FAISAL BASHIR | Lightrocket | Getty image
China’s Deepseek has become the biggest topic of this week. Many of them focus on the industry and Wall Street: $ 6 million.
In DeepSeek Paper The company’s total training cost for its latest artificial intelligence model is US $ 5.576 billion, based on rent Nvidia’s Graphic processing unit. DeepSeek includes a clear warning that the number includes only the “official training” of the model and excluded the “cost of previous research and ablation experiments related to the ablation experiments related to the ablation experiments.”
At the beginning of this week, Deepseek’s AI assistant occupied the mouth -watering attraction of the most coveted free application in the United States appleApp Store, Dethroning Openai ChatGPT. Global technology shares, chip manufacturers NVIDIA and chip manufacturers and chip manufacturers and chip manufacturers Broadcom lose Monday, a total of $ 800 billion in market value.
one New report of semi -analysisA semiconductor research and consulting company adds more backgrounds to Deepseek’s expenditure. The company estimates that the hardware expenditure of DeepSeek is “far higher than the company’s history than 500 million US dollars than the company”, and it is added that R & D costs and total cost of ownership are very important. The semi -analysis writes that it will require “quite a lot of calculations” to generate “synthetic data” for the model.
The report said that anthropomorphic Claude 3.5 14 -line poem “10 million US dollars in training” Amazon and GoogleThis shows how much it costs to run the model and the company.
Half analysis said: “This is because they have to test, propose new buildings, collect and clean data, pay employees, etc.”. “
Deepseek’s own papers do not include estimation of its calculation costs. The company did not respond to the request of the comment immediately.
The semi -analyst wrote: “Obviously, the uniqueness of DeepSeek is that they first reached the cost and ability level.” The company added that the R1 of Deepseek was “a very good model” and “quickly catch up with the reasoning of the reasoning of the reasoning. Objectively impressive. “
Experts and analysts touted the quality of the DeepSeek model this week, and pointed out that the impression of the United States was deeply impressed by the United States Containment chip export I have been to China three times in three years. This has caused people to worry that the United States is behind its main opponent in a market It is expected to be $ 1 trillion Increase within ten years.
Bernstein analysts wrote in a comment on Monday: “According to many (occasional (occasional) heat (on weekends), many (occasionally ashari) meaning range from” this is really interesting “to” to “to” This is the death of the AI ​​infrastructure complex. “We know. ‘”
DeepSeek was established in 2023 by LIANG Weeng. LIANG WENFENG is a co-founder of a quantitative hedge fund High-Flyer for AI. According to reports, AI startups developed from the AI ​​research department of hedge funds in April 2023, focusing on large language models and reaching artificially common information or AGI. This is a branch of AI. The branch is equal Or surpassing human intelligence, and Openai and others are pursuing.
According to Jefferies analysts, Deepseek is still completely owned by high aircraft and is funded by advanced flight.
The buzzing around DeepSeek began steam earlier this month. At that time, the startup released R1, which was a reasoning model that matched OPNAI’s O1. It is open source, which means that any AI developers can use it.
Like other Chinese chat robots, Deepseek is limited to certain themes: when asked about some Chinese leaders XI Jinping’s policy, it is reported that Deepseek is reported Transfer the user to From similar doubts.
Openai CEO Sam Altman publicly praised the model, but the company also said that evidence showed that DeepSeek Improper harvest OpenAI data to build its product.
Altman (Altman) held a event in Washington Special Economic Zone on Thursday, Altman said that Deepseek was obviously a great model.
He said: “This reminds people of the level of competition and the needs of the Democratic Party to win.” He said, which also pointed out “the level of interest in reasoning and the level of open source interest.”
watch: NVIDIA CEO Jensen Huang and President Trump met on AI policy meetings
