Chinese AI company Luchen Technology halts use of DeepSeek, founder reveals the truth.

China’s first wave of official announcements of companies abandoning the use of DeepSeek has emerged. Recently, the “Tsinghua System” AI Infra company “Luchan Technology” announced the suspension of related services, becoming the first company to integrate DeepSeek models to provide API and cloud mirror services to users. According to its founder, You Yang, the actual cost of DeepSeek is much higher than the theoretical cost. After facing online backlash, he openly stated that it is impossible for DeepSeek to not use any American technology in the short term. Why can’t we tell the truth?

According to a report by Sina Technology on March 4th, on March 1st at 17:02, DeepSeek announced that the theoretical profit margin of the online system reached 545%. Subsequently, Luchan Technology released a statement through official channels stating, “We will stop providing DeepSeek API services in a week, please use your balance as soon as possible.”

The report stated that Luchan Technology has not publicly disclosed the reasons for discontinuing the provision of DeepSeek API services. However, You Yang, the founder, has published a significant amount of “rational analysis of DeepSeek costs” on platforms such as Zhihu, which can basically conclude that the cost is the key factor in their decision to abandon providing DeepSeek API services.

According to Sina Technology’s findings, in recent days, You Yang has had numerous reflections on DeepSeek, primarily focusing on three aspects:

1. The latest data released by DeepSeek showing a theoretical profit margin of 545% is based on a scenario where a large number of users overload DeepSeek’s app and website, rather than being a conventional MaaS (Model as a Service) product. To achieve stable output in conventional MaaS, the preparation of machines may be five times the theoretical value.

2. DeepSeek is the top modeling company in China, but its product often goes offline.

3. The paper on DeepSeek shows that it only costs $5.58 million to train a top-tier large model. However, the cost disclosed is for one experiment, and developing a large model may require dozens of experiments. Small and medium-sized cloud vendors should not rely on reselling DeepSeek API to operate at a loss and then boast about improving reasoning by 10 times.

Several individuals close to You Yang have told Sina Technology, “In fact, You Yang’s cost analysis of DeepSeek is aimed at discussing whether companies like Luchan Technology, playing the role of a third-party MaaS product service provider for DeepSeek, have a profitable model.”

A person close to You Yang straightforwardly commented, “It cannot be denied that the MaaS industry is the place where the entire industry bears losses. Major companies and model factories all actually operate their APIs at a loss. The real revenue comes from various collaborations, such as customized development or privatization.”

Both an investor in Luchan Technology and an internal source from the company similarly stated that making such business adjustments was mainly due to cost considerations. These individuals lamented, “Ultimately, it is because the cost of providing stable DeepSeek services is too high. They need to backup a lot of redundant computing resources, and they cannot compete with large cloud vendors in this regard, so they had to give up.”

The report mentioned that as early as February this year, You Yang publicly mentioned: “The training cost of DeepSeek-related models is not as low as rumored.” He calculated that “if the full version of DeepSeek R1 outputs 100 billion tokens per day, the monthly machine cost is 450 million, resulting in a loss of 400 million.”

Subsequently, You Yang was attacked by a group of online supporters of DeepSeek.

In response, in the early hours of March 4th, You Yang continued to post rebuttals on Weibo, saying, “Why don’t I respond to any slander? I am a Chinese national, this is a fact, with no one liking the comment. Comments saying I am Singaporean or American get a lot of likes. Does anyone still care about the truth?”

“You can’t avoid using any American technology in the short term for DeepSeek. Why can’t we tell the truth? I love China. The rational statement is that DeepSeek should be grateful for the open technological ecosystem provided by the United States. Likewise, if American companies use DeepSeek, they should also be grateful for the contributions of Chinese companies to global technology.”