GuShiio.com 
(Photo source: wired)
This weekend, DeepSeek, the AI catfish, stirred up another round of hot battles in the AI industry.
On March 1, DeepSeek published an article titled “Overview of DeepSeek-V3/R1 Inference System” on Zhihu, which comprehensively revealed the key secrets behind the V3/R1 Inference System.
It is noteworthy that the article disclosed key information such as DeepSeek’s theoretical costs and profit margins for the first time. According to reports, DeepSeek uses the large-scale cross-node expert parallelism (EP) method and uses a series of technical strategies to optimize the large-model reasoning system to the greatest extent, achieving amazing performance and efficiency.Assuming that the GPU rental cost is US$2/hour, the total cost is US$87072/day; if all tokens are calculated according to DeepSeek R1 pricing, the theoretical total revenue for one day is US$562027/day, and the cost-profit margin is 545%.
This is the first time DeepSeek has responded to the topic of API profits itself.GuShiio.comAGI calculated that the external API cost of DeepSeek R1 within one year is approximately US$37.64 million, or approximately RMB 270 million.
Earlier, GuShiio.comAGI reported in depth that Tencent, Huawei and other companies connected to DeepSeek and suffered a monthly loss of more than 400 million yuan.
Dr. You Yang, founder and CEO of Luchen Technology, said that in the short term, China’s MaaS model may be the worst business model. Big manufacturers will charge each other with low prices and free prices. The full-blooded version of DeepSeek R1 only charges 16 yuan per million tokens (output).If the daily output is 100 billion tokens, the monthly machine cost of DeepSeek based services is 450 million yuan, with a loss of 400 million yuan; the monthly revenue from AMD chips is 45 million yuan, and the monthly machine cost is 270 million yuan, which means that the loss also exceeds 200 million yuan.
“The more users there are, the more losses they lose. Can the cash flow hold up? Unless there is a free machine, but there is no long free lunch.& rdquo; You Yang said. (See the previous article on GuShiio.comApp: “Tencent, Huawei, etc. connect DeepSeek to lose more than 400 million yuan a month. Will the MaaS model as a service be disrupted?”)
This report attracted attention. On February 20, Tencent, Huawei and others connected DeepSeek to make the topic of losing more than 400 million yuan a month and ranked first on Weibo’s hot search.。Later, You Yang also sent a video to respond to the matter. He said that the loss of 400 million yuan was accurately calculated and that it (MaaS) might have burned too much money.
“With 4 H800 machines + a full-blooded version of DeepSeek, we measured that we can only output 1,000 tokens per second. You can imagine that you have to output 100 billion tokens every day, which is more than 100 million per day. Each machine is calculated at 300 coins per second. Four machines are 100 million yuan per day. Calculated for 4000 and 5000 units, based on the market price of H800 or depreciation, it is about 450 million yuan per month.& rdquo; You Yang said.
However, Yuan Jinhui, founder and CEO of silicon-based mobility, another AI Infra company that competes with Luchen, disagreed with him.
With DeepSeek’s response to costs and profits, the founders of You Yang and Yuan Jinhui posted articles in isolation and exchanged opinions on the circle of friends and Zhihu.
First of all, Yuan Jinhui expressed his gratitude to DeepSeek and commented:
“DeepSeek’s official disclosure of the costs and benefits of large-scale deployment has once again subverted many people’s perceptions. Nowadays, many suppliers are unable to achieve this level. The main reason is that the V3/R1 architecture is too different from other mainstream models. It is composed of a large number of small Experts, which makes systems developed aiming at other mainstream model structures no longer effective. The best efficiency must be achieved according to the methods described in the DeepSeek report. Developing such a system is very difficult and takes time. Fortunately, DeepSeek has opened up the main modules in five consecutive rounds this week, reducing the difficulty of community reproduction. These results fully reflect the first-principles thinking style and strong will of the DeepSeek team. They should first be based on some reasons (?) I thought of using such a model structure, and then found that whether it was training or reasoning, there were very big engineering challenges to do well. However, these problems were not impossible for their engineering team. The key was whether there would be big benefits from spending so much effort. Before the final result came out, no one knew, so they still made a bet, and the result was the right bet. It may also be the other way around, such a brand new model structure was designed based on the starting point of the system. rdquo; Yuan Jinhui said.
Later, You Yang published two articles about the cost of DeepSeek MaaS and the deceptive silicon-based flow.
You Yang said that DeepSeek data has no reference value for calculating MaaS costs. The article adds the number of tokens from DeepSeek web pages, apps and MaaS APIs to calculate them together. But You Yang believes that the MaaS he is talking about is a ToB tool, not a ChatGPT app. If DeepSeek’s MaaS wants to have such a high and full load state, it must keep its apps and web pages overloaded. He also pointed out that during the Spring Festival, the DeepSeek experience realized that it was not a qualified MaaS product at all.
“Before DeepSeek left the circle, I said on Weibo on January 2, 2025 that DeepSeek was the best model in China. I have nothing to belittle DeepSeek.However, the latency performance and experience of DeepSeek apps and websites during the Spring Festival are simply garbage. rdquo; You Yang said that it is impossible to make money selling DeepSeek MaaS.
