**Reuters: “AI Stocks Outside China Could Be Shaken Further”**
[Herald Economy = Reporter Ahn Hyo-jung] Chinese company DeepSeek, which has caused a stir with its “low-cost, high-efficiency” artificial intelligence (AI) models, has released an internal assessment claiming it can achieve a profit margin of over 500%. This is the first time DeepSeek has disclosed information regarding the profitability of its inference operations.
According to Chinese media outlets like Pengpai and Reuters on the 2nd, DeepSeek’s official development team announced in a ‘DeepSeek-R1/V3 Inference System Technology’ description uploaded to the developer platform GitHub the previous day, that by optimizing throughput and latency, DeepSeek’s theoretical total daily revenue is $562,027 (approximately 820 million KRW), with total daily costs at $87,072 (around 130 million KRW). This results in a cost-to-profit margin of 545%.
However, the development team noted that the V3 model’s usage costs are lower than those of the R1 model, and several services remain free on web and app platforms. In addition, developers are paying less for usage during periods of low demand, suggesting that the actual profit margin is much lower than the theoretical margin.
Reuters mentioned, “This announcement could further shake the AI stocks outside of China, which plummeted globally after the web and application chatbots powered by the DeepSeek R1 and V3 models gained popularity in January.” The report added that this selling trend is partly due to claims that DeepSeek spent less than $6 million (approximately 8.8 billion KRW) on the chips used in model training, significantly less than its American competitor OpenAI.
Didi Das, a director at Silicon Valley investment firm Menlo Ventures, praised the theoretical profit margin figures released by DeepSeek on social media, stating, “If this were in the U.S., it would be valued at over $10 billion (approximately 14.6 trillion KRW).”
Despite ongoing skepticism about its low operational costs, DeepSeek has recently been actively releasing open-source code and overviews of its inference systems.