The CTO of Palantir Technologies Shyam Sankar recently criticized Nvidia CEO Jensen Huang’s idea that selling products in China would ultimately serve U.S. interests by making China dependent on American technology. Sankar argued that deepening economic dependence on China would only finance America’s own destruction. Huang, for his part, has consistently cautioned that U.S. export controls on the most advanced AI chips to China could enable Chinese companies to take the lead in developing their own AI chips, noting that about 50% of the world’s AI researchers are based in China. As Huang anticipated, the Chinese government decided to ban its major tech companies from purchasing Nvidia’s AI chips, pressing them to achieve semiconductor self-sufficiency and rely on domestic alternatives.
Unlike U.S. policymakers’ expectations, the goal of U.S. export controls—preventing China’s AI technology from catching up—seems to be slipping away. One researcher who has worked in AI labs in both the United States and China recently shared his experience that while U.S. labs operate with abundant GPU resources and prioritize scaling, Chinese AI labs—faced with restricted access to advanced chips—have fundamentally reoriented their development strategy toward maximizing performance per chip and achieving greater overall efficiency. Specifically, Chinese researchers are substituting algorithmic/software innovation (optimization techniques, compression methods, efficient architectures) for raw computational power (GPU/chip resources) available to their U.S. counterparts. One of the latest examples of this trend is DeepSeek-OCR, announced on October 21st.
DeepSeek OCR: Contexts Optical Compression
When using AI in everyday applications, one rarely stops to consider how many tokens are required for a model to accurately represent a page of text. DeepSeek-OCR, unveiled on October 21st, was developed to address precisely this fundamental efficiency question. Its “Contexts Optical Compression” technology enables models to process large volumes of visual text data with minimal computational cost in terms of GPU usage and advanced AI chips. Typically, AI models process images far less efficiently than text—a page of text as an image might require thousands of “vision tokens” to process, compared to only hundreds of “text tokens” for the same content as digital text. DeepSeek-OCR achieves remarkable compression by accurately extracting text from document images while using only one-tenth the typical number of vision tokens. At this 10× compression ratio, the model maintains 97% accuracy. Even at 20× compression, accuracy remains around 60%. In practical terms, DeepSeek-OCR processes documents using just 100 vision tokens per page, surpassing models that require 256 tokens and dramatically outperforming systems needing nearly 7,000 tokens. This efficiency enables a single compute node to process over 200,000 pages per day—a production scale that would be impossible with traditional approaches.
Figure 1: DeepSeek-OCR achieves 97% accuracy at 10× compression ratio (left) and outperforms existing models while using significantly fewer vision tokens (right). Source: DeepSeek-OCR Technical Report
To understand how DeepSeek-OCR works, imagine you’re trying to remember a long book. The inefficient approach would be to memorize every single word on every page, just as traditional AI models process every tiny detail of a document image. DeepSeek-OCR takes a different approach, similar to how humans actually read. First, you scan the page quickly to identify where the important text is located. Then, instead of remembering every detail, you create a compressed mental summary of what you read. Finally, you use this compressed summary to understand the overall meaning. Just as you can recall the key points of a book without remembering every single word, DeepSeek-OCR can accurately extract text while storing far less information in the AI’s “memory.”
Figure 2: The architecture of DeepSeek-OCR showing how the system processes documents through window attention, compression, and global attention stages. Source: DeepSeek-OCR Technical Report
This compression technique helps models “remember” and process increasingly long documents without requiring exponentially more computational power. Think of it like the difference between carrying around an entire encyclopedia versus carrying a book of well-organized notes—the notes take up far less space but still contain the essential information. By demonstrating that images can serve as an efficient storage format for text, DeepSeek-OCR points toward new possibilities for building AI systems that can handle very long contexts without needing massive amounts of computing hardware—exactly the kind of innovation that emerges when resources are scarce but ingenuity is abundant.
Figure 3: DeepSeek-OCR can parse complex visual content including financial charts and chemical formulas, demonstrating its practical applications beyond simple text extraction. Source: DeepSeek-OCR Technical Report
What Does it Mean for the U.S. AI Industries?
