DeepSeek-R1 is a large language model developed by the Chinese company DeepSeek, released on 20 January 2025. It became a breakthrough success, surpassing ChatGPT as the most downloaded freeware app on the iOS App Store in the United States by 27 January 2025. The model's efficiency and open-source approach triggered a significant market reaction, causing an 18% drop in Nvidia's share price on the same day, as investors reassessed the demand for expensive AI hardware.
The event highlighted a shift in the Artificial intelligence landscape, where smaller, more efficient models could rival larger counterparts. DeepSeek-R1's success was described as "upending AI" and initiating "a global AI space race," according to industry observers. The model's open weights and infrastructure code were praised for contributing to open-source artificial intelligence, while its compliance with Chinese government censorship policies raised concerns about privacy and information control.
Background and Development
DeepSeek began as a series of large language models in November 2023, starting with the DeepSeek-LLM series. The company, based in China, focused on developing efficient architectures that could compete with established players like OpenAI and Google (AI). Early models were similar to Llama, using dense decoder-only transformers, but later iterations introduced innovations such as Mixture of experts and multi-head latent attention.
On 10 January 2025, DeepSeek released a Chatbot based on the DeepSeek-R1 model for iOS and Android. The app quickly gained traction, and by 27 January it had become the most-downloaded freeware app on the iOS App Store in the United States, overtaking ChatGPT. This rapid adoption was driven by the model's free access and strong performance in reasoning tasks.
The release of DeepSeek-R1 was part of a broader strategy by DeepSeek to offer open-weight models. Unlike many competitors that kept their models proprietary, DeepSeek made its infrastructure code and model weights publicly available, allowing researchers and developers to use and modify them. This approach aligned with the principles of open-source artificial intelligence and contributed to its popularity.
Market Impact on Nvidia
The market shock on 27 January 2025 was triggered by the perception that DeepSeek-R1's efficiency could reduce the need for massive Compute resources, which are central to Nvidia's business. Nvidia had become a dominant supplier of GPUs for AI training and Inference (AI), with its market capitalization reaching $3 trillion in 2024. The company's growth was fueled by demand from tech giants and startups alike.
However, DeepSeek-R1 demonstrated that high-performing models could be trained with fewer resources, using techniques like Mixture of experts and efficient attention mechanisms. This raised questions about the sustainability of the AI hardware boom. On 27 January, Nvidia's share price fell by 18%, erasing hundreds of billions of dollars in market value. The drop was one of the largest single-day declines in the company's history.
The impact extended beyond Nvidia, affecting other AI-related stocks and prompting analysts to reassess the RACE for AI dominance. Some experts argued that the market reaction was overblown, noting that efficient models could increase overall AI adoption, potentially boosting demand for hardware in the long run. Others saw it as a signal that the era of "scale at all costs" was ending.
Technical Innovations
DeepSeek-R1 built on several architectural advances introduced in earlier DeepSeek models. The Transformer (architecture) architecture, which underpins most modern large language models, was refined with multi-head latent attention (MLA) and Mixture of experts (MoE) layers. These innovations reduced memory usage and computational costs while maintaining performance.
Multi-head latent attention, introduced in DeepSeek-V2, compressed key-value caches to lower memory requirements. This was a departure from the standard Multi-Head Attention mechanism used in models like Llama. The MoE architecture, which activates only a subset of experts per Token (language model), allowed the model to have a large number of parameters without incurring proportional compute costs.
DeepSeek-R1 also incorporated techniques for improved reasoning, such as reinforcement learning and self-principled critique tuning. These methods enhanced the model's ability to solve complex problems, including mathematical reasoning and code generation. The model's performance on benchmarks was comparable to leading models from American AI companies, according to tests used by those firms.
Usage and Accessibility
DeepSeek-R1 was made available through the official DeepSeek website and mobile applications, free of charge, with no limit on the number of queries. Users could sign up using global email services like Gmail or Yahoo. The web version had a cap of 500 messages per hour to prevent spam.
For developers and businesses, DeepSeek offered API access to the R1 and V3 models with a usage-based pricing model. As of February 2025, API pricing was approximately $0.28 per million input Token (language model)s and $0.42 per million output tokens, making it less expensive than some competing services. This low cost was a key factor in the model's adoption.
The model's open weights allowed for local deployment, which appealed to organizations concerned about data privacy. However, DeepSeek's data collection practices and compliance with Chinese government censorship policies drew scrutiny from regulators in multiple countries. Some governments banned the app on official devices, citing security concerns.
Broader Implications
The DeepSeek-R1 market shock had significant implications for the AI industry. It challenged the assumption that only massive investments in Compute could yield advanced AI capabilities. This prompted discussions about the efficiency of Pretraining and the potential for smaller players to compete.
Nvidia's stock decline was a wake-up call for investors, but the company continued to innovate. Nvidia had a history of acquisitions, including Mellanox and Arm (though the latter failed), and its GPUs remained essential for AI workloads. The company's AI revenue continued to grow, and its market cap eventually recovered.
The event also highlighted the geopolitical dimensions of AI development. DeepSeek's success was seen as a milestone for Chinese AI, which had been advancing rapidly despite export controls on advanced chips. The collaboration between DeepSeek and Tsinghua University on later models underscored the role of academic institutions in AI research.
Subsequent Developments
Following the initial shock, DeepSeek continued to release new models. On 3 April 2025, in collaboration with researchers at Tsinghua University, DeepSeek published a paper unveiling DeepSeek-GRM, which combined generative reward modeling and self-principled critique tuning to improve inference-time scaling. The company said these models would be open-sourced.
On 30 April 2025, DeepSeek released DeepSeek-Prover-V2-671B, a math-focused model for formal theorem proving and mathematical reasoning. On 24 April 2026, DeepSeek released DeepSeek V4 and V4-Pro, followed by DeepSeek-V4.1-Flash in September 2026. These releases continued the trend of efficient, open-weight models.
The market impact of DeepSeek-R1 was a defining moment in the Generative AI era. It demonstrated that innovation could come from unexpected places and that the economics of AI were not set in stone. The event influenced investment strategies, with some companies diversifying their hardware suppliers beyond Nvidia, including AMD and AWS Trainium.
Conclusion
DeepSeek-R1's release and the subsequent Nvidia stock drop marked a turning point in the AI industry. The model's efficiency and open-source nature challenged prevailing assumptions about scale and cost. While the market reaction was dramatic, it also spurred discussions about the future of AI development, the role of hardware, and the importance of accessibility.
The event underscored the rapid pace of innovation in Machine learning and the interconnectedness of technology and finance. As AI continues to evolve, the lessons from DeepSeek-R1 - that efficiency and openness can be as valuable as raw power - are likely to shape the industry for years to come.