A month has passed since China’s DeepSeek, which stirred up a storm earlier this year as a ”cost-effective artificial intelligence (AI),” resumed new downloads in Korea. However, the number of users has actually decreased. DeepSeek has seen its monthly user base grow to about 96 million, primarily in China, India, and Indonesia, but its popularity has waned domestically following security concerns.
According to mobile app data analysis company WiseApp and Retail, DeepSeek’s monthly active users (MAU) last month were 218,699, a 23% decrease from April (284,184). This figure is based on a sample survey of Korean Android and iOS smartphone users conducted by WiseApp and Retail.
Using low-spec semiconductors to show performance comparable to leading AI corporations in the United States at low expense, DeepSeek emerged as a ”game changer” in the AI market at the beginning of the year. As the DeepSeek craze swept through Korea, the number of users surged. Notably, immediately following its launch, it surpassed OpenAI’s ”ChatGPT” to secure the top spot in app market downloads.
Domestic users of DeepSeek recorded 995,709 in January and 949,563 in February, according to the MAU compiled by WiseApp and Retail. Following a halt in domestic downloads due to security concerns, user numbers plummeted to 356,853 in March.
On Feb. 17, the Personal Information Protection Commission confirmed that user information from DeepSeek had been passed to ByteDance, the parent company of TikTok, and prohibited new downloads until improvements and enhancements to the service were implemented. After corrective measures were taken, DeepSeek resumed new downloads on April 28 on Google Play and the Apple App Store, approximately two months later.
However, in the meantime, DeepSeek’s popularity has waned. In the first week after new downloads resumed, the weekly active users (WAU) were 92,115, falling short of 100,000. In the last week of May (May 26 to June 1), the user base decreased to the 70,000 range.
Market analysts suggest that during the hiatus in new downloads for DeepSeek, ChatGPT gained users with its ”Ghibli art style,” while AI corporations like Google and Anthropic revealed next-generation AI models one after another, leading to decreased interest in DeepSeek compared to before.
Recently, DeepSeek also released an upgraded version of the generative AI model ”R1,” but evaluations indicated that it did not have as significant an impact on the global market as before. Business Insider reported on the 7th (local time) that DeepSeek’s latest inference model ”R1-0528” performed at about the third-best level globally based on benchmarks released by the company, but it lacked the disruptive power that it had at the beginning of the year. Reports indicated that the response was so negligible in the Western world that some might not have even noticed the new model was released.
Ross Sandler, a tech analyst at investment bank Barclays, noted that the latest update to DeepSeek R1 ”did not receive much attention.” He emphasized that the stock market ”showed no interest at all.” He explained that the lukewarm response to DeepSeek is attributed to the fact that the latest AI models are no longer as inexpensive as before. DeepSeek’s existing models were about 1/27th the cost of OpenAI’s ”o1” model, but now they are at about 1/17th, reducing their price competitiveness. Furthermore, AI corporations, feeling threatened by DeepSeek's emergence, subsequently released low-cost models with reduced expenses, diminishing the shock factor of ”cost-effective AI.”
Recently, allegations have been raised that the R1 model misappropriated data from Google’s Gemini. International experts believe it is likely that DeepSeek mimicked the thought processes of Gemini based on its inference processes. According to TechCrunch, Sam Fake, an AI developer from Melbourne, Australia, stated on his X (formerly Twitter) that the DeepSeek R1-0528 model tends to prefer vocabulary similar to Google’s Gemini 2.5 Pro. The developer of the AI assessment tool ”SpeechMap” also pointed out that the internal inference process of the DeepSeek model, known as ”traces,” is similar to that of Gemini.
Professor Choi Byung-ho of Korea University’s Artificial Intelligence Research Institute stated, ”Earlier this year, the market was surprised by the ‘DeepSeek shock’ in terms of quality, but now we have moved past that stage,” adding, ”U.S. big tech companies are unveiling innovative AI models and technologies one after another, so DeepSeek’s updates are merely keeping pace with them.”
While DeepSeek has never officially released user numbers, search engine optimization (SEO) firm Backlinko estimated that as of April, DeepSeek’s monthly active user count (MAU) stood at around 96.88 million, a roughly threefold increase from the 30 million in January. The majority of users, over 51.24%, are concentrated in China, India, and Indonesia, according to the firm’s analysis.