DeepSeek AI Chatbot Goes Viral, Disrupts Global AI Market in 2025

DeepSeek became a viral topic this week after its chatbot climbed to the top of the App Store and Google Play charts, drawing the attention of analysts and tech enthusiasts. It underscored a shift in interest in artificial intelligence and demand for powerful AI computing.

DeepSeek’s origins are closely tied to High-Flyer Capital Management – a Chinese quantum fund that uses artificial intelligence to inform its trading decisions.

Key Technologies and Infrastructure

From the outset, DeepSeek built its own clustered data centers for training models. But, like many Chinese AI companies, it came under the sway of U.S. export controls on equipment. To train one of its most advanced models, the company was forced to use Nvidia H800 chips, less powerful than the H100, made available by the U.S. government.

DeepSeek’s team is noted for its young makeup: the company actively recruits AI PhDs from leading Chinese universities and even accepts people without formal computer science backgrounds, so a broader range of knowledge can help in advancing the technology, The New York Times reports.

Key Models and Their Evolution

The initial set of models – DeepSeek Coder, DeepSeek LLM and DeepSeek Chat – was unveiled in November 2023. However, significant industry interest emerged in spring 2024 with the release of the DeepSeek-V2 family, which combines text and image processing and greatly reduces operating costs compared with competitors at the time. This spurred price cuts among some market players and made some models available at lower prices or for free.

In December 2024, DeepSeek-V3 was introduced, further increasing attention to the company. According to internal benchmarking, DeepSeek V3 outperforms open models such as Meta Llama and ‘closed’ API models like OpenAI GPT-4o. Also notable is the ‘thinking’ model R1, released in January, which, according to DeepSeek, operates at the level of the best models on key benchmarks.

In March, DeepSeek ranked second, despite traffic being down 25% from February, according to daily visit data.

R1 is positioned as an “understanding” model that can verify its own conclusions, increasing reliability in physics, science, and mathematics. At the same time, like other DeepSeek solutions, it is subject to regulatory oversight by Chinese regulators to ensure that its responses align with public and social expectations. Thus, in some discussions, R1 may have restrictions on topics related to Tiananmen or Taiwan’s autonomy.

In March, DeepSeek logged over 16.5 million visits. According to Similarweb, DeepSeek ranked second despite a 25% drop in traffic from February – an analyst noted in a industry trends report. However, overall DeepSeek exceeds ChatGPT in activity, which in March approached half a billion weekly users.

In May, DeepSeek released an update for R1 on the Hugging Face platform. In September, an experimental version V3.2-exp appeared, designed to significantly lower inference costs during long-context operations – this again reduced system usage costs.

Divergent Approach and Markets

Speaking about DeepSeek’s business model, the exact formula is not public. The company prices its products and services below the market average, and some solutions are given away for free. At the same time, investment flow is practically non-existent, despite interest from venture funds.

Experts explain that heightened efficiency allows maintaining low costs and competitiveness. However, many doubt the accuracy of the disclosed data. Developers actively use DeepSeek models that are not open source but are available under licenses for commercial use. According to Clement Delangue, CEO of Hugging Face, more than 500 derivative models have been built on top of R1 with millions of downloads.

DeepSeek’s success has put pressure on larger players: its profitability is viewed as an “AI reversal” and sometimes as a “distortion of attention” toward innovation. In some cases this affected Nvidia’s stock and drew responses from leading AI experts. Also U.S. government agencies imposed restrictions on using DeepSeek on government devices, the news agency reports.

Microsoft announced the availability of DeepSeek in its Azure AI Foundry service, while Meta’s leader stressed that AI infrastructure costs will remain a strategic advantage. OpenAI called DeepSeek “state-subsidized” and “state-controlled” and urged the U.S. government to consider banning its models. Nvidia, by contrast, highlighted DeepSeek’s important role as a source of innovation for computing-power needs. Some countries and regions also imposed restrictions: among them South Korea and New York banned the use of DeepSeek on government devices.

In May 2025 a Microsoft representative at a Senate hearing noted that company employees cannot use DeepSeek due to data-security and propaganda concerns. The future of DeepSeek promises new models and heightened regulatory scrutiny by the United States, but the company does not rule out further development and continued reductions in computing costs.

All these developments are shaping a new AI landscape: competition among Western, Asian, and European players is growing tougher, and regulatory requirements are driving more transparent governance of AI development.

This story emerged in late January 2025 and continues to unfold – expect new models and updates on regulatory changes in the future.