A seismic shift has reverberated through the artificial intelligence sector. In a matter of days, DeepSeek, a Chinese startup based in Hangzhou and established less than two years ago, has revolutionized expectations of AI capabilities. Emerging from the artificial intelligence R&D division of High-Flyer, a firm managing $8 billion in assets, DeepSeek actively incorporates AI into stock trading strategies.
The efficiency of this Chinese company is almost magical; experts highlight the model’s profound reasoning and analytical depth. Capable of running on a regular laptop, it removes the boundaries of control and access levels. China, once perceived as merely a catch-up follower, has now boldly challenged the rival, prompting a collective Western reflection reminiscent of 1957. While Meta promotes its LLaMA as a “free alternative,” DeepSeek offers its technology without charge, enabling companies to customize it to their needs.
With an estimated training cost of just $6 million, DeepSeek’s newly unveiled R1 model has achieved performance across several math and reasoning metrics on par with OpenAI’s o1 model, which required investments of tens of billions of dollars from both OpenAI and its backer, Microsoft.
DeepSeek attributes its limitations to access to chips rather than financial resources or talent, asserting it trained its v3 and R1 models using only 2,000 tier-2 Nvidia chips. At the same time, during recent discussions in Davos, Scale AI CEO Alexandr Wang claimed DeepSeek possesses 50,000 H100 processors; however, this statement has yet to be verified and may just reflect a strategy to mitigate exposure to further market declines.
This scenario not only undermines US policies restricting chip exports to China, but also jeopardizes the investments made by US firms in large data centers, endangering multi-billion dollar ambitions in intelligence and military projects. Last year, Amazon, which handles data for intelligence agencies from the UK and the US, placed a $150 billion wager on the data centers essential for the AI boom. Satya Nadella, who has invested $80 billion in AI, is now on the verge of recouping that investment. DeepSeek has erased nearly $1.5 trillion from the valuations of US tech giants, leaving investors in disarray and developers scrambling to rethink their architectures. As industry titans like Google and Microsoft reassess their budgets, emerging startups are reaping the harvest: for the first time, the narrative is shifting toward the notion that the true AI revolution may originate not from Silicon Valley, but from beyond the Great Wall.
Meta, which aims to invest $65 billion in AI infrastructure this year, has already established four war rooms to scrutinize DeepSeek’s models, attempting to find out how the Chinese firm achieved this cost efficiency in training their model and managed to leverage those insights to enhance their own Llama models.
The other day, Trump announced a new $500 billion initiative, backed by OpenAI, Nvidia, SoftBank, Oracle, and MGX, to develop critical infrastructure for creating artificial general intelligence (AGI). However, the viability of this project poses increasing political risks for the American president. The administration has heavily relied on AI to navigate the debt crisis, but the entire market is in danger of going into a tailspin with a tech bubble threatening to burst. Securing exclusive access to users and corporate data is the only way left to help OpenAI ascend to dominance.
The United States risks facing significant geopolitical repercussions, should the nation falter in this competitive race.