Background and Methods
On March 15, 2025, according to multiple reports, Chinese AI startup DeepSeek developed an AI model that competes with Silicon Valley giants such as OpenAI at a very low cost, and its success has triggered discussions about the possible split in the global AI landscape. This report analyzes DeepSeek's achievements and its potential impact on the AI world based on public information as of March 15, 2025. The data sources include technology media reports (such as The New York Times, WIRED, Reuters, etc.) and social media dynamics, focusing on its technical characteristics, market reactions and geopolitical impact.
Data Collection
Gather relevant information from the following sources:
- DeepSeek's technical performance: DeepSeek's founder talks about its unexpected success
- Chinese social media reaction: Chinese companies rush to use DeepSeek
- Technology benchmarking: How Chinese companies publish top AI inference models
- Geopolitical reaction: DeepSeek removed from South Korean app stores
- Silicon Valley Reactions: DeepSeek executive talks about Silicon Valley's reaction
Quantitative analysis
The success of DeepSeek can be quantified in the following aspects, as shown in the table below:
index | DeepSeek | Silicon Valley competitors (such as OpenAI) | Remark |
|---|---|---|---|
Number of training chips | About 2,000 Nvidia H800s | About 16,000 pieces (A100/H100) | DeepSeek is more resource efficient |
Training cost | Approximately US$5 million | Hundreds of millions of dollars (e.g., about $100 million for GPT-4) | Significant cost gap |
GSM8K Math Benchmark Score | 67.2% | GPT-3.5: 50.8%, GPT-4: 80.8% | Close to GPT-4 level |
App store downloads | Apple App Store free app list | No direct comparative data | Reflecting global popularity |
DeepSeek’s success and the AI split
DeepSeek's AI model, DeepSeek-R1, excels in reasoning and mathematical tasks, and its cost is only a fraction of that of Silicon Valley giants. In the GSM8K benchmark test, DeepSeek-R1 scored 67.2%, which is better than OpenAI's GPT-3.5 (50.8%) and close to GPT-4 (80.8%). This shows that DeepSeek has reduced its dependence on high-end chips through efficient model architecture and training algorithms.
Its success could herald a split in the AI world, mainly in the following ways:
- Intensified technological competition : DeepSeek’s low-cost model challenges the U.S.’s technological dominance and could lead to a split in AI development. For example, Silicon Valley companies such as Perplexity began using DeepSeek’s R1 model (Perplexity’s X post), showing the expansion of its global influence, but also potentially intensifying U.S.-China technological competition.
- Geopolitical tensions : The South Korean government has suspended downloads of the DeepSeek app (DeepSeek removed from South Korean app stores), indicating that geopolitical factors may limit its global expansion. This may lead to differences in AI tools in different regions, for example, Chinese models may be more popular with data privacy-sensitive users, while Western models may continue to dominate high-end applications.
- Regulatory and censorship differences : DeepSeek’s models are subject to Chinese regulatory constraints and may be limited on sensitive topics (e.g., Xi Jinping, the war in Ukraine), which could lead to a split in functionality and usability between Eastern and Western AI models.
Geopolitical and regulatory influences
The US-China technological competition could lead to a split in the AI world, especially over differences in data privacy, censorship, and export controls. DeepSeek's open source strategy has attracted global developers, but its restrictions by China's censorship rules could affect its acceptance in Western markets. For example, social media user "Xiaoming" said on his Threads account that if DeepSeek is subject to censorship rules, its global influence could be limited.
In addition, the US export restrictions on high-end chips are aimed at curbing China's AI development, but DeepSeek proves that China can achieve breakthroughs with limited resources. This may prompt Silicon Valley companies to reassess their AI development strategies, but it may also exacerbate geopolitical tensions and lead to the fragmentation of the AI ecosystem.
Future Outlook
In the future, the AI world may see a split between the East and West models, with the Chinese model likely to be more popular with data privacy-sensitive users, while the Western model may continue to dominate high-end applications. This may affect technical standards and collaboration, for example, different standards and protocols may emerge in AI development, hindering the unified development of global AI technology.
In addition, DeepSeek's low-cost model is publicly shared and may be copied by global companies, lowering the threshold for AI development. This may lead to increased participation from small companies and developers, but it may also increase the risk of technology proliferation, such as North Korean hackers or Russian military may build AI systems more cheaply.
Discussion and Conclusion
Research shows that DeepSeek's success may foreshadow the division of the AI world, and the evidence tends to support that its low-cost, high-performance model has exacerbated the division of the global AI landscape, but there are disputes involving geopolitical and regulatory differences. An unexpected detail is that despite the hardware limitations, DeepSeek still attracts Silicon Valley companies such as Perplexity to use its model, showing the expansion of its global influence.
limitation
Due to limited data after March 15, 2025, some analysis is based on trend speculation. The specific channels and quantities of chip acquisition may be controversial.
Key Quotes
Comments
Post a Comment