Sam Altman, Managing Director of Openaai, said that after the progress of the Chinese start-up start-ups from Deepseek, he will quickly follow the lead of the Chinese start-up in a global arms of the Silicon Valley and “deliver much better models”.
Deepseek’s generative KI chatbot, a direct rival of chatt, can perform some tasks at the same level as it recently published by Openai, Anthropic and Meta, despite claims that it cost a fraction of money and time for development .
The publication of Deepseek’s R1 model last week and its increase to the top of Apple’s App Store triggered a tech share sale. Asian Tech shares fell overnight on Tuesday after a Rout of Wall Street.
The Nasdaq fell 3 percent and the US chip maker -Nvidia, which produces the chips with which large AI models were trained. 17 percent dropped and lost $ 600 billion in market capitalization.
On Monday evening, Altman wrote that Deepseek’s model was “impressive, especially with regard to what they can deliver for the price”. He added: “We will obviously deliver much better models and it is also legitimate to have a new competitor!”
Altman, who announced last week that investors, including soft bank, would spend up to 500 billion US dollars to create a network of data centers to supply its AI models with electricity, added that the arithmetic resources ” are more important now than ever ”.
Microsoft, Meta, Alphabet, Amazon and Oracle provided $ 310 billion for investment expenses in 2025, including 310 billion. AI Infrastructure according to data made of visible alpha. Such estimates are based on the premise that large amounts of computing power are required to advance the AI functions.
But Deepseek’s ability to compete with a fraction of the Budget of Openai, which recently worth 157 billion USD and Anthropic, Google and Meta, has raised questions about the enormous sums that are cast into training systems.
“The winners are not those who burn most of the money,” said Aidan Gomez, founder of the Cohere, based in Toronto, the large language models for companies. Instead, he said, they would be those who “find efficient solutions”.
Deepseek’s progress has also uncovered risks for risk capital providers that brought almost $ 100 billion in US ACI start-ups last year. “There is now an open weight model on the Internet, with which you can start every other sufficiently powerful basic model as AI reasons for the AI,” said Jack Clark, co-founder of Anthropic, on Monday in a blog.
“AI skills worldwide just took a one-way rush striker,” he added. “A big praise to Deepseek for being so brave to bring such a change into the world!”
Deepseek’s success complicated the argument These massive cash piles create an unassailable advantage, which has contributed to the fact that the living laboratories in Silicon Valley increase tens of billions of dollars last year.
“If you are anthropic or Openaai, try to stand at the forefront, and someone can serve what you can do with a tenth of the costs is problematic,” said Mike Volpi, who led the investment of Index Ventures in Cohere .
The sudden publication of Deepseek’s latest model surprised some at Meta. “The main frustration is:” Why didn’t we come up with it at first? “If we have thousands of the smartest heads, work on it,” said a meta employee.
The managing director Mark Zuckerberg, who said last week that he would be expected to provide up to $ 65 billion for the expansion of AI teams and the establishment of a new data center, has campaigned for open source and meta in the USA puts in the foreground.
“We want the United States not to define the global AI standard,” said the company in response to Deepseek.
Yann Lecun, the chief scientist of Meta from Meta, said that “the AI assistant services for billions” would still require a high degree of computing power.
Insiders and investors of the competing company have expressed skepticism compared to the low costs for deepseek when developing its models. In December, the company said that its V3 model on which the chat bot of its app will only cost $ 5.6 million for training.
However, it added that this number only for the final training run, not for the full cycle, “which was excluded with previous research work. . . Experiments on architectures, algorithms or data ”.
Deepseek has its success -although he has used inferior chips to his US competitors, to methods that enable the AI model to selectively concentrate on certain parts of input data in order to reduce the cost of executing the model.
For its latest R1 model, it used reinforcement learning technology, a relatively new approach for the AI, in which models can improve without human supervision. The company also used open source models, including Alibabas Qwen and Metas Lama to set its R1 argumentation model.
The technical progress and interest of investors in deepseek progress could illuminate a fire among AI companies. “In general, we expect the bias to be improved to sprint artificial general intelligence faster and reduce the expenses,” said research company Rosenblatt on Monday.
Researchers and investors, including Marc Andreessen, have drawn parallels between the breed between the USA and China about artificial general intelligence agencies and their competition during the Cold War.
Stuart Russell, professor of computer science at the University of California, Berkeley, said the race after Agi was “worse”.
“Even the CEOs that deal in the breed have stated that those who win has a significant probability that the extinction of man will be caused in the process because we have no idea how we systems are more intelligent than ourselves can control, “he said. “In other words, the AGI race is a race to the edge of a cliff.”
Additional reporting by Michael Acton and Rafe Uddin in San Francisco and Melissa Heikkilä in London