The implications of DeepSeek-OCR extend beyond mere document processing efficiency. This development represents a fundamental shift in how AI capabilities can advance under resource constraints. For years, the prevailing assumption in AI development has been that progress requires ever-increasing computational resources—bigger models, more GPUs, larger training datasets. DeepSeek-OCR challenges this paradigm by demonstrating that algorithmic innovation and architectural efficiency can achieve comparable or superior results with only a fraction of the resources.
The efficiency techniques employed by Chinese labs are now well-documented in open-source projects like Tencent’s Hunyuan-Large, which was trained entirely on export-compliant H20 chips yet achieves state-of-the-art performance. These techniques include mixture-of-experts architectures that combine multiple smaller specialized models to match the performance of monolithic large models while using significantly less compute. Likewise, DeepSeek-OCR’s approach to context compression through visual modality opens new avenues for addressing the quadratic scaling problem of transformer models. As language models attempt to process longer contexts, computational costs increase exponentially, creating a practical ceiling on context length even for well-resourced labs. By demonstrating that visual compression can effectively reduce information density while maintaining accuracy, DeepSeek-OCR suggests alternative pathways to long-context AI that don’t require massive computational infrastructure. This is also supported by Tencent’s report that mixed-precision training with bfloat16 can accelerate model training by up to 2.5× while maintaining accuracy, at the same time, quantization methods reduce model sizes by up to 16× with minimal accuracy loss. If DeepSeek researchers can consistently extract 10-20× more performance from limited hardware through optimization techniques, the gap in raw computational power becomes far less decisive.
For U.S. AI companies, DeepSeek-OCR’s technical achievements reveal an inflection point where efficiency innovations challenge the economics of scaling. DeepSeek-OCR compresses visual text into just 100 vision tokens per page while maintaining 97% accuracy at 10× compression—compared to systems like MinerU2.0 requiring nearly 7,000 tokens. It represents a 10-70× efficiency advantage that directly undermines the competitive moat of U.S. companies. In foreseeable future, American companies pursuing ever-larger models may discover that their Chinese competitors—forced to innovate under constraint—have developed systems that are not only cheaper to train and deploy but also more practical for real-world applications where software efficiency directly translates to lower latency, reduced energy costs, and broader accessibility.
Implications for the U.S. Strategists
The emergence of DeepSeek-OCR forces a fundamental reassessment of the assumptions underlying U.S. export controls on China. While Shyam Sankar’s critique of Jensen Huang focused on the dangers of economic dependence on China, the reality unfolding suggests that export controls are not slowing down China’s AI development. In contrast, they are redirecting it toward efficiency innovations that could prove more strategically significant than raw computational scale. DeepSeek-OCR exemplifies how scarcity breeds a different kind of innovation, one that prioritizes doing more with less rather than simply doing more.
The path forward requires moving beyond the binary debate of whether to restrict or cooperate with China, and instead recognizing that export controls itself is built on a misreading of geopolitical and economic reality. Export controls assume that restricting access to advanced chips will slow China’s AI development, but as DeepSeek-OCR shows, this assumption fails to account for how scarcity drives innovation. In this context, Jensen Huang’s argument about maintaining technological interdependence deserves reconsideration—not because economic dependence is desirable, but because complete technological decoupling may accelerate the very divergence in AI development paradigms. The United States has pursued a paradigm based on scaling—throwing more computational power at problems—while China’s resource constraints are driving a paradigm focused on efficiency, optimization, and algorithmic innovation.
From a U.S. strategic perspective, this divergence presents a double-edged sword: while technological decoupling may foster greater diversity and creativity in global AI development, it simultaneously erodes American influence over how AI systems are designed, deployed, and regulated in the world’s second-largest economy and a major technological power. The bigger question is whether the innovations born from those restrictions will reshape the global AI landscape in ways that leave the U.S. strategically disadvantaged despite its overall strength. As DeepSeek-OCR demonstrates, the real competition may not be about who has access to the most advanced chips, but who can innovate most effectively under constraints—a competition where scarcity is a powerful incentive and might paradoxically become an advantage rather than a limitation.