Regarding silicon-based flow, You Yang issued a document saying thatThe reason for the surge in website visits three weeks ago was that at the expense of employees ‘Spring Festival holiday, Huawei first issued public accounts and available DeepSeek API during the Spring Festival holiday. Due to Huawei’s position in China, it is reminiscent of the localization of AI full-stack, which aroused the interest of Chinese people and had a good publicity effect. At the same time, the invitation code is directly sent to the voucher, and the head is quickly spread viral on the little red book. Both invitees and invitees will receive 14 yuan. There are many small red book users who have swiped thousands of yuan.
ButYou Yang pointed out that Silicon-based Mobile claims that it has 3 million users, and many users of Xiaohongshu say that their vouchers have reached 1000 yuan. Assuming an average of 500 yuan per user, there are 1.5 billion yuan in vouchers to be cashed for silicon-based flows, but this company only has 100 – 200 million yuan in cash. The risk is huge. So they had to castrate the model. He also said that it is unreasonable to compare the traffic volume of silicon-based mobile websites with Alibaba Cloud and Volcano Cloud. It is appropriate to compare the traffic volume of silicon-based mobile websites with Kimi Chat or Mita Search.
“Today, DeepSeek had an article pointing to me, and he (Yuan Jinhui) was also fanning the flames there.& rdquo; You Yang said that silicon-based flow has now limited the number of daily calls these students can make, and the API speed is as slow as a snail.
Moreover, Luchen Technology announced that it will suspend the DeepSeek API service.“Dear users, Lu Chenyun will stop providing DeepSeek API services in one week. Please use up your balance as soon as possible. If you don’t use it up, we will refund you in full.& rdquo;
Next, the two began to talk to each other in their circle of friends.
Yuan Jinhui claimed that You Yang slandered the company and pointed out that Luchen Technology had plagiarized the code.
“What’s wrong with our team being willing to work hard and seize an opportunity? What’s wrong with inviting users to give away free coupons? Many applications do this, including overseas; during the Spring Festival, when everyone wanted to visit DeepSeek but couldn’t, we provided the only stable service. What’s wrong with users being willing to come? Too many people came, the website was crowded, and paying users couldn’t use it anymore. They had to create a resource as a Pro version for paying users. What’s wrong with ensuring the paying user experience? We now experience the free version. A few years ago, silicon-based flow engineers opened up a batch of operators faster than Nvidia’s official implementation during OneFlow. They were also copied by Luchen Technology. They did not make it public just to save face for the other party. Now they slander us like this. rdquo; Yuan Jinhui said.
You Yang pointed out that the silicon-based mobile code was the responsibility of the (former) CTO of Lu Chen. After the code plagiarism incident, the CTO of Lu Chen resigned and joined silicon-based mobile.
It is reported that Silicon-Based Mobility and Luchen Technology both belong to domestic AI Infra computing power companies, providing computing power platforms, AI Infra solutions, etc., and creating an AI development and deployment platform.
Among them, in February this year, Luchen Technology completed a new round of financing, and the Beijing Economic and Technological Development Zone Industrial Upgrading Equity Investment Fund participated in this round of investment. The fund is part of the Yizhuang SDIC government investment guidance fund system. The last financing of Luchen Technology took place in September 2024, completing hundreds of millions of yuan in A++ rounds of financing, and Beijing City Artificial Intelligence Industry Investment Fund, Shixi Capital, etc. participated in the investment.
Yuan Jinhui, founder and CEO of Silicon-Based Mobility
SiliconFlow announced at the end of February that it had completed a 100 million yuan Pre-A round of financing. Huachuang Capital led the investment, Pricewaterhouse-followed the investment, and old shareholder Yaotu Capital continued to over-invest. Prior to this round of financing, Silicon-based Flow had introduced Meituan as a strategic shareholder. In addition, SiliconCloud, a large-scale silicon-based mobile model cloud service platform, launched a full-blooded version of DeepSeek-R1 V3 based on Huawei Cloud, which attracted attention. The company said that the total number of users of the SiliconCloud platform has exceeded 3 million, and underwear Tokens are called every day. Prior to this, Yuan Jinhui and Meituan co-founder Wang Huiwen founded the company Outside Light Years, which was later acquired by Meituan.
As of press time, You Yang told.comAGI that“(The company) was a CTO in charge of plagiarism. After leaving the company, he directly joined Teacher Yuan Jinhui’s company. The second person in charge of plagiarism, Lu Chenyun product manager, was directly fired by us. Now he has also joined a friend business. I won’t mention the name to avoid drawing further battles.There was no choice, people were always playing tricks.
GuShiio.comAGI also sought further response from Yuan Jinhui. Yuan Jinhui said that“(Lu Chen) CTO did not join our company. He previously joined Guangnian, and later joined other large companies. Moreover, it was all his own mistakes and blamed others, not others ‘problems.” rdquo;
