Contents
Introduction:
In January alone, two Chinese apps unexpectedly ranked No.1 on the US App Store. On January 27, 2025, DeepSeek became the top free app on the US Apple Store.
The latest AI model, DeepSeek R1, has achieved significant success in the US, surpassing Xiaohongshu (Little Red Book), which previously held the top spot. Today, DeepSeek R1 overtook ChatGPT, becoming the most downloaded free app on the US Apple Store.
This milestone marks DeepSeek’s ability to provide high-performance AI models at an extremely competitive cost, just a fraction of OpenAI’s model training expenses. DeepSeek R1’s breakthrough performance and affordable pricing strategy have disrupted the AI market and sent shockwaves through Silicon Valley.
Following this, NVIDIA’s stock plunged over 12%, wiping out an estimated $300 billion in market value. Microsoft, Google, Meta, and Amazon saw declines of 4% to 6%. Meanwhile, Meta formed multiple DeepSeek research teams. In a public statement, Mark Zuckerberg acknowledged that the AI gap between China and the US is rapidly narrowing.
1. Why did DeepSeek become the Top App?
Why did this app, previously not widely known in the US Apple Store rankings, suddenly become the most downloaded?
The answer lies in its groundbreaking AI model, DeepSeek R1, which has attracted significant attention due to its exceptional performance and affordability.
R1 achieved benchmark results on AIME and MATH-500 comparable to OpenAI’s latest o1 model. This success was made possible through reinforcement learning-based training on advanced hardware. Despite this, its commercial deployment cost is only 5% to 10% of OpenAI’s, significantly lowering the entry barrier for users.
Furthermore, R1 offers flexibility by releasing distilled versions ranging from 150 million to 70 billion parameters, catering to various application scenarios. These models are available on Hugging Face and can be freely accessed under an MIT license, expanding DeepSeek’s influence in the AI development community.
This achievement underscores the company’s ability to deliver high-performance AI models at an incredibly competitive price, costing only a fraction of OpenAI’s training expenses. Its breakthrough capabilities and cost-effectiveness have made it a major disruptor in the AI market.
2. DeepSeek Surpasses OpenAI in the US Apple Store
To the surprise of many Americans, R1’s pre-training cost was only $5.58 million, whereas ChatGPT-4.o’s training cost reached hundreds of millions of dollars.
Additionally, the platform used only 2.8 million GPU hours, spending less than $6 million to train its latest version, V3, which continues to evolve at a rapid pace. This suggests that V3 could potentially surpass Llama3 in performance.
V3 stands out for its cost-efficiency. Compared to other major AI models like OpenAI’s ChatGPT and Meta’s Llama3.1, DeepSeek-V3 not only matches them in English processing but also excels in mathematical applications, making it a leader in AI innovation and cost-effectiveness.
3. DeepSeek R1’s Key Advantages Over OpenAI:
- Superior mathematical and logical reasoning: Achieved higher scores than OpenAI’s o1 model in AIME and MATH-500 benchmarks.
- Free access to its most advanced version: Unlike OpenAI, which requires a ChatGPT Plus subscription ($20/month), DeepSeek offers its most powerful AI model for free.
- Significantly lower training costs: DeepSeek R1’s entire training cost was only $6 million, while OpenAI’s comparable models cost hundreds of millions of dollars.
- No Chinese phone number is required: Foreign users can easily register using an email, making access more convenient.
- Optimized server infrastructure: Faster inference speeds and lower latency ensure a smooth experience on mobile devices.
4. How Did DeepSeek Succeed Despite Chip Restrictions?
Through Chinese Company Search, DeepSeek’s official name is “杭州深度求索人工智能基础技术研究有限公司”, the English name is Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., and it is backed by its parent company, “幻方量化”, the English name is High-Flyer Quant—a Chinese quantitative hedge fund. DeepSeek was founded in April 2023 by High-Flyer Quant, focusing on fundamental AI research and large language model development.
High-Flyer Quant’s official name is “杭州幻方量化投资管理有限公司”, the English name is Hangzhou High-Flyer Quant Investment Management Co., Ltd., and has been a major hardware provider for DeepSeek, supplying massive GPU resources.
However, it faced a significant challenge: the US government’s ban on exporting high-end chips to China. How did it overcome this obstacle? Here are the key solutions:
Securing Chip Resources
- Existing GPU stock: Before the US export restrictions, DeepSeek’s parent company, High-Flyer Quant, had imported around 50,000 NVIDIA H100 GPUs, ensuring sufficient computing power for large-scale AI training.
- Domestic chip development: Although China still faces challenges in producing high-end GPUs, domestic firms like Huawei and Cambricon are actively developing AI accelerator chips, which could eventually replace some NVIDIA products.
Optimizing Algorithms for Greater Efficiency
- Model architecture optimization: The platform adopted advanced operator fusion techniques to reduce computational redundancy, allowing larger AI models to be trained using the same computing power.
- Smart parallel computing: It leverages data parallelism, model parallelism, and pipeline parallelism, enabling lower-end GPUs to efficiently train large models, reducing reliance on H100 GPUs.
- Reinforcement learning-based training: R1 is entirely trained using reinforcement learning, without traditional supervised fine-tuning (SFT). This approach is critical for training top-tier AI models under limited computational resources.
5. DeepSeek’s Strategy for Overcoming Chip Restrictions
For a long time, many believed that AI dominance depended solely on computational power—whoever had the most GPUs would be the strongest. However, this company’s success proves that algorithmic efficiency can compensate for limited computational resources.
Compared to OpenAI and Google DeepMind, which rely on massive GPU clusters, this AI leader has achieved breakthroughs by optimizing reinforcement learning and mathematical reasoning.
Therefore, even if the US continues to tighten chip export restrictions, the company can still maintain its competitive edge through superior algorithmic optimization. This is why it became No.1—not just because of hardware, but because of its powerful software and algorithms.
Conclusion:
This AI giant’s rise not only demonstrates the rapid advancement of global AI technology but also highlights the intensifying AI competition between China and the US.
The release of R1 marks a turning point for Chinese AI labs, proving that they can challenge leading AI companies like OpenAI. This signifies a major shift in the AI industry and global market dynamics.
Ultimately, DeepSeek represents China’s leading position in AI innovation and the power of “Made in China.”
