This week in the world of technology, one of the most discussed topics was the China-based artificial intelligence company DeepSeek. DeepSeek, in an article detailing their latest artificial intelligence model, mentioned that the total training cost of the model was $5.576 million, calculated based on the rental fees of Nvidia’s graphics processing units (GPUs). The company also issued a warning: this amount covers only the “official training” of the model, without including the costs of previous research and trial (ablation) studies related to new architecture, algorithms, or data.

The attention of everyone, from Wall Street to insiders in the industry, was captured by a single number: $6 million. DeepSeek moved ahead of OpenAI’s ChatGPT to claim the title of the most downloaded free application in the U.S. on Apple’s App Store with its “AI Assistant” at the beginning of the week. This led to a selling wave in global technology stocks; particularly, chip manufacturers Nvidia and Broadcom collectively lost $800 billion in market value on Monday.

According to a new report by SemiAnalysis, a research and consultancy firm focusing on the semiconductor industry, additional details on DeepSeek’s expenses were provided. The report indicates that DeepSeek’s hardware spending significantly surpasses $500 million throughout the company’s history. It emphasizes the high costs of research and development (R&D) and the total cost of ownership, stating the substantial need for processing power even to generate “synthetic data”.

The report also highlights that “tens of millions of dollars” were spent to train Anthropic’s Claude 3.5 Sonnet model, but Anthropic managed to secure billions in investments from Amazon and Google. This showcases the significant resources required for artificial intelligence models and the companies developing these models. SemiAnalysis explains these high costs as being attributed to “trying new architectures, gathering and cleaning data, paying employee salaries, and much more”.

Notably, DeepSeek’s article did not include an estimate of how much the company spent on processing power. So far, the company has not responded to requests for comments on this matter.

In the SemiAnalysis report, the statement “DeepSeek’s achievement in reaching this level of cost and capability first makes it unique.” is used. The report describes DeepSeek’s R1 model as “very impressive,” noting that reaching the most advanced level of reasoning in such a short time is objectively impressive.

Throughout the week, experts and analysts praised DeepSeek’s model quality. Despite U.S. imposing restrictions on chip exports to China three times in the last three years, the success of DeepSeek drew more attention. As a result, discussions have begun on whether the U.S. will fall behind its biggest competitor in the artificial intelligence market expected to generate over $1 trillion in revenue.

In a note released on Monday, Bernstein analysts stated that some of the “occasionally exaggerated comments” they heard over the weekend ranged from “This is really interesting” to “The end of the existing artificial intelligence infrastructure”.

Founded by Liang Wenfeng in 2023, DeepSeek is still fully owned and financed by High-Flyer, an AI-based quantitative hedge fund. This AI venture was previously part of High-Flyer’s AI research unit but became independent in April 2023, focusing on large language models and artificial general intelligence (AGI). AGI aims for artificial intelligence to equal or surpass human intelligence in various tasks and is a key goal for several companies, including OpenAI.

The excitement around DeepSeek began with the release of the R1 reasoning model, which competes with OpenAI’s “o1” model earlier this month. Moreover, R1 is an open-source model, allowing any artificial intelligence developer to utilize it.

Similar to Chinese chatbots, some limitations are present in DeepSeek’s chatbot, redirecting queries to different directions when asked about Chinese leader Xi Jinping’s policies, signaling its predefined boundaries.

While OpenAI CEO Sam Altman praised DeepSeek’s model in public, the company also stated concerns regarding DeepSeek allegedly using OpenAI data without permission for its product development.

During an event held by OpenAI in Washington, D.C. on Thursday, Altman described DeepSeek’s model as “absolutely fantastic,” highlighting the need for competitiveness and the importance of “democratic artificial intelligence”. He also drew attention to the significant interest in reasoning and open-source topics.