Column by Secretary-General Chu Chen-Tso

The birth of DeepSeek has sent shockwaves through global AI development. For Silicon Valley in the United States, this is not merely a technological event but a psychological and strategic blow. This marks the first time that core US players have acutely realized that the rise of Chinese AI is not out of reach, and in certain aspects, is showing trends of surpassing them.

DeepSeek did not adopt a brand-new technological architecture. Instead, it utilized "distillation" technology to extract optimized parameters from large language models (LLMs), achieving low-cost, low-energy model operation that demonstrates high potential for deployment. This exemplifies China's adept strategy of "imitation, optimization, and re-innovation"—upgrading existing technologies for rapid industrialization.

Crucially, DeepSeek possesses a dual advantage in data: it indirectly integrates Western linguistic achievements through distillation while possessing massive Chinese-language corpora generated by China's population of 1.4 billion. These are precious resources that Silicon Valley finds difficult to acquire, securing DeepSeek a place in the language model competition.

 

The Resource and Mindset Gap in the US Tech Circle

 

The fundamental reason Silicon Valley did not pursue a similar path is that they "lack no resources." According to the Stanford HAI 2024 AI Index Report, US AI companies received $67.2 billion (approx. NT$2.15 trillion) in private investment, far exceeding China's $7.76 billion (approx. NT$248.3 billion). Companies like OpenAI, Meta, and Google can operate without considering resource constraints, adopting the "Large Model Route" and focusing their development on AGI systems with the largest parameter scales and strongest capabilities.

In contrast, although DeepSeek did not break technological boundaries, it forced Silicon Valley to reconsider: Must AI progress rely on massive capital and computing power? Is a larger model scale necessarily more effective? These questions gained traction after Nvidia's stock plunged 17% on the day DeepSeek released its model, evaporating nearly $600 billion (approx. NT$19.2 trillion) in market value. Investors have begun to question whether the demand for high-priced GPUs still holds predictable growth potential.

 

The Battle Between Open Source and Closed Source

 

With the birth of DeepSeek, AI development has officially entered the era of "Open Source vs. Closed Source."

Models from OpenAI, Google, and Anthropic belong to closed architectures; they provide API access but do not disclose internal training data or model parameters. Users must pay high fees and face limited fine-tuning capabilities. In contrast, DeepSeek fully opens its model weights, training methods, and architecture, allowing developers to perform local deployment and customized adjustments. This is particularly beneficial for small and medium-sized enterprises (SMEs), startups, and developing countries.

Although the open-source model cannot rely on API licensing for profit like closed-source models, it operates on a different economic logic—establishing "ecosystem lock-in" and an "influence economy." By attracting developers to participate and contribute through open source, it builds a powerful community network. For instance, NVIDIA's CUDA platform created a highly sticky ecosystem that is difficult to migrate away from through years of cultivating the developer community. DeepSeek is replicating this model.

Additionally, open-source models carry geopolitical implications. In the race for AI sovereignty, the rise of DeepSeek will deepen the global influence of Chinese models. Its free, locally deployable nature significantly lowers the entry barrier for developing countries that have not yet built their own language models, thereby reshaping the global AI landscape.

 

Information Security and National Security: The Invisible Battleground of Trust

 

However, DeepSeek's development also raises privacy and national security concerns. Government agencies in Taiwan and the United States have successively banned the use of DeepSeek in the public sector, primarily due to fears that user data could fall under the control of the Chinese government. Although DeepSeek lists a privacy policy, it remains vague regarding data storage, query mechanisms, and third-party sharing, lacking complete guarantees.

Chapter 3 of China's Personal Information Protection Law grants the government access rights, causing external doubts about whether model data is secure or if it will be used for intelligence purposes. Unless DeepSeek establishes "data islands" outside of China—as TikTok attempted—these information and national security concerns will be difficult to resolve in the short term.

Currently, DeepSeek appears more inclined to continue expanding its global market and collecting training data rather than implementing distributed deployment for specific markets. This poses a trust challenge in its internationalization process.

 

Challenges and Dilemmas in Building Taiwan's "Sovereign AI"

 

Taiwan holds a global advantage in the AI hardware sector, leading the world in everything from wafer manufacturing to server assembly. However, its development in software and language models is relatively weak. Facing the wave of competition for AI sovereignty, countries like Singapore, Japan, and South Korea have all initiated local language model construction. Taiwan also launched the TAIDE sovereign model project in 2023, utilizing Meta's LLaMA open-source model for Traditional Chinese training.

However, TAIDE still faces three major challenges:

  1. Insufficient Training Data: Traditional Chinese corpora are limited, and laws do not clearly authorize the use of corpora, making data acquisition difficult.

  2. Insufficient Computing Resources: The government has only procured hundreds of GPUs, which pales in comparison to the tens of thousands in South Korea or the hundreds of thousands at OpenAI.

  3. Reliance on Third-Party Models: Taiwan lacks a self-built language model and must rely on architectures like LLaMA, which harbors potential risks regarding information security, national security, and privacy.

 

Conclusion: Finding Balance Amidst Competition and Risk

 

The emergence of DeepSeek symbolizes a shift in global AI development from a closed-source monopoly to diversified competition. It also reshapes the AI industry from a "performance race" to an "ecosystem competition." The controversies surrounding its Chinese background and security issues will force nations to accelerate the construction of sovereign models and push for supply chain diversification and security.

For Taiwan, the only way to consolidate its position in the global AI strategic competition is to face the software gap, address the three major challenges (data insufficiency, lack of computing resources, and reliance on third-party models), and find a balance between open source and national security.

In the future, if Taiwan can simultaneously deepen local language model R&D and leverage its hardware advantages to build a trustworthy and culturally distinct AI sovereign model, it will be able to forge its own path of AI development within this dual-track competition of technology and values.

Source:https://futurecity.cw.com.tw/article/3667

Back to the top