China's AI 'Six Little Dragons' Collapse: The End of an Era

The era of the 'Six Little Dragons' in China's independent large language model (LLM) sector has officially ended. What began as a hype-driven boom in April 2023 has collapsed into a harsh reality check by late 2025.

Two major players have exited the race for foundational model dominance, leaving only four companies to redefine their existence. This shift marks a critical turning point for global AI investors and developers watching the Asian market.

Key Facts: The Great Consolidation

• Sector Contraction: The group known as the 'AI Six Little Dragons' has shrunk from six members to four due to unsustainable costs.
• Strategic Pivots: Yi Inc. abandoned trillion-parameter pre-training, while Baichuan AI shifted focus entirely to medical applications.
• DeepSeek Disruption: DeepSeek-R1 achieved performance parity with top closed-source models at 1/100th of the training cost.
• Market Shock: NVIDIA’s stock dropped 16.97% in a single day following DeepSeek’s release, wiping out $590 billion in market value.
• IPO Activity: Zhipu AI completed its listing on the Hong Kong Stock Exchange in January 2026 amid high demand.
• Valuation Crisis: Previous valuations based on 'Chinese OpenAI' narratives are now being aggressively rewritten by investors.

The Death of the 'Chinese OpenAI' Narrative

For two years, the dominant narrative in China’s tech sector was the creation of domestic equivalents to OpenAI. Venture capital firms poured billions into startups promising to build the next GPT-4. However, this strategy relied heavily on massive computational resources and continuous cash burns.

The label 'AI Six Little Dragons' emerged in April 2023 as these companies raised significant funding rounds. Investors expected rapid monopolization of the local market. Instead, they faced a brutal efficiency crisis that exposed the fragility of their business models.

By late 2025, the illusion shattered. Two of the six companies could no longer justify the immense costs of training foundational models. Yi Inc., founded by Kai-Fu Lee, made the decisive move to stop developing models with over one trillion parameters. This signaled a retreat from the arms race of raw parameter counts.

Baichuan Intelligence had already begun its transformation earlier. In an internal letter dated April 10, 2025, CEO Wang Xiaochuan admitted that the company’s战线 (battlefront) had been too broad. He announced a strategic contraction to focus exclusively on medical AI. This pivot reflects a broader industry trend where vertical specialization replaces horizontal generalization.

DeepSeek’s Cost Revolution Shatters Valuations

The catalyst for this collapse was not internal failure alone, but external disruption from DeepSeek. On January 20, 2025, DeepSeek released its R1 model. Within days, it became clear that this startup, which was not part of the original 'Six Little Dragons,' had cracked the code on efficiency.

DeepSeek-R1 matched the performance of leading closed-source models while costing just 1% of the training expenditure. This achievement dismantled the core justification for the massive valuations of the other startups. If superior or equivalent intelligence can be built for pennies, why invest billions in traditional pre-training?

The financial markets reacted violently. Just seven days after the release, NVIDIA’s stock price plummeted by 16.97%. This single-day drop erased approximately $590 billion in market capitalization. It stands as a record for the largest single-day loss by a US public company.

While NVIDIA’s hardware dominance remains intact, the incident highlighted a fundamental shift in AI economics. The assumption that more compute always equals better AI is no longer universally true. Algorithmic efficiency and architectural innovations are becoming more valuable than raw hardware power.

Implications for Global Tech Giants

This development forces Western companies to reconsider their own strategies. Silicon Valley giants like Microsoft and Google have spent hundreds of millions on GPU clusters. DeepSeek’s success suggests that software optimization can offset hardware dependencies.

For investors, the lesson is stark. Capital allocation must shift from brute-force scaling to innovative engineering. The 'burn rate' metric is no longer a sign of ambition but a potential indicator of obsolescence.

Survivors Pivot to Vertical Integration

The four remaining companies—Zhipu AI, MiniMax, Moonshot AI, and others—are now forced to prove their worth beyond generic chatbot capabilities. They can no longer rely on the allure of being a 'general purpose' foundation model provider.

Zhipu AI took a significant step toward legitimacy by listing on the Hong Kong Stock Exchange in early January 2026. The IPO was oversubscribed 1,159 times, indicating strong retail investor interest. However, this public listing also subjects them to greater scrutiny regarding profitability and sustainable growth.

These survivors are rapidly moving up the stack. Instead of selling API access to raw models, they are building proprietary applications. This strategy aims to capture value directly from end-users rather than competing in a commoditized infrastructure market.

• Focus on Application: Developing industry-specific tools for finance, law, and healthcare.
• Cost Reduction: Optimizing inference engines to lower operational expenses per query.
• Partnerships: Integrating deeply with enterprise software suites to ensure sticky customer relationships.

Industry Context: A Global Correction

This consolidation in China mirrors trends seen globally. In the West, many well-funded AI startups are struggling to find product-market fit. The barrier to entry for LLM development has lowered, but the barrier to profitability has risen significantly.

OpenAI itself faces pressure to monetize its massive investments. Anthropic and Meta continue to compete fiercely, but the era of easy money for AI ideas is over. The market now demands tangible ROI and clear use cases.

The Chinese market’s correction serves as a warning shot to global investors. Hype cycles can inflate valuations temporarily, but technological realities eventually assert themselves. Efficiency, not just scale, will determine the winners of the AI decade.

What This Means for Developers and Businesses

For developers, the landscape is becoming more accessible yet more competitive. High-quality, open-weight models like those from DeepSeek are available at low costs. This democratizes AI development but raises the bar for application-layer innovation.

Businesses should avoid locking themselves into expensive, proprietary foundational models unless absolutely necessary. Instead, leverage efficient open-source alternatives and focus budget on data quality and domain-specific fine-tuning.

• Adopt Hybrid Models: Combine open-source weights with proprietary data for unique advantages.
• Prioritize Inference Optimization: Invest in technologies that reduce token generation costs.
• Monitor Regulatory Changes: Stay agile as governments worldwide adjust AI governance frameworks.

Looking Ahead: The Next Phase of AI

The next 12 to 24 months will define the long-term structure of the AI industry. We expect further M&A activity as smaller players struggle to survive without massive war chests.

Expect to see a bifurcation in the market. One segment will be dominated by hyper-efficient, open-source models used by developers. The other will consist of highly specialized, enterprise-grade solutions offered by the surviving giants.

The 'Six Little Dragons' may be gone, but their legacy will shape how AI is built and sold. The focus has shifted from who has the biggest model to who has the smartest architecture.

Gogo's Take

• 🔥 Why This Matters: This signals the end of the 'compute-is-king' mentality. For businesses, it means AI adoption costs will drop dramatically, allowing smaller players to compete with tech giants using optimized, open-source models rather than expensive proprietary APIs.
• ⚠️ Limitations & Risks: While training costs are down, inference costs at scale remain a challenge. Additionally, the rapid release of open models increases the risk of misuse and makes regulatory compliance harder for enterprises deploying these tools in sensitive sectors like healthcare.
• 💡 Actionable Advice: Stop waiting for the 'perfect' proprietary model. Start integrating efficient, open-weight models like DeepSeek-R1 or Llama-3 into your workflows today. Focus your engineering resources on building robust application layers and data pipelines, as the model layer is becoming a commodity.