Abstract:
According to news on August 27, the Wall Street Journal quoted people familiar with the matter as reporting that DeepSeek is seeking to raise US$7.4 billion. If the transaction is completed, the valuation of the Hangzhou AI company will reach US$74 billion. In June, when investors such as Tencent participated in the last round of financing, DeepSeek was still valued at more than $50 billion.
The new funds will be used for research and development and expansion of computing infrastructure. The report also revealed that DeepSeek’s annual recurring revenue (ARR) has reached US$500 million, and the company is preparing to go public in Shanghai next year. The financing has not yet been announced, and neither ARR nor time to market has been confirmed by DeepSeek.
The day before the financing news arrived, The Information disclosed another set of data closer to operating accounts: from January to July this year, DeepSeek had revenue of approximately 475 million yuan and a net loss of approximately 715 million yuan; during the same period, the comprehensive gross profit margin was 44.6%, and the API business gross profit margin was 82.9%. These figures also come from people familiar with the matter and are not the company's public financial reports.
$500 million in ARR is not the same as actual revenue
USD 500 million and RMB 475 million may seem far apart, but they do not conflict. The Wall Street Journal writes about ARR, which is annual recurring revenue converted based on the current business status; The Information writes about the revenue achieved in the first seven months. The latter also reported in July that DeepSeek's annualized revenue had reached US$400 million to US$500 million. The terms annualized revenue and ARR are similar, but the available materials are not enough to confirm that the two media use exactly the same calculation method.
According to the figures disclosed by the two media, DeepSeek's revenue has accelerated significantly in recent months, but the outside world still cannot see a verifiable full-year revenue and profit statement. According to rough calculations based on data disclosed by the Wall Street Journal, the valuation of US$74 billion is approximately 148 times the ARR of US$500 million.
The API gross profit margin of 82.9% shows that low-price models are not necessarily a loss-making business. If this caliber is true, DeepSeek is not simply exchanging losses for call volume. The model service itself already has a high gross profit margin.
DeepSeek is also adjusting its fees. On August 13, the company officially launched V4 Pro and implemented new peak and valley prices for the V4 series API from the 17th. V4Pro's output token per million has been raised from the previous 6 yuan to 13.5 yuan in off-peak hours and 27 yuan in peak hours; V4 Flash's price is 4.5 yuan in off-peak hours and 9 yuan in peak hours.
Low price is still DeepSeek’s advantage, but the price increase shows that the company is not prepared to rely solely on the lowest unit price forever in exchange for call volume.
Computing power investment is still increasing
The Information stated that DeepSeek’s AI infrastructure-related expenditures in the first seven months of this year were approximately 11 billion yuan, including renting AI servers and purchasing chips and other computing equipment. Model services began to generate gross profits, but the investment in computing power increased first.
In June, DeepSeek stated in its official recruitment announcement that it was working hard to at least double the size of all departments, with recruitment covering algorithms, AI core systems, infrastructure and operation and maintenance. At the end of July, Bloomberg reported, citing people familiar with the matter, that DeepSeek plans to add 1 gigawatt of computing power in Ulanqab, Inner Mongolia, build its own data center while renting external capacity, and strive to bring part of the capacity online by the end of 2027 or early 2028. DeepSeek has not confirmed this project, and it is not clear what chip it will use.
Recruiting people, training the next generation model, building and renting computing power all require spending money before the model can generate stable income. DeepSeek has proven that higher model efficiency can reduce training and calling costs; but higher efficiency does not mean that a cutting-edge model company can buy less computing power.
What to bet on with a $74 billion valuation
The Wall Street Journal stated that existing shareholders such as CATL are preparing to continue investing in the new round, and some local government background funds also plan to participate. In the last round, most investors needed to invest through limited partnerships managed by Liang Wenfeng. This time, DeepSeek is preparing to allow more investors to invest directly.
If these arrangements come to fruition, DeepSeek’s shareholder structure will be closer to that of a company preparing to go public.
DeepSeek’s initial inspiration to the industry was that strong models can also be cheap. What it has to answer in the future is whether cheap models can continue to make money and support the ever-expanding investment in computing power. The US$74 billion valuation is betting that these two answers can be true at the same time.
Comments