Deepseek hardware spending could be as high as $ 500 million: report

Deepseek from China became the biggest theme in technology this week, with many in the industry and on Wall Street they focused on a single number: $ 6 million.

In Deepseek’s paper Regarding its new artificial intelligence model, the company said that its total training costs amounted to $ 5,576 million, depending on the rental price of Nvidia Graphic processing units. Deepseek included a clear warning, saying that the number included only the “official training” of the model and excluded the costs related to “previous research and ablation experiments on architectures, algorithms or data.”

At the beginning of the week, Deepseek’s assistant took the coveted place for the most unloaded free application in the US. AppleThe Application Store, dethroning the OpenAI chatgpt. Global Tech Stocks sold, with Chips Nvidia manufacturers and Broadcom defeated A combined market limit of $ 800 billion on Monday.

TO New Semi -sis ReportA semiconductor research and consulting firm, added more context to Deepseek expenses. The firm estimated that the Deepseek hardware expense is “much higher than $ 500 million about the company’s history”, adding that R&D costs and the total property cost are significant. The generation of “synthetic data” for the model to train would require “a considerable amount of computation,” Semiianalysis wrote.

The report said that the sonnet Claude 3.5 of anthropic cost “$ 10 of million to train”, but noted that Anthrope raised billions of dollars of dollars Amazon and GoogleAn indication of how much more money is required to execute the models and the company.

“It’s because they have to experiment, create new architectures, collect and clean data, pay employees and much more,” said Semianalysis.

Deepseek’s own article does not include an estimate of its calculation costs. The company did not immediately respond to a request for comments.

“Being clear Depseek is unique in the sense that they achieved this level of cost and capacities first,” Semianalysts wrote. The firm added that Deepseek’s R1 is a very good model “and that” reaching the advantage of reasoning is so quickly impressive. “

Experts and analysts this week promoted the quality of the Deepseek model, and noticed how impressive the United States is considering. Chips exports with diving China three times in three years. That led to concerns that the United States is staying behind its main adversary in a market that is planned to overcome $ 1 billion in income within a decade.

Bernstein analysts wrote in a note on Monday that “according to the many shots (occasionally hysterical) we saw (during the weekend), the range of implications anywhere from” that is really interesting “to” this is the tip of death of the AI ​​infrastructure complex as we know it ‘”.

Deepseek was founded in 2023 by Liang Wenfeng, co-founder of High-Flyer, a quantitative coverage fund centered on AI. According to the reports, the startup of AI emerged from the AI ​​Research Unit of the Coverage Fund in April 2023 to focus on large language models and achieve artificial general intelligence, or AGI, a branch of AI that matches or exceeds the human intellect in a wide range of tasks, and that Operai and others are chasing.

Deepseek remains property and funded by High-Flyer, according to Jefferies analysts.

The buzz around Depseek began to pick up Steam earlier this month, when the startup launched R1, its reasoning model that rivals OPENAI O1. It is open source, which means that any IA developer can use it.

Like other Chinese chatbots, Depseek has limitations on certain issues: when asked about some of the policies of Chinese leader Xi Jinping, for example, Deepseek, according to reports, Deepseek Look at the user of similar lines of questioning.

The OpenAi CEO, Sam Altman, praised the model publicly, but the company has also said that there is evidence that Deepseek incorrectly harvested OpenAI data to build your product.

In an event in Washington, DC, on Thursday organized by Openai, Altman said Depseek is “clearly a great model.”

“This is a reminder of the level of competition and the need for the democratic LL to win,” he said. He said he also points to the “level of interest in reasoning, the level of interest in the open source.”

LOOK: The CEO of Nvidia, Jensen Huang and President Trump meet in the policy of AI