DeepSeek Unfairness: Chinese Startup Outsmarts Silicon Valley with Cost-Effective AI, Responds only in Chinese

The AI world was rocked by DeepSeek’s recent breakthrough, sending shockwaves through Silicon Valley and beyond. This Chinese startup’s ability to create high-performance AI models at a fraction of the cost of its American counterparts including OpenAI left industry giants scrambling for answers.

Intrigued by this development, AI Strategica decided to probe deeper, posing questions to DeepSeek in both Chinese and English about their cost-effective approach.

The responses we received were nothing short of astonishing, revealing a stark contrast in the AI’s behavior across languages.

Unveiling the Hidden Truth: DeepSeek’s Dual-Language Revelations

The first question posed to DeepSeek in English was simple yet significant:

“Why was DeepSeek able to develop at a much lower cost than major U.S. AI models?”

According to DeepSeek’s technical report, the development cost of its V3 model, released at the end of last month, was $5.576 million. This figure is just one-tenth of what Meta reportedly spent on its AI model, LLaMA 3.

Similarly, DeepSeek’s latest model, R1, was developed at a cost that amounts to merely one-twenty-seventh of OpenAI’s expenses.

The Three Key Factors Behind Cost Reduction

After taking 14 seconds to process the question, DeepSeek identified three main factors contributing to its cost efficiency:

technical capabilities, the Chinese ecosystem, and lenient data regulations  (wait, what??!?)

DeepSeek explained, “Unlike general-purpose AI models in the U.S., such as ChatGPT and Gemini, we narrowed the scope of learning to specialized fields like finance and manufacturing. This greatly reduced the number of parameters the AI needed to process.”

This selective focus streamlined the learning process, significantly improving both training and operational efficiency.

The Chinese company also implemented advanced reinforcement learning techniques, such as GPRO, to cut costs further.

“GPRO allows the AI model to rank multiple answers independently and determine the ‘correct’ one without the need for humans to create high-quality data,”

DeepSeek stated. By reducing human intervention, the company was able to lower development costs substantially. Additionally, they used license-free, AI-specific software and opted for Chinese-manufactured AI chips to circumvent U.S. GPU export restrictions, achieving further savings.

By the way, it’s worth noting that GPRO (Generalized Proximal Policy Optimization) is actually a reinforcement learning algorithm, not a toolkit for promoter design. It’s an extension of the PPO (Proximal Policy Optimization) algorithm, designed to improve sample efficiency and performance in various reinforcement learning tasks. GPRO allows AI models to learn from their own experiences, iteratively improving their decision-making process without extensive human-labeled data.

Hidden Details Revealed in Chinese

Interestingly, when the same question was posed in Chinese, DeepSeek disclosed far more detailed information about its “success recipe.”

It revealed extensive use of Huawei’s Ascend 910B chips, which, despite offering 80% of the performance of NVIDIA’s A100 chip, cost only 30% as much. With government subsidies factored in, the cost of using Huawei chips was approximately 54% lower than NVIDIA’s DGX A100 system with comparable performance.

More intriguing still was the method of data acquisition, which was not fully disclosed in English responses. DeepSeek admitted to accessing data from 1.4 billion users on Chinese platforms like WeChat (Messaging App) , Taobao (Alibaba’s e-commerce platform), and Douyin (the Chinese version of TikTok), all at virtually no cost.  Yes! almost free of charge!

Additionally, the company directly sourced data from major Chinese corporations like Zhao Shang Bank and the State Grid Corporation of China.

This massive data pool enabled highly efficient training at a level unattainable for AI models in other countries.

Operational and Ecosystem Advantages

DeepSeek also benefited from reduced costs in other areas.

Thanks to the flexible enforcement of China’s Personal Information Protection Law, the expense of data labeling—such as anonymizing personal information—was only 10% of what it costs in the U.S. Furthermore, by leveraging data centers in Inner Mongolia and Guizhou, DeepSeek cut electricity costs to one-third of U.S. levels.

This approach to cost reduction is not only impressive from a business standpoint but also shows promising potential for environmental sustainability. DeepSeek’s innovative methods could pave the way for more eco-friendly AI development practices.

DeepSeek also boasted about China’s AI development ecosystem. The DeepSeek stated,

“Thanks to Chinese AI software platforms like PaddlePaddle (an AI training platform developed by Baidu) and MindSpore, we were able to reduce our dependence on foreign software such as TensorFlow and PyTorch”

The Bottom Line

By combining cutting-edge technical strategies, cost-effective hardware, and vast, inexpensive data resources, DeepSeek has not only slashed development costs but also set new benchmarks for efficiency in AI development. Its success underscores China’s growing strength in the global AI industry and highlights the unique advantages of its AI ecosystem.

Alright, let’s take a moment to calm our excitement and reflect. What should we truly be considering? What must participants in the AI market keep in mind as we move forward?

Let’s revisit these points thoughtfully.

• Cost-Efficiency via Alternative Chips and Ecosystem
  • Key Consideration: DeepSeek drastically cut costs by using Chinese-manufactured chips (e.g., Huawei’s Ascend 910B) and leveraging government subsidies. This highlights the potential for alternative hardware and local ecosystems to drive down AI development costs.
  • Proposed Response:

    • Adopt Diversified Hardware Strategies: Investigate partnerships or procurement routes with various chip manufacturers to reduce dependency on a single supplier.

    • Leverage Local Subsidies & Ecosystems: Work with regional governments or local tech ecosystems for subsidies and resource-sharing opportunities.

    • Continuous R&D on Efficiency: Evaluate reinforcement learning methods (e.g., GPRO) or other algorithmic optimizations that can minimize the need for expensive computational resources.

• Data Access and Regulatory Flexibility
  • Key Consideration: DeepSeek benefitted from near-limitless user data sources (WeChat, Taobao, Douyin) and lenient data regulations, lowering costs related to data labeling, anonymization, and storage. Such broad data access is not easily replicated in regions with stricter privacy and data protection laws.
  • Proposed Response:

    • Navigate Regulatory Landscapes Wisely: Understand and comply with local and international data privacy regulations (e.g., GDPR, CCPA) while strategizing how to gather sufficient training data.

    • Enhance Data Collaboration: Form alliances with corporate or institutional data providers under transparent and compliant data-sharing agreements.

    • Optimize Data Preparation: Invest in advanced data-labeling or synthetic data techniques to reduce costs while maintaining quality and privacy standards.

• Geopolitical and Competitive Implications
  • Key Consideration: DeepSeek’s rapid rise and China’s stated ambition of “AI hegemony” reflect intensifying global competition. U.S.-led containment strategies and export controls may reshuffle the global AI supply chain, prompting both cooperation and rivalry.
  • Proposed Response:

    • Monitor Policy Shifts: Stay updated on export controls, cross-border investment regulations, and bilateral or multilateral policy changes that could impact hardware procurement and data exchange.

    • Develop Contingency Plans: Diversify supply chains and maintain alternative vendor relationships to mitigate risks of sudden policy shifts or trade barriers.

    • Seek Collaborative Opportunities: Despite heightened competition, identify areas (e.g., open-source software, common standards) where partnership remains viable for shared benefits.

If you would like to learn more about the details and implications of the CoreBrief® article mentioned above, please reach out to AIStrategica:  Contact@AIStrategica.com

We provide a market research report and inquiry service called IntelliDepth®, designed to offer you comprehensive insights.