The Emergence of a Chinese AI Superpower? Reflections on the Deepseek Case

4 Reading time

“A cheaper and more efficient AI than its American competitors.” This, in sum, is how the global political and media sphere has been offering a panegyric on the advancements of the Chinese company DeepSeek. This persuasive force quickly spilled over into the financial sector: the American semiconductor giant Nvidia saw its stock collapse by 17% in a single day – a record. But should we take at face value the legitimately advantageous claims made by the Chinese firm? More importantly, are we truly witnessing a turning point in the technological competition between the United States and China?

Many elements suggest that DeepSeek, an artificial intelligence startup based in the city of Hangzhou, is disrupting the global geopolitical and geo-economic equation. On 20 January 2025, the company published large language models (LLMs) that rivalled the performance of American giants (OpenAI, Google…) at an unbeatable cost. Each user request is said to cost 27 times less than its competitor, GPT-4, while the development cost of DeepSeek’s R1 model is estimated to be 96% lower than that of its direct rival, OpenAI’s o1 model. At the root of this discrepancy is, in particular, the use of significantly fewer technological resources than those required by its American competitors. Yet DeepSeek achieved these spectacular results despite the American embargo on the export of next-generation semiconductors to China, which forced Chinese players to fall back on lower-performance technologies. It is this feat and apparent frugality that led to a nearly $590 billion loss in Nvidia’s stock market valuation on Wall Street… in a single day (27 January 2025) – an absolute historical record. The race for gigantism among American players, in terms of computing power – which is at the heart of the White House’s “Stargate” project, with a budget target of $500 billion – thus appears to be showing all its limits. By comparison, the training cost of DeepSeek’s V3 model is said to be no more than $5.5 million, whereas GPT-4 cost over $100 million.

However, it would be a mistake to see in this success the mark of China’s absolute technological independence. In fact, DeepSeek was able to acquire certain Nvidia graphics processing units (GPUs) before the implementation of the American semiconductor embargo. Among the chips in question is the H100, the American firm’s most advanced model. DeepSeek seems to have benefited from the informal semiconductor market that developed in response to American sanctions while relying on other supply (and sanction circumvention) channels, notably via India, Taiwan, and Singapore. Moreover, it is not inconceivable, as OpenAI suggests, that DeepSeek may have “distilled” the American company’s model – that is, trained its own model using responses from OpenAI’s model.

Nonetheless, the Chinese company’s ability to achieve performance levels comparable to American firms while requiring drastically lower computing power and energy consumption is a true tour de force – and perhaps a prelude to a much less unequal international technological competition than the American development model, highly capital-intensive (financial, technical, energy-related…), might suggest. Indeed, this achievement must be placed in the context of the American Stargate project, an artificial intelligence infrastructure initiative set to receive, over the next four years, a total budget of $500 billion (an initial investment of $100 billion has already been announced). This project takes the form of a company created by OpenAI, SoftBank (a Japanese telecom holding company and also the largest shareholder of Chinese giant Alibaba), Oracle (an American cloud computing firm), and MGX (a UAE investment company specialising in AI) – with the support of various technological partners (Arm, Microsoft, and Nvidia). At the centre of this project is the construction, in a first phase, of about ten data centres specifically dedicated to AI development. Others would be built later. However, critics from the Department of Defense have pointed out the unrealistic nature of this project from an energy perspective: the national power grid would not, as it stands, be able to support such a surge in electricity consumption, according to Roy Campbell, Deputy Director of Advanced Computing at the Office of the Under Secretary of Defense for Research and Engineering. Indeed, electricity demand in the United States is skyrocketing: it is expected to increase by 12.5% by 2030, according to consultancy Rystad Energy, due in particular to the circulation of more than 30 million electric vehicles over the same period and the country’s reindustrialisation driven by the Inflation Reduction Act, potentially leading to many highly electricity-intensive factories (for example, for battery manufacturing). Even before the announcement of the Stargate project, data centres were already projected to account for 9% of total U.S. electricity production by 2030 – twice the current levels. What will happen once this vast infrastructure project is implemented – if it ever materialises?

As venture capitalist Marc Andreessen suggests, are we witnessing a “Sputnik moment” – a reference to the launch of the first artificial satellite by the Soviet Union, which marked the beginning of the space race with the United States? The question boils down to whether China’s AI advances are altering the balance of power and threatening American hegemony. Indeed, it is common in the United States to link dominance in international relations to supremacy in technology – a framework of analysis that political scientist Stanley Hoffmann referred to as “expert thinking.” In fact, AI has been endowed with the fantastical attributes of the “ultimate weapon,” capable of granting whoever masters it, beyond a certain threshold, the highest degree of sophistication (sometimes referred to as “strong” or “general” AI), an unshakable leadership on the international stage. This is the full meaning of Vladimir Putin’s famous 2017 statement: “Whoever becomes the leader in this field will rule the world.” Behind this idea lies the possibility of achieving immeasurable productivity gains across all sectors of human activity (precisely through the widespread substitution of high-performance “machines” for human beings), which fuels the ambitions of economic and state powers. Regardless of the validity of such a hypothesis, the power of fantasy lies in its effects, often far more tangible than its causes.