It took just a few days for the app Chinese DeepSeek to climb the ranks and rise to the first place in the App Store ranking in many parts of the world. This has been underlined in the last few hours by several Chinese media highlighting the fact that the latest version of theartificial intelligence assistant of DeepSeek Startup, ChatGPT rival, is, according to them, cheaper and it might be even better based on some parameters than that of Meta and OpenAI.
La market reaction it was very strong on both sides of the globe, even if then some analysts they believe that perhaps it was excessive and that the prices could recover in the next few hours. In any case, the tsunami has hit the Nasdaq 100 futures which continued to show signs of worsening: at 12.00 Italian time they are at -4,66%. Obviously this has had an impact on the stocks of high tech companies, leading Asian ones on the rise e US ones down.
The Hang Seng index of Hong Kong closed with a gain of 0,7%, the CSI 300 of the stock exchanges Shanghai Shenzhen closed near breakeven, the day before closing for a week-long holiday to celebrate the Chinese New Year (until Tuesday of next week). In particular, companies that investors connect to the Deepseek system are rising: Iflytek +4,8%; TRS Information Tech +20%; Merit Interactive + 20% Hangjin Technology + 10% Range Intelligent Computing Technology Group + 5.2%.
On the contrary, the shares of Nvidia's supplier, Advanced fell 8,5 percent in Tokyo. SoftBank Group, an investor in AI-focused startups, slid more than 8%, posting its biggest daily decline since Sept. 30. The chipmaking equipment giant Tokyo electron fell 5%
DeepSeek: Who is the engineer who founded it?
The company was founded by Liang Wenfeng, an engineer born in 1985 in Guangdong who graduated in machine vision systems at Zhejiang University. In 2015, together with two of his fellow students, he founded the High-Flyer hedge fund, capable of becoming a giant in the IT and hi-tech sector in China in less than 10 years: it is from that fund, which today manages around 8 billion dollars in assets, that in 2023 Wenfeng created DeepSeek with the intention of investing all its resources to build his own cutting-edge artificial intelligence models based on his desire for scientific research. “I can’t find a commercial reason,” the engineer told the Chinese 36kr. “Basic research, in fact, has a very low return on investment.”
In the suo team, as he remembers Wired, did not choose expert engineers, but only PhD students from the country's top universities, from Peking University to Tsinghua University, eager to demonstrate their capabilities. "Our top technical positions are mostly filled by people who have graduates in the last two years“, Wenfeng explained to the Beijing press in 2023. “Most people, when they are young, can devote themselves completely to a goal without utilitarian considerations. What do I ask of potential employees? Solve the world’s most difficult problems.”
More efficient and definitely less expensive than US Big Tech
DeepSeek It has established itself with the release, in late December, of a large language model (LLM) open source. The company's flagship model, DeepSeek R, has impressed with the efficiency of the results achieved in relation to the development costs. According to the research team, the project required less than two months and a lower budget to 6 million dollars, a paltry sum when compared to the billionaire investments from companies like OpenAI and Google.
Ironically, it is a tool created by the Chinese precisely in reaction to the years of sanctions imposed by the United States on chips, which encouraged Beijing to strive for self-sufficiency. So, while American companies struggle to obtain advanced chips due to the restrictions imposed by the United States government on exports to China, DeepSeek has managed to demonstrate that it is possible to achieve competitive results even with less powerful hardware. The strength of DeepSeek, according to experts, lies precisely in theHardware optimization: The Chinese lab managed to develop its model using Nvidia H800 chips, reduced-capacity versions of the powerful H100 chips currently used by US companies for their flagship AI models.
THEDeepSeek Assistant shows its work and reasoning, while responding to the user's written question or request, a candor and transparency that is highly appreciated by users. In a series of third-party tests, DeepSeek's model outperformed Llama 3.1 by Meta, GPT-4o of OpenAI and Claude Sonnet 3.5 of anthropic for accuracy, from complex problem solving to mathematics and coding, according to industry assessments. The American investor Marc Andreessen said, speaking of DeepSeek, that we are facing "one of the most surprising and impressive discoveries". Applause also from Yann LeCun, Chief AI Scientist of Meta: “The real lesson is in open source.”
Trump didn't mention the potential threat to his half-trillion-dollar artificial intelligence initiative, but traders are keeping a close eye on his Truth Social account. The sector was boosted this week by news that SoftBank, OpenAI and Oracle have formed a $100 billion joint venture to fund AI infrastructure. Meta's CEO announced ambitious plans to invest up to $65 billion in AI-related projects by 2025.
DeepSeek news now increases attention on health state with Big Tech which will announce the quarterly results this week. It will start on January 29th with Microsoft, Meta and TeslaThe next day, January 30th, it will be the turn of Apple Friday 31st January it will be the turn of Amazon.
Some analysts cautious
- Bernstein analysts They noted that the new models are impressive in their ability to compete with leading OpenAI and Meta models, while claiming to use a fraction of the investment and chip infrastructure used by their competitors. They did, however, push back on claims that DeepSeek built OpenAI for $5 million and that the market panic seems overblown. “While we acknowledge that DeepSeek may have reduced the cost of achieving equivalent model performance by, say, 10x, we also note that current model cost trajectories are still increasing by about that amount each year (the infamous “scaling laws…”) that cannot continue forever,” Bernstein analysts wrote in a note. In conclusion, Bernstein recommends against buying into the “doomsday scenarios currently being bandied about and that it still has a positive view on Nvidia and Broadcom, both of which are rated Outperform for their AI prospects.
