept China’s censorship within their jurisdictions remains an open question which will soon be tested with DeepSeek. In the particular case of the EU, the AI Act provides a regulatory framework which aims at ensuring that AI systems are transparent, accountable, and respect human rights, including freedom of expression and political speech. Thus the AI Act could stand between DeepSeek and its European users. A second concern is the location in which DeepSeek stores the data, namely mainland China.

    However, the biggest concern is geopolitical. US President Donald Trump’s comments on how DeepSeek may be a wake-up call for US tech companies, which some have named a Sputnik moment, indicate clearly that the AI race between the US and China will be at the forefront of the strategic competition for decades to come.

    This competition is pushing both governments to devote as many resources as needed to lead, as AI not only has commercial, but also military capabilities in cyberspace, autonomous motion and beyond. The latest endeavour, on the US side, was the announcement of “Stargate” on Jan 18, with an investment of up to US$500 billion to accelerate AI research. Previous endeavours on both sides focused on hardware, with former US president Joe Biden’s Chips and Science Act and China’s three semiconductor funds.

    The importance of DeepSeek is not so much about its technical ability to substitute US AI platforms as it is about making clear that the AI race is open. This creates the necessary anxiety for both sides to put even more resources into this field. On the US side, the most likely outcome of such a realisation is a doubling down in the US protection of its core technologies by tightening export controls and other means.

    Another likely consequence of this accelerated race will be the drying out of any scientific cooperation between the US and China on any critical technology. Republican Senator Josh Hawley has already pushed forward with a proposed Bill to decouple America’s AI capabilities from China’s.

    Given the above, Europe must move beyond the positive aspects brought about by the commoditisation of AI after DeepSeek’s success, and realise that an even-harsher technological competition between the US and China for AI dominance will have consequences for Europe.

    The most immediate one is the likely bifurcation into two AI worlds as a consequence of tighter export controls and more limited scientific cooperation. In a non-Trumpian world this could have been read positively; less cooperation with China might have implied more cooperation with US allies, including the EU. But that does not seem to be the direction the Trump administration is taking.

    Second, the EU’s own AI regulations may make it very difficult to benefit from the efficiency gains from DeepSeek and other AI platforms from China, based on censorship and data protection concerns.

    All in all, beyond the hype, DeepSeek should bring an important realisation to European policymakers – namely that the EU is caught between a rock and a hard place in an increasingly geopolitical AI race. The EU’s own strategic autonomy is harder than ever. But the reliance on either the US or China is also challenging for different reasons – more difficult tech transfer from the US, and China’s censorship and data issues.

    Alicia Garcia Herrero is chief economist (Asia-Pacific), Natixis CIB. Michal Krystyanczuk is a data scientist.

    Share with us your feedback on BT's products and services