There are those who speak of an ‘upside-down world’, those who speak of a ‘prophecy of a trigger for armaments worse than the one that began with the deregulation signed a few days ago by Donald Trump’, those who attempt comparisons with what happened with the Soviet Sputnik and the birth of NASA, and finally those who think it is ‘an alternative to the monopoly of American big tech’.
This is what is being stirred up after the news of DeepSeek, the Chinese start-up specialising in open-source language models and general artificial intelligence (AGI), founded in 2023 by Liang Wenfeng and High-Flyer Capital Management, which has just launched DeepSeek-V3, its most advanced AI model: this open-source model with 671 billion parameters, offers performance comparable to GPT-4o and Claude 3.5, using reduced computational resources.
Currently open-source and free, it has recently overtaken ChatGPT as the most downloaded free app in the US, attracting developers due to its low cost and customisation, and caused the stock market shares of its US competitors to plummet.
In a context of barely two months, we have already experienced the new US president Donald Trump being ‘crowned’ by the so-called American ‘oligarchs’ at the inauguration of his second presidency: Tesla and SpaceX owner Elon Musk, Meta CEO Mark Zuckerberg, Alphabet CEO Sundar Pichai and Amazon founder Jeff Bezos all sat in the seats of honour in the front row during the ceremony: a real break with tradition, as the best seats at presidential inaugurations are usually reserved for family members and former US presidents. An important signal: the combined wealth of Musk, Zuckerberg and Bezos alone has grown to almost a trillion dollars in the last decade, while the federal minimum wage has remained unchanged since 2009.
At the same time, the announcement that some of these companies, the first being Google and YouTube, will no longer carry out checks on content, but will rely solely on user reports. A no to fact-checking preceded by the inauspicious fire in California, the scene of Silicon Valley, the American cradle of technology companies. An episode that triggered doubts about American sustainability policies and start-up culture.
Then Donald Trump’s announcement on a $500 billion artificial intelligence joint venture called Stargate, with OpenAI, SoftBank and Oracle to advance the AI infrastructure in the US. Next up is another action in his new term that sees the US president overturn Biden’s executive order on AI security. On 21 January 2025, his first day in office, Trump in fact repealed an executive order regulating artificial intelligence that had been issued by his predecessor Joe Biden in 2023. This order aimed to mitigate the risks that artificial intelligence could pose to consumers, workers and national security.
Subsequently, a letter of dismissal was sent to all members of the US PCLOB (Privacy and civil liberties oversight board), thus tipping the delicate balance that allowed European and US laws to interact. In fact, also based on the PCLOB, the European Commission allows European personal data to flow freely to the US under the so-called Transatlantic data privacy framework (TADPF). Dismissing them would bring the number of appointed members below the threshold necessary for the PCLOB to function, and would call into question the independence of all other executive redress bodies in the US. The European Union has relied on these US judicial bodies to believe that the United States offers ‘adequate’ protection of personal data. Now, in the absence of that supervisory authority, one of the risks foreshadowed could be that of no longer being able to use American platforms.
DeepSeek thus arrives during a geopolitical and financial tsunami (and perhaps this is no coincidence). Its rapid rise would prove that Beijing can hold its own against the US in the AI race despite the heavy technological restrictions imposed by Trump and Biden in recent years on Chinese companies.
DeepSeek also suddenly puts Nvidia’s billion-dollar profits into question by offering more efficient and cheaper AI models. On the Nasdaq, the stock of the company led by Jensen Huang went as far as to lose 10% in the pre-market (and then went on to lose almost 17%, i.e. almost USD 600 billion, more than the cost of the Stargate project itself). But that is not all.
It is worth mentioning the news from a few days ago that unfortunately seems to have passed unnoticed by many: China had put Nvidia and its artificial intelligence chips in its sights, opening an antitrust investigation for suspected violation of China’s anti-monopoly law. According to Beijing, Nvidia had also disregarded commitments made during the 2020 acquisition of Mellanox. In early December 2024, following the opening of the investigation, Nvidia’s shares fell by 2.6 per cent.
China and the US have clashed in recent years over exports of key chip production technologies, of which Nvidia is a major player. In early December 2024, Beijing had declared that it would restrict exports to the US of certain key semiconductor components, after Washington had announced restrictions on China’s ability to produce advanced chips. And the Chinese Ministry of Commerce had declared ‘national security concerns’ in a statement.
These objections relate to the agreement reached in 2020 when Beijing gave Nvidia conditional approval for its $6.9 billion acquisition of Mellanox Technologies. Washington wanted and wants to restrict the export of cutting-edge chips to China precisely because they can be used in advanced weapons systems and artificial intelligence. China still relies heavily on Nvidia’s chips for the development of artificial intelligence, but with such bans, Invidia, as a Californian company, can only sell downgraded versions of its AI products. China accounted for about 15.4 per cent of Nvidia’s total sales in the quarter ending October 2024.
“Deepseek is one of the most amazing and impressive innovations I have ever seen and, as open source, a wonderful gift to the world,” Marc Andreessen, known for co-founding Netscape, one of the first successful web browsers, and Andreessen Horowitz, one of Silicon Valley’s leading venture capital firms, wrote on social media. His R1 model is among the first in the world to combine real-time web search and advanced reasoning capabilities, something that OpenAI’s o1 model lacks to date: o1 in fact relies on pre-existing data, while R1 accesses up-to-date information from the web. Moreover, R1 makes its argumentation explicit, describing to the user the steps it takes to arrive at its answers. The user’s experience of its use, therefore, sees it as similar to a human stream of consciousness.
But the Deepseek phenomenon is not just that: while many AI companies depend heavily on advanced hardware, DeepSeek has focused on maximising its resources on the software side. It has had to do so because of the export restrictions imposed by the US, which limit Chinese access to the most advanced semiconductors. As a result, DeepSeek had to adapt to less efficient Nvidia chips, such as the H800, designed to comply with these restrictions.
And, despite these constraints, the company has perfected resource optimisation via software. An innovative, phenomenal approach, a model that many could now take inspiration from to achieve high performance at a much lower cost than competitor models.
ALL RIGHTS RESERVED ©