DeepSeek Claiming 700 Million Won Daily Profit... "Actual Profit Margin of 545% Is Much Lower"

"Profit Margin of 545% Compared to Costs"
DeepSeek Official Development Team Announces
"Actual Profit Margin Will Be Much Lower Than This"

Chinese AI startup 'DeepSeek' evaluated the profit margin of its service to be over 500%.

According to local media such as Pengpai and Reuters on the 2nd, DeepSeek's official development team stated in a technical explanation of the DeepSeek-R1/V3 inference system posted on the developer platform GitHub the day before that they "optimized throughput and latency." Theoretically, DeepSeek's total daily revenue is $562,027 (approximately 820 million KRW), and the total daily cost is $87,072 (approximately 130 million KRW), resulting in a profit margin of 545% compared to costs.

Chinese Artificial Intelligence (AI) DeepSeek. Reuters · Yonhap News

However, the development team estimated that the actual profit margin would be much lower than the theoretical profit margin. This is because the usage cost of the V3 model is lower than that of the R1 model, some services are offered for free on the web and applications, and developers pay less during periods of low usage.

This is the first time DeepSeek has disclosed the profitability of its inference operations. Reuters explained, "The global popularity of web and application chatbots powered by DeepSeek's R1 and V3 models since January could further shake AI stocks outside China, which have plummeted." It added, "This sell-off is due to claims that DeepSeek spent less than $6 million (approximately 8.8 billion KRW) on chips used for model training, far less than its U.S. competitor OpenAI." Didi Das of Silicon Valley investment firm Menlo Ventures also mentioned on his social media, "If DeepSeek were in the U.S., it would be a company valued at over $10 billion (approximately 14.6 trillion KRW)."

Meanwhile, amid some skepticism about DeepSeek's low operating costs, the company has recently been releasing open-source code and overviews of its inference system consecutively.