Chinese AI Redefines Efficiency: DeepSeek Releases OCR That Digitizes 200,000 Pages/Day, but Washington Suspects Links to Surveillance and Military Apparatus

The announcement today, October 20, 2025, by DeepSeek regarding the open-source release of its DeepSeek-OCR model—focused on two-dimensional optical compression—occurs within a geoeconomic scenario marked by fierce competition in artificial intelligence. However, the news of this technical innovation clashes with a much hotter chapter whose roots go back some time: the serious accusations leveled by Washington against DeepSeek, suspecting it of supporting Beijing’s military and intelligence operations and attempting to circumvent strict export controls to access advanced chips.

GettyImages

The Efficiency of Chinese AI and the Challenge to Western Hegemony

DeepSeek-OCR, an OCR (Optical Character Recognition) model with 3 billion parameters, consists of two main components: DeepEncoder and the DeepSeek-3B-MoE-A570M decoder. This architecture signals a clear shift in the distribution of global economic and digital power, where Chinese innovation efficiently and strategically challenges the technological hegemony of the United States.

DeepSeek has stated that the DeepSeek-OCR model demonstrates operational and cost superiority:

• Compression and Accuracy: The model efficiently compresses long text contexts via two-dimensional optical mapping, compressing text content into visual pixels. It maintains a low activation rate while achieving a high compression ratio. Experimental data shows that when the number of text tokens is 10 times that of visual tokens, the model's decoding (OCR) accuracy can reach 97%.

  
  

• Benchmark Performance: Its innovative architecture demonstrates strong performance potential in practical applications. In the OmniDocBench benchmark, the model surpassed GOT-OCR2.0 (1256 tokens per page) using only 100 visual tokens. Similarly, using less than 800 visual tokens, it surpassed MinerU2.0 (over 6,000 tokens per page on average).

  
  

• Economic and Production Advantage: In a real production environment, a single A100-40G GPU can generate over 200,000 pages of LLM/VLM training data per day, a workload equivalent to that of 100 professional data entry operators.

  
  

This computational efficiency and the declared performance translate into a significant economic and production advantage, strengthening China's position as a leading player in the global AI ecosystem.

Strategic Applications and Implications:

DeepSeek has highlighted that DeepSeek-OCR has broad application potential in critical sectors:

• Financial Sector: It can instantaneously convert voluminous financial reports into structured data.

• Medical Sector: It can rapidly digitize historical medical records.

• Publishing and Archives: The efficiency of digitizing ancient books will increase by tens of times.

• LLM Memory Management: The "visual memory" properties shown by this model offer a new approach to overcoming the context length limitations of large language models.

  
  

However, the rise of DeepSeek, like other emerging Chinese technologies, raises questions about privacy, data security, and geopolitical implications. While open source promotes diffusion, DeepSeek's capability to operate in critical sectors like finance and medicine intensifies technological competition, fueling the debate on "decoupling" between Sino-American technology supply chains.

DeepSeek Under Scrutiny: Military Ties and Evasion of Controls

Despite the claims of technical superiority, DeepSeek's Hangzhou headquarters is at the center of a geopolitical crisis whose origins date back some time. A senior US State Department official stated that DeepSeek "has willingly provided and will likely continue to provide support to China's military and intelligence operations".

Evidence previously presented by Washington includes:

• PLA Procurement Records: Over 150 references to DeepSeek were found in the procurement records of the People's Liberation Army (PLA), suggesting the provision of direct technological services to military research institutes.

• Evasion of Chip Controls: The company is accused of attempting to circumvent export restrictions to access "large volumes" of Nvidia H100 chips, using shell companies in Southeast Asia and attempting remote access to data centers. Nvidia, however, stated that their internal review indicates the use of H800 products, not H100.

• Privacy and Surveillance: DeepSeek is suspected of sharing user information and statistics with Beijing's surveillance apparatus. US lawmakers had previously expressed fears that the company was transmitting American user data to China. Code analysis also revealed direct links to the authentication systems of state-owned giant China Mobile.

The Chinese Geoeconomic Response: A Game of Cat and Mouse

The DeepSeek case is part of an escalating tension. Washington, while not yet having announced specific new sanctions, views the model as an example of Beijing's broader geoeconomic strategy to overcome technological restrictions.

China continues to show great adaptability, employing several strategies to maintain its competitive advantage in AI:

1. Circumvention (Black Market and Shell Companies): Use of foreign subsidiaries, often in Southeast Asia, to legally import chips and forward them to China, or reliance on the flourishing black market to obtain advanced semiconductors.

2. Remote Access: Leasing computing capacity in data centers located in non-sanctioned countries to train their LLMs on US-equipped servers.

3. Technological Autonomy: Massive long-term investment in the development and production of domestic chips (like Huawei's Kirin 9000s chip) and the optimization of less powerful chips through techniques like the "chiplet" approach.

Ultimately, the dual nature of DeepSeek—an innovation that promises efficiency versus a company suspected of geopolitical opacity—crystallizes the escalating clash between superpowers, defining the new boundaries of global economic and digital power.