Faisal Bashir | Lightrocket | Getty Images
Chinese Deepseek has become the biggest topic in technology this week, and many in the industry and Wall Street are focused on one number: $ 6 million.
In Deepsek’s paper About its latest artificial intelligence model, the company said its total training costs amounted to $ 5,576 million, based on rental price Nvidia Processing graphic units. Deepseek included a clear warning, saying that the number included only the “official training” of the model and turned off the costs related to “previous exploration and ablation experiments on architecture, algorithms or data”.
At the beginning of the week, Deepseek’s AI helper took a wished place for most of the redecorated free applications in the US on AppleApp Store, Dethroning Openai’s Chatgpt. Global technological shares sold out, with nvidia chips and Broadco losing Combined $ 800 billion in market limit on Monday.
AND A new Semianize reportThe semiconductor’s research and counseling company added more context. The company estimated that hardware Deepseek consumption “significantly exceeding $ 500 million in the history of the company”, adding that the cost of research and development and the total cost of ownership is significant. Generation of “synthetic data” for a model that would train would require “a significant amount of calculation,” Semianalysis wrote.
The report says that Claude 3.5 sonnet from anthropic cost “$ 10 million for training,” but noted that Anthropopi collected billions for dollars from dollars from Amazon and Googlean indicator of how much more money is needed to start models and businesses.
“This is because they have to experiment, come up with new architecture, collect and clean data, pay employees and more,” Semianalysis said.
Deepseek’s own work does not include an assessment of his calculation costs. The company did not immediately respond to the comment request.
“To be clear Deepseek is unique in that they have first achieved this level of costs and abilities,” Semianalysts wrote. The company added that Deepseek’s R1 is a “very good model” and that “catching the edge of the explanation is quickly impressive.”
Experts and analysts this week advertised the quality of Deepseek’s model and noticed how impressive the US is an impressive consideration Chip exports to China three times in three years. This has led to concern that now is behind his main opponent on the market that is foreseen to be made of $ 1 trillion in revenue within a decade.
Bernstein’s analysts on Monday wrote in a note that “according to many (occasionally hysterical) heat we have seen [over the weekend,] Implications move anywhere from “It’s really interesting” to “This is the death of the knee and the infrastructure complex we know.”
Deepseek was founded by Liang Weenfeng in 2023, co-founder of high-flyer, a quantitative hedge fund focused on AI. Startup AI allegedly grew from the Hedge Fund’s research unit in April 2023 to focus on large linguistic models and achieve artificial general intelligence, or agi – a branch of Ai that is equal to or outweighs the human intellect on a wide range of tasks and which Openai and others The following are.
Deepseek is still fully owned and funded by a high flyer, according to Jefferies’ analysts.
The sound around Deepseek started picking up Steam earlier this month, when the startup published R1, his reasoning model that rival Openai’s O1. It is an open brawl, which means that every AI developer can use it.
Like other Chinese Chatboti, Deepseek’s have restrictions on certain topics: for example, when asked about some of the Chinese leader XI Jinping, Deepsek saves user From similar test lines.
Openai Executive Director Sam Altman has publicly praised the model, but the company also said he believes there is evidence that Deepsek irregularly harvested OPENAI data for the construction of your product.
At the Washington event, DC, on Thursday, hosted by Openii, Altman said Deepseek was “obviously a great model”.
“This is a reminder of the level of competition and the need to win the democratic al,” he said. He said this also indicates “the level of interest in reasoning, the level of interest in an open source.”
WATCH: CEO Nvidia Jensen Huang and President Trump meet in AI politics
