This week, the Chinese artificial intelligence company Deepseek has been increasing the price by claiming that the new AI model exceeds Openai.
Specifically, the claim that Deepseek’s large -scale language model training costs only $ 5.6 million is currently expanded in the computing in frustration required to train and execute advanced AI workloads. I caused concerns about the total wake -up.
Investors are afraid of Deepseek’s destructive impact, and has been erased nearly $ 600 billion from NVIDIA’s market capitalization on Monday. This is the largest decrease in the US history.
But not everyone is convinced of DeepSeek’s claim.
CNBC asked the industry expert for opinions on DeepSeek and asked how to compare it with Openai, the creator of Viral ChatBot Chatgpt, which actually caused the AI revolution.
What is DeepSeek?
Last week, DeepSeek released R1 A new inference model comparable to O1 of OPENAI. Progress model is a large language model that is disassembled, and we consider multiple approaches before generating a response. It is designed to process complex problems like humans.
Deepseek was established in 2023 by Liang Wenfeng, a co -founder of high -central hedge funds in AI, focusing on large -scale language models, artificial general information, and AGI.
The AGI as a concept relys on a wide range of tasks or an AI idea that is equal to human intelligence or beyond that.
Many of the technologies behind R1 are not new. However, it is worth noting that DeepSeek first developed the high -performance AI model, according to the company, with significant reduction in power requirements.
“There is a lot of possibilities to develop this industry. The high -end chip/capital -intensive method is one technical approach,” Xiaomeng Lu, the director of the Eurasian Group’s Geotechnology Cleans, Xiaomeng Lu. I mentioned it.
“But DeepSeek has proven that we are still in the early stages of AI development, and the route established by Openai may not be the only route to a very competent AI. ”
How is it different from Openai?
DeepSeek has two major systems that have been talked about from the AI community. V3 is a large language model that does not use the product and R1, R1.
Both models are open source. In other words, the basic code is free, and other developers are published to customize and redistribute.
The model of DeepSeek is much smaller than many other large language models. V3 has a total of 671 billion parameters or variables that the model learns during training. Openai does not disclose parameters, but the experts estimate that they have at least one trillion of the latest models.
Deepseek from the viewpoint of performance say The R1 model quotes a benchmark that verifies the AIME 2024, CodeForces, GPQA Diamond, Math-500, MMLU, and SWE benches, which is comparable to O1 of OPENAI.
In a technical report, the company stated that V3 model training costs were only $ 5.6 million. This is a part of the billions of dollars spent by Western AI labs such as Openai and humanity to train and operate basic AI models. However, the amount of execution cost of DeepSeek is not yet clear.
However, if the training cost is accurate, the model means that it was developed in just a part of the cost of Openai, human rival models. Google others.
Daniel Newman, the CEO of Tech Insight Firm The Futurum Group, stated that these development suggested a “large breakthrough,” but he questioned the exact numbers.
“Deepseek’s breakthrough shows the meaning of the scaling method, and I think it’s a true need,” he said. “Nevertheless, because it is related to the development of DeepSeek, there are still many questions and uncertainty about the overall cost picture.”
Meanwhile, Paul Triolio, a senior VP of the advisory company DGA GROUP and the technical policy lead Senior VP, states that it is difficult to derive direct comparisons between Deepseek’s model cost and the model costs of major US developers in the United States. I did it.
“The 5.6 million numbers of Deepseek V3 are only for execution of one training, and the company emphasized that it did not express the overall cost of R & D to develop a model,” he said. 。 “Since then, the overall cost was significantly higher, but it is more likely to be lower than the amount spent by major US AI companies in the United States.”
When I was contacted by CNBC, DeepSeek could not comment immediately.
Compare the price of Deepseek and Openai
Both DeepSeek and Openai disclose the price setting of the model on the website.
According to Deepseek, R1 is 55 cents per million tokens of “token”, which refers to individual textbooks processed by the model, and $ 2.19 per million tokens.
In comparison, the O1 Openai price setting page indicates that the company charges $ 15 per million input tokens and $ 60 per million output tokens. In the case of GPT-4O mini, a small and low-cost language model of Openai, companies claim 15 cents per million input tokens.
Skepticism of chips
Deepseek’s R1 clearly, the US export control has been restricted from the use of AI chips, which is advanced, but the model has been built, so it is intense over the truth of the claim. It has already led to the public discussion.
Deepseek claims that there was a breakthrough using mature NVIDIA clips, including H800 and A100 chips. This is not more advanced than the cutting -edge H100 of chip manufacturers that cannot be exported to China.
However He said that the scales of AI CEO’s Alexandr WANG, a comment on the CNBC last week, believed that Deepseek used a banned chip.

NVIDIA came out later that DeepSeek’s GPU was completely compliant with exports.
Is it true?
Industry experts seem to have widely agree that DeepSeek has been achieved, but some have skeptical of some of the Chinese companies.
“Deepseek is legally impressive, but the hysteria level is a great deal of accusation,” said Palmer Luckey, a US entrepreneur, who established Oculus and Anduril.
“The $ 5 million number is fake. To delay investment in emerging companies in the United States, provide services to Titan in the United States such as NVIDIA, and hide the sanctions to carry over. It is promoted by Chinese hedge funds.
Netmind’s SEDA REJAL, SEDA REJAL, stated that it was a London headquatter startup that provides access to DeepSeek’s AI model via a distributed GPU network, and there is no reason not to believe DeepSeek.
“Even if it was a specific factor, it was still very efficient,” Rejal told CNBC this week. “The logic of what they explained is very wise.”
However, some people argue that DeepSeek technology may not be built from zero.
“DeepSeek has a strong sign that O1 has made the same mistake and torn technology,” said the billionaire investor Vinod Khosla in X.
It is reported that Openai itself implies that CNBC may have developed an AI model called “distillation” in the CNBC on Wednesday, and the output data used by DeepSeek “inappropriately” from the model. It is a claim that you are doing it.
“We are taking aggressive and aggressive measures to protect technology, and will continue to work closely with the US government to protect the most capable models built here,” he said. A spokeswoman told CNBC.
AI commercialization
However, since the scrutiny surrounding DeepSeek is shaking, AI scientists have widely agree that it is a positive step for the industry.
Yann Lecun, the chief of AI scientists MetaAlthough DeepSeek’s success represents the victory of the open source AI model, China’s victory on meta in the United States is behind the popular open source AI model called lama.
“Looking at DeepSeek’s performance, people who think that” China is beyond the United States with AI. ” You read this by mistake: “The open source model is beyond its own model,” he said in Linkedin’s post.
“Deepseek is profitable from open research and open source (Meta’s Pytorch, LLAMA, etc.). They came up with new ideas and built them on other people’s work. It is open. The power of research and open source.
clock: Why DeepSeek is at risk of American AI lead

-Cnbc’s Catholy Navi Shop and Haydenfield contributed to this report