All you should know about the Chinese AI model that drops the markets

Such a day have not seen Wall Street for a long time: The trade week opens with sharp declines, with the person in charge of shaking is a small Chinese startup called Deepseek. In an investment of less than $ 6 million, using relatively weak chips, the startup has succeeded in developing an artificial intelligence model that exceeds American models with huge investment.

The big drama included a 3% decrease in the NASDAC index led by Anbia – which proved that even with a huge $ 3.5 trillion market value, you can easily drop in a double -digit rate – and along with a similar trend among other chip shares, including Marvel, Broadcom, AMD and ARM.

The incredibly cheap new model undermines the premise that needs strong computational power to develop the best artificial intelligence, and overwhelmed the investors who rushed to get rid of shares in the field. The trend also came to the Asian and more so -to -date stock exchanges in Tel Aviv, here the dual -chip shares Kuttak and Nova plummeted sharply and pulled the flagship index.

What is special about Chinese development, is it really threatening to blow up the artificial intelligence bubble, which an organia and other companies at the chip club have enjoyed it in recent years, and what is the chance of the Chinese AI app to penetrate the American market? Globes makes order.

What sparked the storm?

Deepseek recently released the V3 model, that analysts defined as “one of the strongest open models.” The model was released with an open source license, which allows developers to adjust it for most applications. Construction of the model took about two months and was invested in less than $ 6 million. For comparison, Google has invested hundreds of millions of dollars to train its Jamini model.

The model performance is impressive not only because of the humble investment, but also because of the weak hardware behind it. According to reports, due to the US export limitations, the company used 2,000 low -capable ana chips, while the competition develops a on -top with about 16,000 chips at least.

Last week, the company released the R1 model that revealed in late November, and should compete with the OpenAi Reasoning model, the GPT O1. The company claims that the model transcends or at least compares to the O1 model in some standards – such as math, coding and reasoning. According to publications, the R1 model consists of six more focused and smaller version, each with 1.5 billion parameters.

“OpenAi’s model is the best in performance, but we also don’t want to pay for abilities we don’t need,” said Wall Street Journal Anthony Fu, founder of a startup that uses artificial intelligence to predict financial returns, and moved to Deepseek from a model of the model Claude. He said the company tests showed that the new model performs the tasks in the same way and in the quarter of the cost.

Who is behind the company?

There is still not much information about Chinese society, but the head is a fairly anonymous character named Liang Wonpang. On the network reports, the company was born from a Chinese hedge fund called HIGH-FLYER QUANT, which runs $ 8 billion. Wonpang, according to Bloomberg, set up Deepseek with a registered capital of 10 million yuan (equivalent to $ 1.4 million).

Wonpang, which is no longer anonymous, even met on January 20 with Lee Kachiang, head of the China State Council, to discuss the reduction of Chinese companies in front of the US, published Wall Street Journal. In an interview he provided the body Communication 36kr said Wonpang that what led to progress was not fundraising – but US restrictions on access to the best chips.

Why did the market respond to declines?

One of the main concerns that analysts point to is the high pricing of technology giants. If a small Chinese company has succeeded in developing successful models with poor equipment and modest investment, why the giants spend more and more? “Their massive artificial intelligence expenses need re -evaluation,” CNBC experts say.

The investment banks also addressed the drama. In JP Morgan, the model has dramatically reduces the costs, and quoted Wenpang’s “subversive claim” that “more investments are not worth more innovation.” Jeffris also referred to this, noting that “the appearance of effective training models from China may further enhance the race for artificial intelligence between the US and China. The key question is whether the strategy is to build at all costs, or whether the focus will move to a higher efficiency – which exerts pressure on the big players. ”

In contrast, City actually thought it didn’t change the perception today. “While the achievement of Deepseek may be a breakthrough, we doubt the perception that the achievements have been made without the use of advanced GPU processors to build the basic language models based on the final model that has undergone a refinery technique. The dominance of the American companies in AI modes can be undermined, but we estimate the environment. Of the US for advanced chips is an advantage. Therefore, we do not expect leading AI companies to move away from more advanced GPU slaves. ”

Why does Anabiya drop?

The chip giant, which has become the world’s largest company in terms of market value in the past year, opened trading on Monday with a decrease of almost 12%, thus imitating close to $ 400 million. She is already expected to finish the end of the day as the third largest Wall Street company, after Apple and Microsoft.

“The big threat to Anabiya,” Sargi Vasaconock, a senior analyst at the Oppenheimer Investment House, “is that technology companies will understand that they do not need Anbia’s strong processors and can also be successful with lower efficiency. Wacchonuk also adds that “if Deepseek’s case shows that you can do a lot with less, perhaps the chip arms race decreased, and then the entire echo -basket built on Anbia, will be a serious problem.”

Hanan Zakai, CEO of Customer and Technology in Codevalue, which deals with the cloud and transformation, agrees and explains to Globes that “Deepseek breaks the paradigm that needs a lot of computing power to process artificial intelligence. If it turns out to be right, it is a human blow to the anonymous. ”

What are US concerns?

Beyond the big competition with American models, the US has set significant restrictions on advanced chip technologies to China, partly to stop the state’s artificial intelligence progress. Deepseek has proven that engineers overcame restrictions and focused effectively despite the meager resources. Advanced to coach the models, but at the end of the limit, the restrictions were found to be effective.

It was also announced that Deepseek models are censorship to prevent criticism of the Communist Party in China, which sets a challenge to adopt it in the world. Such censorship can indicate the involvement of the Chinese administration or at least its powerful power on the company.

What is special about open source model?

The Chinese model allows the developer to take a software item and use it freely. OpenAI models for comparison are open to use, but if the key wants to reach an advanced stage and certainly for commercial use – it will be required to pay for the use of the code. Codevalue is entitled to Globes that “the significant gospel is that the code can be downloaded to you for systems and servers and use.”

Will the model be restricted in the US?

Vacconock from Opanheimer cools the enthusiasm, saying that “in a world where Tiktok and Wavy are boycotted, no Chinese artificial intelligence models enter the United States.” In addition, Wacchonuk says the Chinese model does not receive access to American information dads, so it can hurt its efficiency.

Reports indicate that the model refuses to answer sensitive political questions about China and leader Shai Jinping, and provides comments on the official propaganda in China. However, this is open source so users can easily change it.